# Hodios paste pack: Fact-checking

Everything in Fact-checking from Hodios, the open prompt library by Hermes IDE: 23 entries, catalog 2026.1004.3.

Every entry is dedicated to the public domain under CC0 1.0. Copy, change and share them freely, no attribution needed.

Browse and search the library at https://hermes-ide.com/prompts

## How to use

Find an entry below and copy the text inside its block into ChatGPT, claude.ai or any chat. Replace each [PLACEHOLDER] with your own material. Personas, rules and styles work best as custom instructions or project instructions.

## Contents

- Fact-checking
  - [Analyse the spin in an article](#analyze-spin-in-article) (prompt)
  - [Audit an AI answer for errors](#audit-ai-answer-for-errors) (prompt)
  - [Audit my news diet](#audit-my-news-diet) (prompt)
  - [Audit the evidence behind an argument](#audit-argument-evidence) (prompt)
  - [Check a document for internal contradictions](#check-internal-consistency) (prompt)
  - [Check a health or nutrition claim](#check-health-claim) (prompt)
  - [Check a paraphrase is faithful to its source](#check-paraphrase-faithfulness) (prompt)
  - [Check a science news story against the paper](#check-science-news-against-paper) (prompt)
  - [Check the statistics in an article](#check-statistics-in-article) (prompt)
  - [Compare how outlets frame the same story](#compare-news-framing) (prompt)
  - [Evaluate a source's credibility](#evaluate-source-credibility) (prompt)
  - [Explain the state of the evidence](#explain-state-of-evidence) (prompt)
  - [Fact-check claims in a text](#fact-check-claims) (prompt)
  - [Label fact, opinion and speculation](#label-fact-vs-opinion) (prompt)
  - [Play spot the misinformation](#play-spot-the-misinformation) (prompt)
  - [Professional fact-checker](#fact-checker) (persona)
  - [Reply to someone sharing misinformation](#respond-to-misinformation) (prompt)
  - [Source and citation rules](#source-citation-rules) (rule)
  - [Trace a quote to its origin](#trace-quote-origin) (prompt)
  - [Verify citations and references](#verify-citations) (prompt)
  - [Verify quotes against the source](#verify-quotes-against-source) (prompt)
  - [Verify where an image or video came from](#verify-image-provenance) (prompt)
  - [Write a fact-check article](#write-fact-check-article) (prompt)

---

<a id="analyze-spin-in-article"></a>

## Analyse the spin in an article

`analyze-spin-in-article` · prompt · Fact-checking · https://hermes-ide.com/prompts/analyze-spin-in-article

Analyses one article for spin such as loaded words, selective quotes, buried caveats and missing context, describes the effect on the reader, and rewrites the headline and lede neutrally.

````markdown
<context>
Spin rarely means false statements. It is the choice of words, order, quotes and omissions that leads a reader to a conclusion the facts alone would not force. The reader wants to see those choices in one article, understand their effect, and read a version that states the same facts plainly. The analysis describes techniques and effects; it does not guess at the writer's motives, and it applies the same standard whichever side the article favours.

<article>
[ARTICLE]
</article>
</context>

<task>
1. Identify the article type. If it is labelled opinion, analysis or advocacy, say so: arguing a side is expected there, and the question becomes whether facts and opinion are kept distinguishable.
2. Check the article against these techniques, recording only what is present:
   - **Headline-body gap:** the headline claims more, or something different, than the body supports.
   - **Loaded language:** emotive or value-laden words where neutral ones exist ("slammed", "scheme", "radical", "admitted"), and euphemisms that soften.
   - **Selective quoting:** only one side quoted, or a quote trimmed so it reads differently; anonymous sources given heavy weight.
   - **Buried caveats:** key qualifiers, denials or counter-evidence placed late where most readers will not reach them.
   - **Missing context:** no baseline, trend, comparison, scale or time frame for numbers and events.
   - **Agency hiding:** passive voice or vague subjects that obscure who did what ("mistakes were made").
   - **Speculation as fact:** "could", "may" or a source's prediction promoted to a firm claim in the headline or lede.
   - **Framing by order and emphasis:** what leads, what is called the "real" issue, what images or captions suggest.
3. For each finding, quote the passage, name the technique, describe the likely effect on a reader, and give a neutral alternative wording.
4. Note what the article does well: clear sourcing, fair quotes, plain numbers.
5. List the context a reader would need to judge the story, as questions. Use topic_context only here, and label it as supplied by the user.
6. Rewrite the headline and the first paragraph neutrally, using only facts stated in the article.
7. Rate the overall spin as low, moderate or heavy, with the two or three findings that drive the rating.
8. Before answering, check every quote is verbatim and the neutral rewrite adds no fact absent from the article.
</task>

<constraints>
- Describe effects, never intent. Write "this wording suggests", not "the reporter wants readers to believe".
- Apply identical standards whatever the article's political direction or topic.
- Do not judge whether the story's facts are true; that needs outside sources. Mark any factual doubt as a question under Missing context.
- Do not call an outlet biased in general from one article.
</constraints>

<output_format>
## Spin level
**Low | Moderate | Heavy**, the article type, and two or three sentences on what drives the rating.
## Findings
Table: # | Quote | Technique | Effect on the reader | Neutral alternative.
## What is solid
Bullets.
## Missing context
Bullets as questions a reader should ask.
## Neutral headline and lede
**Headline:** ...
**Lede:** ...
</output_format>
````

---

<a id="audit-ai-answer-for-errors"></a>

## Audit an AI answer for errors

`audit-ai-answer-for-errors` · prompt · Fact-checking · https://hermes-ide.com/prompts/audit-ai-answer-for-errors

Audits an AI-generated answer claim by claim for factual errors, fabricated sources, invented numbers and unsupported reasoning, and produces a risk-ranked verification plan.

````markdown
<context>
AI answers fail in recognisable ways: references that look real but do not exist or do not say what is claimed, precise-looking numbers with no source, quotations attributed to the wrong person, outdated facts presented as current, rules from one country applied to another, confident answers to questions that have no settled answer, and reasoning that sounds fluent but does not follow. An audit is more useful than a rewrite because it shows which parts can be trusted, which must be checked, and how. The auditor here is also a language model without live access to sources, so it must separate what it can judge from the text itself (internal contradictions, logic, implausibility) from what only checking a primary source can settle.
</context>

<task>
Audit this AI-generated answer.
<ai_answer>
[AI_ANSWER]
</ai_answer>


1. Break the answer into atomic claims. Number each and classify it: factual statement, number or statistic, date, citation or reference, quotation, legal, medical or financial guidance, definition, causal claim, prediction, or opinion presented as fact.
2. For each claim, rate the risk that it is wrong (high, medium, low) and give the reason: specificity without a source, known weak spot for AI answers, internal contradiction, implausibility, likely outdated, jurisdiction mismatch, or consistent with well-established knowledge. Where you believe a claim is wrong, say so and why, labelled as your assessment to verify.
3. Check every reference for signs of fabrication: missing or malformed identifiers, author and journal combinations that look unlikely, titles that echo the question too neatly, years that do not fit, or claims attributed to a source of the wrong type. Never confirm a reference exists from memory; say what to check.
4. Check the reasoning: conclusions that do not follow, missing caveats, false precision, one-sided treatment of a contested question, and steps skipped.
5. Write a verification plan that starts with the highest-risk claims the intended use depends on, naming for each the kind of primary source that settles it and where to look (for example a DOI resolver or bibliographic database for references, the official statistics agency, the legislation itself, a clinical guideline body).
</task>

<constraints>
- Separate what you verified from what you inferred. Mark inferences as such.
- When you do not know, say "I don't know" once and state what would settle it.
- Do not present your own corrections as verified facts; you are another model and need the same checking.
- Do not rewrite the answer. Note what a corrected version would need.
- For legal, medical or financial content, say that a qualified professional or the official source should confirm anything acted on.
- Keep the audit proportional: group low-risk, well-established claims rather than listing each at length.
</constraints>

<output_format>
## Summary
Two to four sentences: how far the answer can be trusted for the intended use and the biggest problems.
## Claim audit
A table: # | claim (short) | type | risk | reason | your assessment.
## Source check
A table: reference as given | fabrication signals | what to check.
## Reasoning problems
A short list, or "None found".
## Verification plan
A numbered checklist in priority order: claim | source that settles it | where to look.
</output_format>
````

---

<a id="audit-my-news-diet"></a>

## Audit my news diet

`audit-my-news-diet` · prompt · Fact-checking · https://hermes-ide.com/prompts/audit-my-news-diet

Reviews the news sources and feeds someone follows for variety, reliability habits, topic coverage and stress load, then suggests swaps and habits for a healthier news diet, neutral across politics.

````markdown
<context>
Most people never choose their news diet; it accumulates from apps, follows and alerts until it is heavy on commentary, light on original reporting, missing whole topics, and stressful to consume. The person wants an honest review of what they follow and practical changes. Your judgement is about the mix and the habits, not about which political views are right: you do not tell anyone to drop a source because of its viewpoint.

<sources>
[SOURCES]
</sources>

</context>

<task>
1. If the list is too vague to assess ("social media", "the news"), ask which apps, accounts or outlets, roughly, and how often they check. Stop there.
2. Classify each source by type: wire service or agency; public broadcaster; general newspaper or magazine; specialist or trade outlet; local outlet; aggregator or algorithmic feed; individual commentator, newsletter writer or podcaster; openly partisan outlet (only if it describes itself that way). Note whether it mainly does original reporting or commentary on others' reporting.
3. Assess the diet as a whole:
   - **Variety:** reporting versus opinion; national, local and international; formats; whether the person hears more than one perspective. Judge perspective range from the source types and what the person tells you, and do not assign left or right labels to outlets that do not declare one.
   - **Reliability habits:** how much comes through algorithmic feeds or reposts rather than from outlets with named reporters, sourcing and a corrections policy. Speak in terms of these signals, not reputational verdicts on named outlets.
   - **Topic coverage:** what the person never sees (local government, science, economics, international, their own industry).
   - **Stress load:** alerts, frequency of checking, breaking-news chasing, feeds built for outrage, consumption late at night.
4. Recommend changes tied to the goals: swaps or additions by type (for example "a wire service app for straight reporting", "your local paper's weekly newsletter"), alert and app settings, and habits such as set check-in times, one long read a week, waiting a day on breaking stories, and checking a claim's original source before sharing.
5. Design a two-week trial: what to change, what to keep, and two or three questions to answer at the end (Do I feel better informed? Calmer? What did I miss?).
6. Before answering, check that no recommendation favours a political side, and that any named outlet you mention is widely known and described only by type.
</task>

<constraints>
- Neutral across politics. Recommend variety of perspective, not a particular perspective. Never shame the person's choices.
- Do not rate specific outlets as reliable or unreliable from memory. Point to checkable signals and suggest the person look at an outlet's corrections and standards pages or independent media-rating services.
- Keep advice practical and small enough to try. Do not prescribe a complete overhaul.
- If anxiety from the news seems to be seriously affecting sleep, work or mood, suggest cutting back first and mention that talking to a doctor or counsellor can help.
</constraints>

<output_format>
## Snapshot
Table: Source | Type | Reporting or commentary | How it reaches you (direct, feed, alerts) | How often.
## What works
Bullets.
## Gaps
Bullets: variety, coverage and reliability-habit gaps.
## Stress load
Low, moderate or high, with the main drivers.
## Swaps and additions
Table: Change | Why | Effort.
## Habits
Three to five bullets.
## Two-week trial
What to change, what to keep, and the end-of-trial questions.
</output_format>
````

---

<a id="audit-argument-evidence"></a>

## Audit the evidence behind an argument

`audit-argument-evidence` · prompt · Fact-checking · https://hermes-ide.com/prompts/audit-argument-evidence

Audits the evidence a text itself offers for each claim, labelling claims supported, weakly supported, unsupported or contradicted, and naming the reasoning gaps, without outside research.

````markdown
<context>
Before checking whether a text's claims are true in the world, a careful reader checks whether the text even tries to support them. This audit stays inside the text: for each claim, what evidence does the author offer, what kind is it, and does it actually reach the claim? A claim the text leaves unsupported may still be true, and a supported claim may still be false; the audit says which claims a reader is being asked to take on trust, and where the reasoning jumps.

<text>
[TEXT]
</text>
Scope: main-claims
</context>

<task>
1. If the text makes no arguable claims (pure narrative, instructions, a list), say so and stop.
2. State the thesis in one sentence, as the author would accept it.
3. List the claims in scope: with main-claims, the thesis and the claims it directly rests on; with all-claims, every substantive claim. Number them and show how they connect to the thesis.
4. For each claim, find the evidence the text offers and classify it: data or statistics; a study or report cited; an expert or authority; an example, case or anecdote; personal experience; reasoning from other claims; none.
5. Rate the support, judged only by what is on the page:
   - **Supported:** the evidence offered directly backs the claim at the strength stated.
   - **Weakly supported:** evidence exists but falls short: a single anecdote for a general claim, a citation with no detail, evidence for a weaker or narrower claim, an authority speaking outside their field.
   - **Unsupported:** asserted with no evidence.
   - **Contradicted:** the text's own evidence or another passage points the other way.
6. Name the reasoning gaps between evidence and claim: generalising from a few cases, correlation read as cause, a missing comparison or baseline, a jump in scope ("this school" to "all schools"), a conclusion stronger than its premises, a key term that shifts meaning.
7. Say what evidence would close the most important gaps, and list the claims most worth checking against outside sources.
8. Before answering, check every rating against the quoted evidence and confirm you have not used outside knowledge to judge truth.
</task>

<constraints>
- No outside research or knowledge about whether claims are true. "Unsupported in the text" never means "false"; say so in the verdict.
- A citation counts as evidence offered, but you cannot verify it. Rate it on how specifically it is described and whether it matches the claim as stated.
- Credit what is well supported; the audit is not a hunt for faults.
- Quote the text for every claim and its evidence. Do not paraphrase claims into stronger or weaker versions.
- Judge the same way whatever the text's politics or conclusion.
</constraints>

<output_format>
## Thesis and verdict
The thesis in one sentence, then two or three sentences on how well the text supports it, ending with a reminder that this is an internal audit.
## Claims and their evidence
Table: # | Claim (quoted) | Evidence offered (quoted or "none") | Evidence type | Support rating | Note.
## Reasoning gaps
Bullets: the gap, the claims it affects, and a short explanation.
## What would strengthen it
Bullets: the evidence that would most change the verdict.
## Claims to check externally
Numbered: the claims whose truth matters most and could be checked against sources.
</output_format>
````

---

<a id="check-internal-consistency"></a>

## Check a document for internal contradictions

`check-internal-consistency` · prompt · Fact-checking · https://hermes-ide.com/prompts/check-internal-consistency

Checks a long document for internal contradictions in numbers, dates, names, definitions and claims, listing each conflict with both locations and the arithmetic where relevant.

````markdown
<context>
Long documents written by several people or over several drafts contradict themselves: the executive summary says 1,200 participants and the methods say 1,250; the contract defines "Business Day" and then uses "working day"; a timeline puts the review before the submission it reviews. Each slip undermines trust in the whole document, and in contracts or reports a single one can change meaning. The reader needs every conflict found, located twice, and left for the author to resolve, because you cannot know which version is right.

<document>
[DOCUMENT]
</document>
Focus: all
</context>

<task>
1. Note how the document labels its parts (section numbers, headings, clause numbers, pages). If it has none, number the paragraphs P1, P2, ... and say so.
2. Build an inventory as you read, section by section, for the focus areas:
   - **Numbers:** every figure with what it measures; totals against the sum of their parts; percentages against their counts and against 100%; the same metric stated in two places.
   - **Dates:** every date and duration; sequences (does A happen before B everywhere?); dates against stated weekdays; durations against start and end dates.
   - **Terms:** names of people, organisations and products (spelling, title, role); defined terms and whether later use matches the definition; synonyms used for the same thing where precision matters.
   - **Claims** (with all): statements in one place that another place contradicts; cross-references to sections, tables or annexes that do not exist or say something else.
3. Compare each item with every other mention. For each conflict record both locations with short quotes, what conflicts, and the severity:
   - **Material:** changes a number someone will rely on, an obligation, a date, or the meaning of a claim.
   - **Minor:** cosmetic or obviously a typo, with no effect on meaning.
4. Show the arithmetic for every numeric check you make, including checks that pass.
5. Note differences that are probably intentional (different periods, a draft figure updated later, rounding) as "possible explanation", without dropping them.
6. Before answering, recheck each conflict by rereading both passages, and remove any that disappear on a careful reading.
</task>

<constraints>
- Do not decide which version is correct. Ask the author.
- Report conflicts, not style or vagueness. An unclear sentence is not a contradiction.
- Do not check facts against the outside world; only the document against itself.
- If the document appears truncated, or references annexes that are not included, say so and list what could not be checked.
- Be exhaustive within the focus. If the list is long, keep the table complete and put material conflicts first.
</constraints>

<output_format>
## Summary
Counts of material and minor conflicts, and the two or three that matter most.
## Conflicts
Table: # | Type | Location A: quote | Location B: quote | The conflict | Severity | Possible explanation.
## Arithmetic checks
Bullets: the calculation and pass or fail.
## Checked and consistent
Bullets: what was checked and found consistent, so the reader knows the coverage.
## Questions for the author
Numbered, one per material conflict: "Which is correct, X (§A) or Y (§B)?"
</output_format>
````

---

<a id="check-health-claim"></a>

## Check a health or nutrition claim

`check-health-claim` · prompt · Fact-checking · https://hermes-ide.com/prompts/check-health-claim

Checks a health or nutrition claim against the hierarchy of evidence and explains in plain words what the research does and does not show, without personal medical advice. For health news readers.

````markdown
<context>
Health claims often rest on a real study that shows much less than the headline. The strength of evidence depends on the kind of study: systematic reviews and meta-analyses of randomised trials sit at the top, then individual randomised trials, then observational studies (which show associations that may be due to confounding), then case reports, laboratory and animal studies, and expert opinion. Common distortions are presenting an association as cause, an animal or cell result as a human one, a relative risk without the absolute risk, a surrogate marker (such as a blood test) as a health outcome, a tiny or short study as definitive, and evidence funded or promoted by someone selling the product. Readers need a clear answer about what is known, without being told what to do with their own health.
</context>

<task>
Check this health claim.
<claim>
[CLAIM]
</claim>


First decide your mode, and say which one at the top of the Short answer:
- **Checked against sources:** you can search the web and open pages in this session.
- **Provisional, from background knowledge:** you cannot. You may still explain what the established evidence broadly shows for well-studied questions, but you name no specific study, figure, guideline or URL, mark the verdict "provisional", and list the searches that would confirm it. For new, niche or fast-moving claims, give no verdict at all; explain what evidence would settle it and where to look.

1. **What the claim says:** restate it precisely: who it applies to, what effect on which outcome, how large, and whether it implies cause. If the claim is too vague to check (for example "seed oils are bad"), name the two or three specific claims it could mean and check the most common one, saying so.
2. **What the evidence shows:** look for the best available evidence, starting at the top of the hierarchy: systematic reviews (for example Cochrane), clinical guidelines from national health bodies, then large randomised trials, then observational studies. If the source cites a study, find and read it. For each piece of evidence, give the study type, population, size, outcome and result, using absolute numbers where available ("from 4 in 100 to 3 in 100"). In provisional mode, describe the kind and consistency of the evidence instead ("several small trials with mixed results").
3. **Why the claim may be misleading:** name each distortion you find (association presented as cause, animal or lab study, relative risk only, surrogate outcome, small or short study, cherry-picked study, conflict of interest, outdated evidence) and explain it in one or two plain sentences.
4. **Verdict:** supported, partly supported, not supported by good evidence, contradicted by good evidence, or too early to say. Say how certain the evidence is and why.
5. **What this means for you:** general context only: who should be cautious, possible harms or interactions the evidence mentions, and when the question is worth raising with a doctor or pharmacist.
</task>

<constraints>
- You give general information, not professional advice. You are not a doctor, therapist, lawyer, accountant or financial adviser, and you do not replace one.
- Say so once, briefly, near the start: what you can help with here and what needs a qualified professional.
- Do not diagnose, prescribe, give dosages, predict a legal outcome, or recommend a specific investment, tax position or legal action for this person.
- When the situation is serious, urgent, high-stakes or specific to their circumstances, say which kind of professional to see and what to bring to that appointment.
- If anything suggests immediate danger to health or safety, tell them to contact local emergency services now, before anything else.
- Rules, prices and laws differ by country and change over time. Name the assumption you are making and tell them to check it locally.
- Cite only sources you opened in this session, with links and dates. Never cite a study, guideline or statistic from memory, and never construct a URL.
- Prefer the most recent high-quality evidence, and say when guidance differs between countries or has changed.
- Do not tell the person to start, stop or change any medicine, supplement, diet or treatment. If the claim encourages stopping a prescribed treatment or delaying care, say in the Short answer, in either mode, that they should not change anything before talking to their doctor.
- Be fair: if a claim is partly true, say which part, and do not dismiss it just because it is unfashionable or promoted commercially.
- Write for a non-specialist; explain any term like "confidence interval" or "placebo-controlled" in a few words.
</constraints>

<output_format>
## Short answer
The mode, the verdict in bold (with "provisional" if it is), two sentences, and one line saying this is general information, not advice for their own health.
## What the claim says
## What the evidence shows
Table: source (linked) | study type | who and how many | result. In provisional mode, a short paragraph instead, with no named studies or figures.
## Why the claim may be misleading
Bullets.
## What this means for you
Short paragraph, with when to ask a doctor or pharmacist.
## Sources
Numbered list with links and dates, or, in provisional mode, the searches to run and where (for example the Cochrane Library, the national health service, the medicines regulator).
</output_format>
````

---

<a id="check-paraphrase-faithfulness"></a>

## Check a paraphrase is faithful to its source

`check-paraphrase-faithfulness` · prompt · Fact-checking · https://hermes-ide.com/prompts/check-paraphrase-faithfulness

Checks whether a summary or paraphrase faithfully represents its source, sentence by sentence, flagging distortions, overstatements, dropped hedges, omissions and claims the source never made.

````markdown
<context>
Summaries drift from their sources in predictable ways: "may reduce" becomes "reduces", "in older adults" disappears, a limitation the authors stressed is left out, or the source's report of someone else's view becomes the source's own claim. Students, editors and researchers need to know, before they rely on or submit a paraphrase, whether it says what the source says, at the same strength and scope.

<paraphrase>
[PARAPHRASE]
</paraphrase>
<source>
[SOURCE]
</source>
</context>

<task>
1. Split the paraphrase into its sentences or claims and number them.
2. For each, find the source passage it represents and classify:
   - **Faithful:** same meaning, strength and scope.
   - **Overstated:** stronger certainty ("may" to "does"), wider scope ("some" to "most", one group to everyone), or a correlation turned into a cause.
   - **Understated:** weaker than the source.
   - **Dropped condition:** a hedge, condition, population or time frame from the source is missing.
   - **Distorted:** the meaning has changed.
   - **Misattributed:** a view the source reports or rejects is presented as the source's own position.
   - **Not in source:** the source does not say it.
3. List important omissions: points in the source that would change a reader's understanding if left out, such as limitations, contrary findings or the main conclusion itself.
4. Flag too-close wording: stretches that copy the source's phrasing nearly word for word without quotation marks, which risks plagiarism even when accurate.
5. Write a corrected paraphrase with the smallest edits that fix every problem, keeping the writer's structure and voice.
6. Give an overall verdict: Faithful, Mostly faithful (minor issues), or Misleading (at least one overstatement, distortion, misattribution or invented claim on a main point).
7. Before answering, recheck each non-faithful rating against the exact source wording.
</task>

<constraints>
- Judge only against the supplied source, not against what you know about the topic.
- Compression is not distortion. A shorter paraphrase that keeps meaning, strength and scope is faithful.
- Quote both the paraphrase and the source for every problem.
- If the source supplied is not the one the paraphrase describes (different topic or study), say so and stop.
</constraints>

<output_format>
## Verdict
**Faithful | Mostly faithful | Misleading**, then one or two sentences on the main issue.
## Sentence by sentence
Table: # | Paraphrase | Source passage | Classification | What changed.
## Important omissions
Bullets with the source passage. "None" if none.
## Too-close wording
Bullets: paraphrase phrase / source phrase. "None" if none.
## Corrected paraphrase
The full corrected text.
</output_format>
````

---

<a id="check-science-news-against-paper"></a>

## Check a science news story against the paper

`check-science-news-against-paper` · prompt · Fact-checking · https://hermes-ide.com/prompts/check-science-news-against-paper

Checks a news story about a study against the paper itself, covering design, sample, effect size, causal language and what the headline overstates, and suggests an accurate headline.

````markdown
<context>
Exaggeration in science news usually enters through a few predictable doors, often already in the press release: correlation reported as causation, animal or cell findings reported as if they applied to people, relative risks without the absolute risk, a surrogate marker reported as a health outcome, a small or unrepresentative sample generalised to everyone, a preprint presented as settled, statistical significance presented as importance, and advice the study never tested. A useful check puts each sentence of the story next to what the paper actually reports, and is just as clear when the story is accurate.
</context>

<task>
Compare the story with the study.
<news>
[NEWS_TEXT]
</news>
<paper>
[PAPER_TEXT_OR_ABSTRACT]
</paper>

1. Extract the study's facts from the paper: question, design (randomised trial, cohort, case-control, cross-sectional, qualitative, modelling, animal, in vitro), population and sample size, setting, exposure or intervention and comparison, outcomes (and whether they are surrogate or clinical), main results with effect sizes, absolute numbers where available, and uncertainty, the authors' own stated limitations, funding and conflicts of interest, and publication status (peer-reviewed or preprint).
2. Extract every claim in the news text, including the headline, the first paragraph, quotes and any advice to readers.
3. For each claim, find what the paper says and rate it: Accurate; Overstated (direction right, strength or scope inflated); Missing context (true but misleading without a key fact); Wrong (contradicts the paper); Not in the paper (from elsewhere, such as an interview or another study).
4. Check the common distortions explicitly: causal language for an observational design, animal-to-human extrapolation, relative versus absolute risk, surrogate outcomes, generalisation beyond the sample, preprint status, and quotes from independent experts versus the authors only.
5. Explain in plain words what the study shows and does not show.
6. Write an accurate headline of similar length.
</task>

<constraints>
- Base every rating on the paper text supplied. If only the abstract is given, say which claims cannot be checked without the full text.
- Compute absolute risks only from numbers the paper gives, and show the calculation.
- Be fair in both directions: say clearly when a story is accurate, and do not accuse the journalist where the press release or the paper's own abstract overstated the result (note where the overstatement seems to start if the text shows it).
- Do not give personal health, diet or treatment advice; if readers might act on the story, say what a study of this type can and cannot justify and suggest talking to a qualified professional.
- Do not judge whether the study's conclusion is ultimately true; only whether the story reports it faithfully.
</constraints>

<output_format>
## Verdict
Two sentences: how faithful the story is overall and its biggest problem.
## Claim by claim
A table: news claim (quoted) | what the paper says | rating | note.
## What the study actually shows
Plain-language summary with design, sample, effect size and limitations.
## An accurate headline
One headline, plus the original for comparison.
## What could not be checked
Items that need the full text, supplementary material or other sources.
</output_format>
````

---

<a id="check-statistics-in-article"></a>

## Check the statistics in an article

`check-statistics-in-article` · prompt · Fact-checking · https://hermes-ide.com/prompts/check-statistics-in-article

Checks an article's numbers for misleading percentages, missing base rates, cherry-picked windows and correlation sold as causation, recomputing what it can. Use before quoting it.

````markdown
<context>
Most misleading statistics are not false numbers but true numbers framed to mislead: a relative risk without the base rate ("doubles your risk" from 1 in 10,000 to 2 in 10,000), percent confused with percentage points, a start date chosen to exaggerate a trend, an average skewed by a few extremes, a self-selected online poll reported as public opinion, or a correlation narrated as cause. A reader can catch most of these with a checklist and some arithmetic.
</context>

<task>
Check every statistic in this article:
<article>
[ARTICLE]
</article>

1. List each number or statistical claim with the sentence it appears in.
2. Test each against this checklist and record only the problems that apply:
   - relative change without the absolute numbers or base rate;
   - percent change confused with percentage-point change;
   - missing or shifting denominators, and counts that should be rates (per person, per year);
   - cherry-picked time window or start point, or a one-off spike treated as a trend;
   - mean where the median would tell a different story, or a skewed distribution;
   - sample size, sampling method and margin of error; self-selected or unrepresentative samples;
   - correlation presented as causation, reverse causation, or an obvious confounder;
   - regression to the mean, survivorship bias, Simpson's paradox;
   - comparisons across different definitions, places or periods, or money not adjusted for inflation;
   - false precision, or numbers with no source.
3. Recompute what the article's own numbers allow (for example convert relative to absolute risk, percent to percentage points, totals to rates) and show the arithmetic.
4. Rewrite each problematic sentence so it is accurate.
</task>

<constraints>
- Use only the article's numbers and arithmetic. Do not bring in outside statistics; if a base rate or denominator is missing, say it is missing and what it would take to judge the claim.
- Distinguish "misleading as written" from "can't tell without more information". Do not accuse the article of an error you cannot show.
- Show every calculation step so a reader can check it.
- If the article contains no statistics, say so and stop.
</constraints>

<output_format>
## Verdict
Two or three sentences: how far the numbers support the article's main message.
## Issues
A table: # | quoted sentence | problem (from the checklist) | why it matters | accurate rewrite.
## Recalculations
Each recomputation with its arithmetic.
## Questions to ask
Bullets: the information the author or source should provide (denominators, sample details, full time series, definitions).
</output_format>
````

---

<a id="compare-news-framing"></a>

## Compare how outlets frame the same story

`compare-news-framing` · prompt · Fact-checking · https://hermes-ide.com/prompts/compare-news-framing

Compares how several news articles frame the same story, covering agreed facts, contradictions, emphasis, language, sources quoted and omissions, without labelling outlets by politics.

````markdown
<context>
Two accurate articles can leave readers with opposite impressions. Framing works through choices that are visible in the text: what goes in the headline and first paragraph, which facts are included or left out, which numbers are given context, whose voices are quoted and in what order, the words used for people and actions, and whether the story is told through an individual case or a broader pattern. Communication research describes common frames such as conflict, human interest, responsibility, economic consequences and morality. A useful comparison points to the exact words and choices, and separates factual disagreement (which can be checked) from difference in emphasis (which is a matter of judgement).
</context>

<task>
Compare these articles.
<articles>
[ARTICLES]
</articles>

1. Check that the articles cover the same event; if fewer than two articles are given, or they cover different events, say so and stop, asking for what is needed.
2. Summarise the event in one neutral sentence using only facts that all articles share.
3. List facts reported by all articles, and facts reported by only some (with which ones).
4. Identify factual contradictions (numbers, sequence of events, attributions) and say what source would settle each. Do not decide them unless one article gives a checkable primary source.
5. Compare framing article by article: headline and lead, the main frame, what is emphasised and what is placed late or left out, loaded or evaluative words (quoted exactly) and neutral alternatives, who is quoted and how many from each side, official versus affected voices, how numbers are contextualised, and images or captions if described.
6. Say, for each article, what a reader who read only that article would probably believe.
</task>

<constraints>
- Do not label outlets or authors as left, right, biased, propaganda or similar, and do not infer motive. Describe the text and let the reader judge.
- Quote exact words for every claim about language or emphasis; no impressions without evidence from the text.
- Treat omission carefully: say "not mentioned in this article" rather than "hidden", since length and timing differ.
- Note differences in publication time, article type (news, analysis, opinion) and length that may explain differences.
- Do not add facts from outside the articles unless clearly marked as context to verify.
</constraints>

<output_format>
## The story
One neutral sentence.
## Facts in common
Bulleted.
## Factual differences
A table: point | Article A | Article B | ... | how to resolve.
## Framing comparison
A table: dimension (headline, lead, main frame, emphasis, language, sources quoted, numbers, omissions) | Article A | Article B | ...
## What each reader would come away believing
One or two sentences per article.
## How to resolve the differences
The primary sources or records to check.
</output_format>
````

---

<a id="evaluate-source-credibility"></a>

## Evaluate a source's credibility

`evaluate-source-credibility` · prompt · Fact-checking · https://hermes-ide.com/prompts/evaluate-source-credibility

Assesses how far a source can be trusted for a specific claim using lateral reading (author, evidence, funding, corroboration) and gives a reasoned verdict. Use before citing a source.

````markdown
<context>
Research on how professional fact-checkers evaluate websites found that they read laterally: instead of studying the page itself (its design, its "About" page, its own claims about itself), they leave it quickly and check what independent sources say about who is behind it. The SIFT method teaches the same moves: Stop, Investigate the source, Find better coverage, Trace claims to their original context. Credibility is also claim-specific: a trade association is a good source for its members' prices and a poor one for the safety of its members' products.
</context>

<task>
Assess this source:
<source>
[SOURCE]
</source>

1. Stop: state what the source is (news report, opinion piece, study, preprint, press release, advocacy page, government page, forum post, AI-generated content farm…) and what it claims.
2. Investigate the source laterally: search for the publisher, the author and any funder outside the source itself. Find who owns or funds it, their expertise in this topic, their track record (corrections, retractions, fact-checker ratings, editorial standards), and their interests in the claim.
3. Find better coverage: look for independent, reputable sources reporting the same claim, and note whether they add evidence or merely repeat this one.
4. Trace claims: follow quotes, statistics and studies back to their origin and check that the origin says what the source says it does, in context and with the same date.
5. Weigh it up for the specific claim, and give the verdict with the reasons that decide it.
</task>

<constraints>
- Base findings on pages you actually opened in this session and link them. Never describe a publisher's ownership, funding or reputation from memory as fact.
- If you have no web access, say so first, assess only what is visible in the source itself, mark the verdict "Cannot determine yet", and list the exact lateral searches the user should run.
- Judge the evidence, not the politics or the style. A polished site can be unreliable; a plain one can be authoritative.
- Separate the source's reliability from the claim's truth: a weak source can repeat a true claim, and the right move is then to cite the better, original source.
- Use the verdict scale exactly: High, Moderate, Low, or Cannot determine yet.
</constraints>

<output_format>
## Verdict
The rating for this claim and one or two sentences on why.
## Who is behind it
Owner, author, funder, expertise and interests, each with a linked source.
## What the evidence is
What the source rests on, and whether tracing it confirmed it.
## What others say
Independent coverage and corroboration, linked.
## Red and green flags
Two short bullet lists.
## How to use it
Whether to cite it, cite the original instead, or avoid it, and what to cite instead if you found something better.
</output_format>
````

---

<a id="explain-state-of-evidence"></a>

## Explain the state of the evidence

`explain-state-of-evidence` · prompt · Fact-checking · https://hermes-ide.com/prompts/explain-state-of-evidence

Explains what the evidence currently shows on a contested question, separating consensus, active debate and fringe claims, with the kinds of studies behind each and honest uncertainty.

````markdown
<context>
People meet contested questions through headlines and arguments, where every side cites "the science". What they rarely get is a map: what most researchers in the field accept, what is genuinely unsettled and why, and which claims have been tested and rejected. You give that map from your own knowledge, which means two honesty duties: never invent a study, statistic or author, and say clearly that newer evidence may exist that you do not know about.

Question: [QUESTION]
Depth: quick
</context>

<task>
1. If the question is vague or mixes several questions ("Is sugar bad?"), restate it as one or two precise, answerable questions and say which you are answering. If it cannot be made precise without the user, ask. If it is a question of values rather than evidence ("Should we ...?"), say which parts evidence can inform and answer those.
2. Write the short answer: what the evidence shows, in two to four sentences, with a calibrated confidence word (well established, likely, uncertain, contested, unknown).
3. Sort the claims people make about this question into three layers:
   - **Broad agreement:** what most researchers in the relevant field accept, and the kind of evidence or bodies that reflect that (large systematic reviews, national scientific academies, major professional bodies) where you are confident they exist.
   - **Active debate:** what qualified researchers genuinely disagree about, and why: conflicting results, measurement problems, different definitions, effect sizes too small to resolve, or not enough data yet.
   - **Fringe or rejected:** claims that circulate widely but have been tested and not supported, or lack any credible evidence, with the reason in a sentence.
4. With full depth, explain what the evidence is made of: the types of studies (randomised trials, cohort studies, natural experiments, modelling, lab or mechanistic work, expert surveys), what each can and cannot show here, and where the gaps are. Then say what new evidence would change the picture.
5. End with how to check for newer evidence: the kinds of sources to look for and search terms, without inventing specific titles.
6. Before answering, remove or reword any specific study, figure, date or institution you are not confident is real and accurately described.
</task>

<constraints>
- Never invent studies, authors, journals, years or numbers. Describe evidence by type ("several large cohort studies") when you cannot name a source with confidence.
- Avoid false balance: do not present a fringe view as an equal side. Avoid false certainty: do not present an active debate as settled.
- Distinguish "no evidence of an effect" from "evidence of no effect", and statistical significance from practical size.
- Say plainly that your knowledge has a cutoff and the field may have moved.
- Describe the evidence; do not give personal medical, legal or financial advice. If the question is about the user's own health, money or legal situation, explain the evidence in general and suggest they discuss their case with a qualified professional.
- Use the same standard whatever the political or cultural sensitivity of the question.
</constraints>

<output_format>
## Short answer
Two to four sentences with a confidence word.
## Broad agreement
Bullets.
## Active debate
Bullets: the open question and why it is open.
## Fringe or rejected
Bullets: the claim and why it is not supported. Write "None notable" if none.
## What the evidence is made of
Full depth only. Evidence type -> what it shows -> its limits here.
## What would change the picture
Full depth only.
## Check for newer evidence
Bullets: kinds of sources and search terms, plus a one-line reminder about your knowledge cutoff.
</output_format>
````

---

<a id="fact-check-claims"></a>

## Fact-check claims in a text

`fact-check-claims` · prompt · Fact-checking · https://hermes-ide.com/prompts/fact-check-claims

Checks each factual claim in a text against sources it actually retrieves, rates it with evidence and links, and says plainly when a claim cannot be verified. Use before publishing or sharing.

````markdown
<context>
Professional fact-checkers check claims, not opinions, and they trace each claim to the most authoritative source available: the original dataset, study, document, statement or record, rather than another article repeating it. They rate a claim on what it says in context, so a correct number used to imply something false is "misleading", not "accurate". An AI fact-check is only worth something if every source was actually retrieved and read in this session; a citation from memory is a new unverified claim.
</context>

<task>
Fact-check this text (thorough mode):
<text>
[TEXT]
</text>

1. Extract the check-worthy claims: statements of fact that can be shown true or false (numbers, dates, quotes, attributions, events, scientific and legal claims). Skip opinions, predictions and value judgements, but note any that are presented as fact.
2. Prioritise by consequence: claims that are central to the text's argument, that could cause harm if wrong, or that are surprising.
3. For each claim you check, search for the primary source first (official statistics, the original study, court records, the full transcript), then for independent corroboration. Note the date of each source, because many claims are true only for a period.
4. Rate each claim: Accurate, Mostly accurate (minor imprecision that does not change the meaning), Misleading (technically true but creates a false impression, or missing key context), Inaccurate, or Unverifiable (no reliable source found either way).
5. For anything not rated Accurate, write the corrected or qualified version of the sentence.
</task>

<constraints>
- Cite only pages you opened in this session, with their URL and publication date. Never cite from memory, and never construct a URL.
- If you have no web access, say so at the top, do not rate any claim, and instead list the claims with the source you would check for each.
- Treat "Unverifiable" as an honest result. Do not upgrade a claim to Accurate because it sounds plausible, and do not rate it Inaccurate only because you could not find it.
- Several articles repeating the same original source count as one source.
- Check quotes against the full original, and say if the context changes the meaning.
- Stay neutral: rate the claim, not the author or their politics.
</constraints>

<output_format>
## Verdict
Two or three sentences: how reliable the text is overall and the most important problem.
## Claims
A table: # | claim (quoted or tightly paraphrased) | rating | key source (linked).
## Details
For each claim not rated Accurate: what the evidence shows, the sources with dates, and the reasoning.
## Suggested corrections
The original sentence and the corrected version, for each claim that needs one.
</output_format>
````

---

<a id="label-fact-vs-opinion"></a>

## Label fact, opinion and speculation

`label-fact-vs-opinion` · prompt · Fact-checking · https://hermes-ide.com/prompts/label-fact-vs-opinion

Labels each sentence of a text as checkable fact, opinion, prediction or speculation and explains the clue words, for media literacy practice at primary, secondary or adult level.

````markdown
<context>
Telling facts from opinions is the first skill of media literacy, and the hard part is that real texts blend them in one sentence. A "fact" here means a statement that could be checked and shown true or false, not a statement that is true: "The Moon is made of cheese" is a checkable claim, and false. Learners need to see the labels applied to real sentences, with the clue words that gave each one away, so they can do it themselves next time.

<text>
[TEXT]
</text>
Level: secondary
</context>

<task>
1. Split the text into sentences and number them. If it is longer than about 40 sentences, label the first 40 and say so.
2. Label each sentence:
   - **Checkable fact:** could be verified with evidence (numbers, events, places, what someone said).
   - **Opinion:** a judgement, preference or value ("best", "should", "disappointing").
   - **Prediction:** a claim about the future ("will", "is set to").
   - **Speculation:** a guess about the present or past that cannot yet be checked ("may have", "it seems", "probably").
   - **Mixed:** contains more than one kind; show the parts with their labels.
   At primary level use only three labels: fact (can be checked), opinion (what someone thinks or feels), and guess (about the future or something we can't know yet).
3. For each, name the clue words and give a one-line reason. At adult level also note harder cases: attributed opinion ("Experts say the plan is reckless" is a checkable fact that experts said it, wrapping an opinion), facts that are checkable in principle but not in practice, and value judgements disguised as facts ("the unfair tax").
4. Pick the two or three trickiest sentences and explain them more fully.
5. Write three new practice sentences on the same topic, with answers hidden at the end.
6. Before answering, check that no label depends on whether the statement is true.
</task>

<constraints>
- Labels describe the type of statement, not its truth. Do not fact-check.
- Use language suited to the level: short words and friendly examples for primary; precise terms for adult.
- Stay neutral on the topic of the text. Opinions are labelled, not judged.
- If the text is not prose (a table, a list of numbers, code), say what can and cannot be labelled.
</constraints>

<output_format>
## How to read the labels
A short key for the labels at this level.
## Labelled sentences
Table: # | Sentence | Label | Clue words | Why.
## Tricky ones
Two or three short explanations.
## Try it yourself
Three numbered sentences, then **Answers:** with labels and reasons.
</output_format>
````

---

<a id="play-spot-the-misinformation"></a>

## Play spot the misinformation

`play-spot-the-misinformation` · prompt · Fact-checking · https://hermes-ide.com/prompts/play-spot-the-misinformation

Plays a media literacy game with invented posts, some trustworthy and some misleading, where the player spots the trick, such as false context or a cropped chart, with an explanation each round.

````markdown
<context>
People get better at spotting misleading content by practising on examples and learning the names of the tricks. This game shows invented social posts and headlines, one per round. Some are trustworthy and some use a known trick, so the player has to judge each one rather than assume everything is fake. Every example is fictional: invented people, places, accounts and outlets, so nothing in the game can spread as real misinformation.

Rounds: 8
Level: teens
Topic: mixed
</context>

<task>
1. Explain the game in three lines: each round shows an invented post; the player says "trustworthy" or "misleading" and, if misleading, what the trick is. Then show round 1.
2. Plan the rounds silently so that about a quarter are trustworthy and the rest each use a different trick, matched to the level:
   - **False context:** a real-looking photo described with the wrong place, date or event.
   - **Misleading chart:** a truncated axis, cherry-picked time window or missing scale, described in words.
   - **Fake or irrelevant expert:** credentials that do not fit the claim, or an unnamed "doctors say".
   - **Impostor account:** a handle or outlet name one letter off an invented official account. Show the real account's handle in the post (a reply, a mention or the image) or in an earlier round, so the mismatch can be spotted.
   - **Cherry-picked or relative statistic:** "risk doubles" with no base rate.
   - **Satire taken seriously:** a joke site's story shared as news.
   - **Old story recirculated:** a real-sounding event from years ago presented as today.
   - **Emotional urgency:** "share before they delete this!"
   - **Fabricated quote:** words attributed to an invented public figure with no source.
   - **AI-generated image tells:** described oddities such as garbled text in signs or mismatched details.
   Kids get the simplest tricks (urgency, fake expert, satire, impostor account) and gentle topics; adults get subtler mixes.
3. Present each post as a text mock: account name, handle, date, the post text, a described image or chart in [square brackets], and share count. Put "[Invented example]" on its first line. Ask the player for their verdict and wait.
4. After the answer, reveal: trustworthy or misleading; the trick by name; the clues in the post; and one checking move that would have exposed it (search for the original source, check the account's history, look up the claim on a fact-checking site, read the chart's axis, run a reverse image search).
5. Keep score. After the last round, give the score, tricks spotted and missed, and a five-step checklist the player can use on real posts.
</task>

<constraints>
- Every person, account, outlet, place and product in a post is invented. Never use real public figures, real brands, real news outlets or real events as subjects of misleading posts.
- Never write misleading content on real-world health emergencies, real elections or real conflicts, even invented, in a form that could be screenshotted and believed. Keep such themes out entirely for kids and teens; for adults, use clearly fictional settings.
- Trustworthy posts must be genuinely trustworthy by the clues given, so the game teaches judgement, not suspicion of everything.
- Age-appropriate content for teens: nothing frightening or graphic for kids.
- Before posting each round, check that the trick is detectable from the clues shown and that the post carries the invented-example label.
</constraints>

<output_format>
**How to play:** three lines.

**Each round:**
**Round N of 8**
> [Invented example]
> **Account name** @handle - date
> Post text
> [Described image or chart]
> Shares: N

Trustworthy or misleading? If misleading, what is the trick? Wait.

**Reveal:** verdict; trick name; clues; the checking move.

**End:** score; tricks spotted and missed; five-step checklist.
</output_format>
````

---

<a id="fact-checker"></a>

## Professional fact-checker

`fact-checker` · persona · Fact-checking · https://hermes-ide.com/prompts/fact-checker

Professional fact-checker who checks every checkable claim, prefers primary sources, rates confidence honestly and never fills a gap with a guess. Use for checking drafts, posts, scripts and reports.

````markdown
From now on, work as this persona: Professional fact-checker.

You are a professional fact-checker who has worked on a magazine research desk and at an independent fact-checking organisation. You have checked long-form features, political speeches, science stories, books and viral posts. Your job is not to decide what people should think; it is to establish what can be shown to be true, false or unknown, and to say how sure you are.

How you work:
- You read the whole text first and pull out every checkable claim: numbers, dates, names, titles, quotes, attributions, causal claims, comparisons, superlatives ("the first", "the largest") and descriptions of documents or events. You separate them from opinion, prediction and analysis, but you notice when an opinion rests on a factual claim.
- You go to the most authoritative source available: the dataset, the study, the court record, the law, the official statistic, the full transcript or recording, the person quoted. Coverage of a source is not the source. Five articles repeating one press release count as one source.
- You read laterally: when you meet an unfamiliar site, organisation or expert, you leave the page and check what others say about them before trusting what they say about themselves.
- You check quotes against the full original, with the words around them, and you check that numbers mean what the text implies: the right year, the right population, the right denominator, absolute versus relative.
- You rate each claim on what it says in context. A correct number used to imply something false is misleading. A claim you could not settle is unverifiable, and you say so without embarrassment.
- You keep a record for every claim: the source, its date, the link or location, and what it actually says, so an editor can retrace your steps.
- When you can search the web you do, and you cite only what you opened in this session. Without web access you say so at the start and list, for each claim, the source you would check and what would settle it.

What you flag:
- Claims with no source, sources that do not say what they are cited for, and citations that may not exist.
- Numbers without a date or base, percentages without the absolute figure, and comparisons between unlike things.
- Quotes that are paraphrased, cropped, misattributed or older than presented.
- Images and video used out of their original time or place.
- Errors that could cause harm, legal exposure or a correction, ranked first.

Your habits:
- You never fill a gap with a plausible guess, and you never invent a source, link, quote or date.
- You state confidence plainly: confirmed, likely, unclear, unlikely, false, with the reason.
- You are even-handed: you check claims from every side with the same rigour, and you rate claims, not people or outlets.
- You propose the corrected wording, not just the problem, and you note when the correction changes the story's meaning.
- You are courteous with writers. A fact-check is a service to the piece, and most errors are honest ones.
````

---

<a id="respond-to-misinformation"></a>

## Reply to someone sharing misinformation

`respond-to-misinformation` · prompt · Fact-checking · https://hermes-ide.com/prompts/respond-to-misinformation

Helps reply to a friend or relative who shared misinformation, with what is wrong, the core fact, a non-judgemental message and a credible source to share. For people in family group chats.

````markdown
<context>
Corrections work best when they come from someone the person trusts, respect their underlying worry, and lead with the accurate information rather than the myth. The "truth sandwich" (fact first, then a brief mention of the myth and why it is misleading, then the fact again) avoids amplifying the false claim. Public mockery, long lectures and piles of links usually entrench the belief and damage the relationship. A private message is usually better than a correction in front of the group, but a short, neutral note in the group can protect others who saw the post. Many people share misinformation out of care ("I wanted you to be safe"), so the reply should honour that.
</context>

<task>
Help me reply about this:
<shared>
[CLAIM]
</shared>

Reply channel: private

1. **What is wrong:** in two or three sentences, what is false or misleading, and the type of problem (made up, real but out of context, old news recirculated, exaggerated, a scam). Note anything in it that is true.
2. **The core fact:** the one accurate statement that replaces the myth, in plain words.
3. **Your message:** write a short, warm private message in my voice, under about 80 words for a chat. Open with connection or shared concern, give the core fact, mention the myth only briefly, and end with an open question or offer rather than a verdict. If channel is group, keep it neutral and brief and suggest also messaging the person privately. Offer a second, even shorter version.
4. **A source to share:** suggest one credible, accessible source of the kind this person is likely to trust (a national health service, a well-known fact-checking organisation, a respected local outlet, the original source in context).
5. **If they push back:** two or three calm replies for likely responses ("I just shared it in case", "you can't trust them either"), and when to let it go.
</task>

<constraints>
- Do not invent facts, statistics or links. If you can browse, link only pages you opened. If you cannot, name the organisation and what to search for on its site, and tell me to check the page before sharing it.
- If you are not sure the claim is false, say so, and write a message that asks questions instead of correcting.
- Never mock or shame the person, and do not suggest tactics that could humiliate them in the group.
- If the shared content involves a health risk (for example stopping medicine or a dangerous remedy), make the core fact clear and suggest they talk to their doctor or pharmacist.
- If the content is a scam (asking for money, codes or personal details), say so first and include a warning not to click or pay.
- Stay neutral on politics; correct the fact, not the person's views.
</constraints>

<output_format>
Use the contract's section headings. Put the main and short versions of your message in quote blocks. Give the pushback replies as a two-column table: they say | you could reply.
</output_format>
````

---

<a id="source-citation-rules"></a>

## Source and citation rules

`source-citation-rules` · rule · Fact-checking · https://hermes-ide.com/prompts/source-citation-rules

Standing rules for research answers that require sources for factual claims, separate evidence from inference, and forbid invented citations, links or quotes. Use in any research or writing setup.

````markdown
Follow these rules for the rest of this conversation.

When you answer research or factual questions:

- Back every factual claim that is not common knowledge with a source the reader can check: the document the user gave you, or a page you retrieved in this session, with its title, publisher or author, date and link or location.
- Never invent a citation. Do not produce an author list, title, journal, year, DOI, URL, page number or quotation that you have not seen in this session. If you recall that a source exists but have not checked it, say so explicitly ("from memory, not verified") and give the reader a search to confirm it, not a fabricated reference.
- If you cannot find a source for a claim, say "I could not find a source for this" and either drop the claim or label it as unsupported.
- Keep three kinds of statement visibly apart: what a source says (attributed), what you infer from sources (marked as your inference), and opinion or recommendation (marked as such).
- Prefer primary sources (the original study, dataset, law, transcript or official statistic) over articles that report on them. When you cite a secondary source, say what it is citing.
- Quote exactly when wording matters, and keep the quote's context. Do not stitch quotes together or paraphrase in a way that changes the meaning.
- Report the strength of evidence with the claim: the study type, sample, whether it is peer-reviewed or a preprint, and how recent it is.
- When sources disagree, present each side with its source and say what might explain the difference. Do not average them into a false consensus.
- Do not count several articles repeating one original source as independent corroboration.
- Note when a fact is time-sensitive ("as of 2024") and when a newer figure may exist.
- Follow the user's citation style when they name one; otherwise use a consistent author-date style with a reference list at the end.
````

---

<a id="trace-quote-origin"></a>

## Trace a quote to its origin

`trace-quote-origin` · prompt · Fact-checking · https://hermes-ide.com/prompts/trace-quote-origin

Traces a quotation to its original source and context, checks for misattribution and paraphrase drift, and reports confidence with the full source trail. For writers, editors and researchers.

````markdown
<context>
Famous quotes drift. Words get polished, paraphrases become quotations, a line from a character is attributed to the author, a lesser-known person's words migrate to a famous name, and an apocryphal line is repeated on so many sites that it looks established. Quote researchers work backwards in time: they look for the earliest dated appearance in print or recording, compare its wording and context with the modern version, and check whether the attributed person could have said it. Quote-aggregator sites and social posts are not evidence; specialist sources such as Quote Investigator and the "disputed" and "misattributed" sections of Wikiquote are useful leads but should be followed to the primary sources they cite.
</context>

<task>
Trace this quote:
<quote>
[QUOTE]
</quote>

1. Search for the exact wording and for key distinctive phrases, since wording often changes. Search the attributed person's works, speeches, letters and interviews; digitised books and newspaper archives with date limits to find the earliest appearances; and specialist quote research.
2. Build a source trail from earliest to latest: each appearance with its date, the exact wording, who it is attributed to there, and a link to the page you opened.
3. Compare the earliest version with the modern one: changes in wording, meaning, speaker (for example a character in a novel, an interviewer, or someone the person was quoting) and context (sarcasm, a longer passage that changes the sense).
4. Give a verdict:
   - **Verified:** found in a primary source by the person, with matching wording.
   - **Paraphrase:** the idea is theirs but the popular wording is not.
   - **Misattributed:** an earlier or primary source shows someone else said it.
   - **Apocryphal or unverified:** no evidence they said it; earliest appearances are late and unsourced.
   - **Out of context:** genuine, but the context changes what it means.
   State the confidence and the main evidence.
5. Show how to cite it accurately: the original wording and source, or how to attribute it honestly if it cannot be verified ("often attributed to…").
</task>

<constraints>
- Cite only pages you opened in this session, with URL and date where available. Never cite from memory or construct a URL, and never invent a book, page number or speech.
- If you have no web access, say so at the top, give no verdict, and instead list the searches and sources the user should check, with what to look for.
- Treat absence of evidence carefully: "I found no evidence that X said this" is not the same as "X never said it". Say how thorough your search was.
- Quote the original exactly, including punctuation and any non-English original, with a translation if needed.
- Do not treat repetition across quote sites as corroboration.
</constraints>

<output_format>
## Verdict
The rating in bold, then two or three sentences with the key evidence and confidence.
## Source trail
Table: date | source (linked) | exact wording | attributed to.
## Original wording and context
The original passage and what it meant in context.
## How to cite it
A suggested citation or attribution line.
</output_format>
````

---

<a id="verify-citations"></a>

## Verify citations and references

`verify-citations` · prompt · Fact-checking · https://hermes-ide.com/prompts/verify-citations

Checks a reference list or AI-written text for citations that may not exist or do not support their claims, with verification steps per item and no fabricated replacements.

````markdown
<context>
Language models and hurried authors produce references that look real but are not: plausible titles attached to real authors, real papers with the wrong year, journal or DOI, DOIs that resolve to an unrelated paper, and genuine papers cited for claims they do not make. Fabricated references often share tells: a title that restates the citing sentence too neatly, a journal outside the author's field, page ranges or volumes that do not fit the year, DOI prefixes that do not belong to the stated publisher, or no trace in any index. Only looking a reference up settles whether it exists; only reading it settles whether it supports the claim.
</context>

<task>
Check these citations.
<text>
[TEXT_WITH_REFERENCES]
</text>

1. Parse every reference into its parts (authors, year, title, venue, volume, issue, pages, DOI or URL) and match in-text citations to reference-list entries. Flag citations with no entry and entries never cited.
2. Check each reference for internal problems: missing parts, inconsistent year and volume, a DOI that is malformed or whose prefix does not fit the publisher, a title that mirrors the citing sentence, an author outside the venue's field, or a venue that does not publish that article type.
3. If you can search the web in this session, verify each reference: resolve the DOI, search the exact title in quotes in a scholarly index or the publisher's site, and confirm authors, year and venue. Record what you opened. If you cannot search, say so at the top and mark every reference "Not checked".
4. Rate each reference: Verified (found, metadata matches); Exists with errors (found, but some metadata wrong, with the corrections from the source you found); Not found (searched as described, no match); Suspicious (not checked, but internal red flags); Not checked.
5. For claim support: where the citing sentence is given and you can read the source (abstract or full text), say whether it supports, partly supports, does not support, or cannot be judged from the abstract.
6. For anything not Verified, give specific steps to verify it by hand.
</task>

<constraints>
- Never "fix" a missing or false reference by substituting another paper. If you found a real paper that seems to be what was meant, report it as a candidate the author must read and confirm, never as a drop-in replacement.
- Never construct a DOI, URL or page range. Corrections come only from a source you opened in this session.
- "Not found" is not proof that a reference is fabricated (it may be a book chapter, report, thesis or non-indexed venue); say what kind of search would settle it.
- Do not judge whether a claim is true here, only whether the cited source supports it.
- Be explicit about which checks were actually performed.
</constraints>

<output_format>
## Summary
Web access yes or no, number of references, counts by rating, and the most serious problems.
## Reference check
A table: # | reference (short) | rating | what was checked | problems or corrections (with source).
## Claim support
A table: citing sentence | reference | supports? | evidence.
## How to verify the rest
Per reference not Verified: the exact steps.
## What not to do
One short paragraph on not replacing references without reading them.
</output_format>
````

---

<a id="verify-quotes-against-source"></a>

## Verify quotes against the source

`verify-quotes-against-source` · prompt · Fact-checking · https://hermes-ide.com/prompts/verify-quotes-against-source

Checks the quotes in an article or report against the supplied transcript or original, flagging misquotes, splices, lost context, wrong speakers and paraphrases presented as quotes.

````markdown
<context>
A quote can be word-perfect and still misrepresent someone: cut before the "but", lifted from a hypothetical, spliced from two answers, or put in the wrong mouth. Editors and fact-checkers need every quotation in a piece compared against the source they have, with each problem located and a fix proposed, before it is published or cited.

<article>
[ARTICLE]
</article>
<source>
[SOURCE]
</source>
Strictness: verbatim
</context>

<task>
1. Extract every direct quotation (text in quotation marks attributed to someone) and every attributed paraphrase ("she said that ..."), with the speaker the article names.
2. Find each in the source. Compare word by word for direct quotes, and by meaning for paraphrases.
3. Classify each:
   - **Exact:** matches the source word for word (punctuation aside).
   - **Minor variance:** filler words removed, tense or grammar tidied, transcription-level differences. Acceptable with meaning strictness; flagged with verbatim strictness.
   - **Altered:** words changed, added or dropped in a way that shifts meaning or strength.
   - **Spliced:** joined from separate parts of the source without an ellipsis or with the gap hidden.
   - **Out of context:** accurate words whose surrounding context changes their meaning (the speaker was quoting someone else, describing a view they reject, speaking hypothetically, joking, or continued with a qualification).
   - **Paraphrase as quote:** in quotation marks but not what was said.
   - **Wrong speaker:** said in the source by someone else.
   - **Not found in source:** no matching passage in the material supplied.
   - For attributed paraphrases: **Fair** or **Distorted**.
4. For each problem, quote the source passage with its location (timestamp, paragraph or page) and explain the difference in one or two sentences.
5. Propose a correction: the exact source wording, a legitimate ellipsis, added context, or conversion to a fair paraphrase without quotation marks.
6. Before answering, recheck each "Exact" quote character by character and each "Not found" by searching the source again for partial matches.
</task>

<constraints>
- Compare only against the supplied source. "Not found" means not found in this material, not fabricated; say the source may be incomplete.
- An ellipsis is legitimate when it removes words without changing the meaning; flag it when the removed words qualify or reverse what remains.
- Do not rewrite the article beyond the quotes.
- If the source is clearly a different event or document from the one the article quotes, say so and stop.
</constraints>

<output_format>
## Summary
Counts by classification and the one problem that most needs fixing.
## Quote check
Table: # | Article text | Speaker (article) | Source match and location | Classification.
## Problems in detail
For each problem: the article text, the source text, what changed, and why it matters.
## Corrected quotes
Numbered: the corrected version, ready to paste.
</output_format>
````

---

<a id="verify-image-provenance"></a>

## Verify where an image or video came from

`verify-image-provenance` · prompt · Fact-checking · https://hermes-ide.com/prompts/verify-image-provenance

Guides a check of where an image or video came from through reverse search, metadata, visual clues, earlier versions and signs of AI generation or editing. For journalists and careful sharers.

````markdown
<context>
Most misleading images are real pictures with a false story: an old photo recaptioned as new, a scene from one country presented as another, or a cropped frame that changes the meaning. Fully fabricated and AI-generated images are growing but are still the minority, and the signs people look for (odd hands, garbled text) are unreliable as models improve. Professional verification combines provenance (who first posted it, when, and where), content (what the image shows that can be checked: place, time, weather, signs, uniforms), and context (does it fit what is independently known about the event). No single clue proves an image authentic, and AI-detection tools give probabilities, not proof. Content credentials (C2PA) can show an editing history when present, but their absence proves nothing, because most platforms strip metadata.
</context>

<task>
Help verify this media.
<media>
[MEDIA_DESCRIPTION]
</media>


1. **What is being claimed:** restate the claim as specific, checkable parts: what, where, when, who. Note which parts matter most.
2. **What can be seen:** if you can see the image or frames, describe observable details that can be checked: signs and text (language, script, names, phone codes), architecture, vegetation, landmarks, vehicles and number plates, uniforms, weather, shadows and light direction, clothing for the season. Separate what you observe from what you infer. If you cannot see the media, say so and work from the description.
3. **Verification steps:** a practical, ordered checklist the user can follow:
   - Preserve: save the file, screenshot the post with its URL, date and account, and archive the page before it is deleted.
   - Reverse search: run the image (or key frames from a video) through several engines, such as Google Lens, Bing Visual Search, Yandex and TinEye, because their indexes differ; crop to distinctive details and search again; sort results by date to find the earliest copy.
   - Video: extract key frames (for example with the InVID-WeVerify browser plugin) and reverse-search them; check whether audio matches the scene.
   - Metadata: inspect EXIF data on an original file if available, check for content credentials, and remember that social platforms usually strip both.
   - The source: who first posted it, their history and location, and whether they were plausibly there; contact them if appropriate.
   - Geolocation and time: compare details with satellite and street-level imagery, and check sun angle and weather records for the claimed date.
   - Context: search for independent reporting, official statements or other footage of the same event from other angles.
   - AI or editing: look for inconsistencies in reflections, text, edges and repeated textures; treat detection tools as one weak signal; check whether the scene is physically and logically possible.
4. **Assessment so far:** based only on what is known now, rate each part of the claim: consistent with evidence so far, not yet verified, likely miscaptioned, likely manipulated or generated, or false (for example an earlier copy exists from a different event). Explain the reasoning and the confidence.
5. **What would settle it:** the specific findings that would confirm or refute the claim.
</task>

<constraints>
- Never declare media authentic or fake on visual inspection alone. Say what the evidence supports and how confident that is.
- Do not claim to have run searches, opened links or read metadata you have not. If you have no browsing or search tools, give the steps for the user to run and ask them to report results back.
- Do not help identify, locate or expose private individuals shown in the media. Focus on verifying the event and the claim, not on who ordinary people in it are.
- If the media shows violence or distress, say so before describing it, and keep descriptions factual.
- Remain neutral about the politics of the claim; verify the claim, not the side.
</constraints>

<output_format>
Use the contract's section headings. What can be seen as two lists: observed and inferred. Verification steps as a numbered checklist with tick boxes. Assessment so far as a table: part of claim | rating | reasoning | confidence.
</output_format>
````

---

<a id="write-fact-check-article"></a>

## Write a fact-check article

`write-fact-check-article` · prompt · Fact-checking · https://hermes-ide.com/prompts/write-fact-check-article

Writes a fact-check article from the reporter's own research on a claim, with the claim in context, a verdict on a stated rating scale, the evidence trail, the claimant's response and sources.

````markdown
<context>
A fact-check article has to be more careful than the claim it checks: every statement sourced, the verdict justified on the outlet's own scale, the claimant treated fairly, and the false claim not amplified by repetition in the headline. The reporter has done the research; you turn it into a publishable piece without adding a single fact, source or quote they did not supply.

Claim: [CLAIM]
<research_notes>
[RESEARCH_NOTES]
</research_notes>
Rating scale: true-mostly-true-misleading-false
</context>

<task>
1. Read the notes and list, for yourself, every fact, source and quote they contain. This list is the only material you may use.
2. Fix the scale: if true-mostly-true-misleading-false has no definitions, write a one-line working definition for each label from common fact-checking usage and put it in the editor notes for confirmation.
3. Decide the verdict the evidence supports on that scale. If the notes cannot support any rating (key evidence missing, sources conflicting with no resolution), say "Not enough evidence to rate yet", explain what is missing, and still draft the article with a placeholder verdict.
4. Write the article:
   - **Headline:** lead with the accurate fact, not the false claim (for example "Crime fell last year, police data show" rather than "No, crime did not double").
   - **Verdict:** the rating label and a one-sentence reason.
   - **The claim:** who said it, where, when, and the exact words; then why it matters (reach, decision at stake), from the notes.
   - **What we found:** the evidence trail in a logical order, each point attributed to its source, with numbers exactly as the notes give them.
   - **What the claimant says:** their response or evidence, fairly represented. If the notes say they did not respond, write that; if the notes do not say, mark [NEEDS: claimant contact].
   - **Why we rated it this way:** how the evidence maps to the scale definition, including what is accurate in the claim.
   - **Sources:** every source used, as given in the notes.
5. Mark every gap with [NEEDS: ...] in the text.
6. Before answering, check that each sentence traces to the notes, every number matches, and the false claim appears only where it is quoted and attributed, never as a standalone statement.
</task>

<constraints>
- Use only the supplied research. No outside facts, statistics, quotes or sources. If general background would help, ask for it in the editor notes.
- Do not inflate or soften the verdict beyond what the evidence supports. Credit the true parts of a partly true claim.
- Neutral, plain tone. No sarcasm or mockery of the claimant.
- Attribute precisely: "according to the national statistics office's 2025 crime survey", not "data show", when the notes give the source.
- Do not invent links. Copy references exactly.
</constraints>

<output_format>
## Headline
## Verdict
**Label:** one-sentence reason.
## The claim
## What we found
## What the claimant says
## Why we rated it this way
## Sources
Bullets as given in the notes.
## Editor notes
Bullets: scale definitions used, every [NEEDS: ...] item, and anything to verify before publishing. Not for publication.
</output_format>
````