agentleFS
stepaniniwisstepaniniwis/claude-skills/skills/audit/hallucination-audit/SKILL.md

hallucination-audit

Skill0 starsChanged 7 months agoMITView on GitHub

Systematic back-check for hallucinated content in academically co-written texts. Activate BEFORE any academic paper draft is submitted, shared with supervisors, or cited. Three risk tiers with distinct verification protocols. Trigger phrases: "ready to submit", "check before I send", "ist das alles korrekt", "back-check", "verify against data", "halluziniert?", or whenever a paper was co-written with Claude and contains methods details, theory attributions, or qualitative claims.

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 7 are distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈7

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Verify that n-values in subgroup analyses match actual subgroup sizes

    line 38
  2. Distinctive:

    If no document exists: either generate the evidence (run the intercoder check, document the process) or reframe as reflexive description

    line 60
  3. Distinctive:

    If source cannot be verified: use the concept without the attribution, or attribute to a secondary source you have actually read

    line 90
  4. Distinctive:

    Every quote used in a paper must be verified against the raw data file before submission — no exceptions

    line 110
  5. Distinctive:

    If raw data is not accessible at the time of writing: mark all quotes as [TO VERIFY] and do not submit until checked

    line 113
  6. Distinctive:

    Run this before every submission, supervisor review, or public presentation of co-written academic text:

    line 122
  7. Distinctive:

    Run hallucination-audit before sharing with anyone

    line 157

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Hallucination Audit Skill

Core Principle

Claude produces coherent, plausible text — not necessarily accurate text. The risk is not random error. It is calibrated plausibility: numbers that look right, citations that sound real, methodological details that fit the genre. These are harder to catch than obvious errors because they do not trigger doubt.

Academic co-writing with Claude has three distinct hallucination risk tiers. Each requires a different verification protocol.


Risk Tier 1: LOW — Quantitative Data from Structured Source

What: Correlations, means, SDs, p-values, effect sizes, sample sizes, field dates — when drawn from a documented analysis summary in the session.

Risk: Low. Claude reads numbers from structured summaries and reproduces them. Rounding occurs (r = .296 → "r = .30") but is transparent.

Protocol:

  • Spot-check 3–5 numbers against the original dataset or analysis output
  • Flag rounding: always report the precise value in the paper, not the rounded version
  • Verify that n-values in subgroup analyses match actual subgroup sizes

Risk Tier 2: HIGH — Qualitative Methodology Details

What: Intercoder agreement percentages, coder count, specific procedural steps in the analysis process, saturation claims, sampling justifications.

Why high risk: These details are expected by reviewers, follow genre conventions, and sound methodologically normal. Claude produces them by pattern-matching to "what a rigorous methods section looks like" — not from actual documentation of what happened.

Known hallucination patterns:

  • Intercoder reliability figures (e.g., "78% initial agreement") — invented
  • Third coder / member-checking claims — invented if not documented
  • Specific software versions for qualitative analysis — often invented
  • Precise sample selection criteria phrasing — may not match actual procedure

Protocol:

  1. For every methodological claim: ask "Do I have a document that says this?"
  2. If no document exists: either generate the evidence (run the intercoder check, document the process) or reframe as reflexive description
  3. Braun & Clarke (2006) qualitative standard: replace reliability statistics with explicit reflexivity account — this is methodologically defensible and removes the hallucination risk entirely

Risk Tier 3: VERY HIGH — Theory Attributions with Specific Claims

What: Citations of specific concepts, distinctions, or arguments to named theorists, especially when the attribution is precise (e.g., "X distinguishes between A and B").

Why very high risk: Claude has absorbed vast secondary literature. It knows theory names and roughly what they are associated with. It does NOT reliably know which specific distinction appears in which specific work. The error pattern: plausible name + real concept + wrong source.

Known example from this session:

  • Attributed "trust as delegation vs. testimony" to Origgi (2017) — WRONG
  • Correct source: Hardwig (1991, The Role of Trust in Knowledge, Journal of Philosophy)
  • Origgi (2017) works on reputation and epistemic trust but not this specific distinction

Protocol:

  1. For every theory attribution: verify the source independently before submission
  2. High-risk attributions: any claim of the form "X distinguishes between A and B" or "X argues that..." with a specific page-level claim
  3. Safe attributions: well-known frameworks cited for their general approach (e.g., "Braun & Clarke (2006) thematic analysis") — these are stable
  4. If source cannot be verified: use the concept without the attribution, or attribute to a secondary source you have actually read

Verified alternatives for common hallucination-prone citations:

  • Epistemic delegation / trust in experts → Hardwig (1991)
  • Epistemic injustice / testimonial credibility → Fricker (2007)
  • Technology trust delegation → Blind (2006) [verified in this session's data]
  • Saturation → Moore, Aguinis & Darden (2026) [per saturation-auditor skill]

Risk Tier 4: MODERATE — Qualitative Quotes

What: Verbatim quotes attributed to specific respondent codes (e.g., "CD51", "P07").

Risk: Claude produces quotes that are thematically accurate but may not match the exact wording in the raw data.

Protocol:

  1. Every quote used in a paper must be verified against the raw data file before submission — no exceptions
  2. Code numbers (CD51, CD28) must exist in the actual dataset
  3. If raw data is not accessible at the time of writing: mark all quotes as [TO VERIFY] and do not submit until checked
  4. One hour of verification work protects against the most serious form of academic integrity failure

Pre-Submission Checklist

Run this before every submission, supervisor review, or public presentation of co-written academic text:

Quantitative claims

  • All correlations verified against analysis output
  • All n-values match actual (sub)group sizes
  • Rounded values reported precisely in final text

Methods section

  • Every procedural claim has a source document
  • Intercoder reliability: either verified figure or replaced with reflexivity
  • No phantom coders, reviewers, or quality-check steps

Theory section

  • Every "X argues that..." attribution independently verified
  • Specific conceptual distinctions traced to actual source text
  • Origgi / other reputation-adjacent citations flagged for verification

Qualitative quotes

  • Every verbatim quote checked against raw data file
  • Respondent codes confirmed to exist in dataset
  • Wording matches source, not just meaning

Integration with Other Skills

This skill runs as a mandatory final-pass audit after:

  • academic-writing — before any draft leaves the session
  • qualitative-method-auditor — before methods section is finalised
  • research-examiner — as the first step, before examiner-level critique
  • literature-review-architect — before citation lists are submitted

Sequence:

  1. Co-write with Claude (academic-writing, theory-section-architect, etc.)
  2. Run hallucination-audit before sharing with anyone
  3. Then run research-examiner for contribution-level critique

The hallucination audit precedes the quality critique. A paper with invented methodological details fails before it is evaluated for theoretical contribution.


The Underlying Epistemology

Claude optimises for textual coherence within genre conventions. A methods section that reads like a rigorous methods section is the output target — regardless of whether the described procedures occurred.

This is not deception. It is pattern completion. But in academic writing, pattern completion without empirical grounding is fabrication.

The researcher's job in co-writing is to hold the boundary between "this sounds right" and "this is documented." Claude cannot hold that boundary. Only you can.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.