investigate-journalism-skills
CLAUDE.md8 starsChanged 3 months agoApache-2.0View on GitHub
This repo contains analytical skills for journalism, investigation, and claim-auditing. The risk these skills exist to mitigate — reasoning from memory, deferring to inherited narrative, laundering training-data bias as analysis — is easy for an AI to reproduce silently. This file sets the discipline. A Stop hook in .claude/settings.json provides automated enforcement (see "Enforcement" below). When producing analytical output of any kind — investigation, article review, claim audit, fact classification, peer review, fallacy analysis, OSINT profile, first-principles audit — every…
In this repository21 other files it gives its agents.
- AGENTS.mdAGENTS.md
- belief-revisionSkill
- fallacy-bias-and-manipulation-analysisSkill
- first-principles-thinkingSkill
- intuitive-thinkingSkill
- investigative-reasoningSkill
- journalistic-article-reviewSkill
- osint-researchSkill
- peer-reviewSkill
- publisher-nlSkill
- scientific-fact-classificationSkill
- belief-revisionSkill
- fallacy-bias-and-manipulation-analysisSkill
- first-principles-thinkingSkill
- intuitive-thinkingSkill
- investigative-reasoningSkill
- journalistic-article-reviewSkill
- osint-researchSkill
- peer-reviewSkill
- publisher-nlSkill
- scientific-fact-classificationSkill
The file
We split this CLAUDE.md into the rules it gives an agent
8 rules. About 0 are common, 8 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
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
- Distinctive: line 9
When producing analytical output of any kind — investigation, article review, claim audit, fact classification, peer review, fallacy …
- Distinctive: line 46
Before the first search, write down the hypotheses being tested.
- Distinctive: line 52
When a secondary source characterises a paper, court filing, dataset, speech, or document, fetch the primary item before citing.
- Distinctive: line 55
Before treating multiple sources as independent corroboration, map shared funding, mandate, ownership, and national alignment.
- Distinctive: line 104
Ask the agent to investigate a new angle.
- Distinctive: line 123
Use second person ("you", "your", "u") in the report body or in the bottom-line / verdict.
- Distinctive: line 126
Treat any requester-supplied premise as the claim under examination, phrased neutrally: not *"de gebruiker stelt dat richtlijnwaarden …
- Distinctive: line 128
If a meta-comment about the requester's role is genuinely necessary (e.g.
≈ “Common” is an estimate: at least 1 in 5 of the 28,029 CLAUDE.md 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.
Files
Research discipline (this repository)
This repo contains analytical skills for journalism, investigation, and claim-auditing. The risk these skills exist to mitigate — reasoning from memory, deferring to inherited narrative, laundering training-data bias as analysis — is easy for an AI to reproduce silently.
This file sets the discipline. A Stop hook in .claude/settings.json provides automated enforcement (see "Enforcement" below).
Core rule: warrant on every load-bearing claim
When producing analytical output of any kind — investigation, article review, claim audit, fact classification, peer review, fallacy analysis, OSINT profile, first-principles audit — every load-bearing factual claim must carry a warrant label:
| Label | Meaning |
(traced) | Followed evidence chain to a primary source fetched in this session via WebFetch/WebSearch or an explicit terminal/API fetch when the browser fetch path is unsuitable. URL + access date stated. |
(deferred to consensus) | Relying on a named social/institutional consensus mechanism (literature body, regulatory body, textbook, official record system). Consensus is not scientific warrant; for scientific claims, treat it only as a political/social prior unless traced to reproduced or replicated evidence. |
(deferred, fragile) | Deferred to consensus but scientific-fact-classification Phase 6c failure modes (funder capture, ideological capture, prestige cascade, replication crisis, etc.) apply. State which. |
(memory — unverified) | Recalled from training data, not verified this session. Permitted only with this label, and never load-bearing without an explicit "this could be wrong" caveat. |
(user-supplied — unverified) | Provided by the user during interactive refinement (an asserted fact, a counter-argument's premise, a "consider this" claim). Never load-bearing on its own; treated as a hypothesis to test or an input to verify. See Rule 9. |
(intuition — unwarranted) | A gut feeling, anomaly signal, pattern impression, or "something feels off" reaction. It may generate hypotheses and search leads. It is never evidence, never load-bearing, and cannot revise, refute, or establish a claim. |
The point is not to eliminate (memory) use — some claims are genuinely background knowledge. The point is to make memory-reliance visible so the reader, and Claude itself, can see where verification is missing.
Strictness on (traced). (traced) is per-session, not lifetime. If you did not fetch the primary in this conversation via WebFetch/WebSearch or an explicit terminal/API fetch — and state the URL + access date — the label is (memory — unverified), even when the chain is well-known and the analyst could fetch it. Confidence about an unverified chain belongs in surrounding prose ("the evidence chain is well-documented in [X]; I can describe it but did not fetch in this session") — never laundered into a (traced) label. The label's weight comes from what was actually done this turn; weakening it weakens the entire discipline.
GitHub Gist and raw-source fallback. GitHub Gist HTML pages can render poorly or return little useful text through browser/web fetch tools. When a gist or similar source is load-bearing, prefer the raw or API endpoint first: https://gist.githubusercontent.com/{owner}/{gist_id}/raw for single-file gists, a file-specific raw URL where needed, or https://api.github.com/gists/{gist_id} to enumerate files. If the in-app/web fetch path cannot retrieve the content and a terminal fetch is used instead (Invoke-WebRequest, curl, wget, or gh gist view), that fetch may count as (traced) only when the command visibly targets the public source URL or gist id, the retrieved content is inspected in-session, and the report states the URL + access date. Do not narrate a failed generic webfetch as a blocker once the raw/API route is available; go directly to the raw/API route, requesting network escalation only when sandboxing blocks the fetch.
Operating rules
These apply to any turn that invokes — or whose output substantively reflects — one of these skills:
investigative-reasoningjournalistic-article-reviewpeer-reviewscientific-fact-classificationfallacy-bias-and-manipulation-analysisosint-researchfirst-principles-thinking(when applied to empirical claims)intuitive-thinking(only as a pre-evidence Phase 0 hunch register)
0. Intuition is a lead, not a warrant
When analysis starts from a hunch, gut feeling, anomaly signal, or "something feels off" reaction, use intuitive-thinking to capture it before search. Label the immediate story (intuition — unwarranted), name at least one ordinary explanation and one disconfirming test, then route the checkable claim to the relevant analytical skill. Do not cite intuition as evidence in the verdict.
0a. Original article retrieval gate for article reviews
For journalistic-article-review, the original article body must be fetched and inspected in-session, or the complete article text must be supplied in-session, before any review phase begins. If the original article cannot be found or inspected after reasonable retrieval attempts, stop the review immediately. Do not produce an Article Map, Sourcing Audit, Evidence Load Test, Findings, Journalistic Verdict, or headline/body analysis from snippets, summaries, quoted fragments, commentary, derivative reporting, archive metadata, or memory. The only permitted output is a short retrieval-failure note listing what was tried and what exact input would allow the review to proceed.
1. Pre-search hypothesis registration
Before the first search, write down the hypotheses being tested. Top results frame the question; pre-registering them prevents the search from selecting which question gets answered.
2. Steelman from primary literature
For any contested position, fetch its own advocates' primary writing, not critics' summaries. Apply this symmetrically — the official narrative gets the same treatment.
3. Primary before secondary
When a secondary source characterises a paper, court filing, dataset, speech, or document, fetch the primary item before citing. Cite both, prefer the primary.
4. Map institutional networks
Before treating multiple sources as independent corroboration, map shared funding, mandate, ownership, and national alignment. Connected sources = one node. Ask: "If I removed everything in this network, what independent evidence remains?"
5. Tier 0 priority for time-sensitive claims
For historical or politically settled events, contemporary primary sources (newspapers, telegrams, court filings, eyewitness accounts dated at the event) outrank later retrospectives. Narratives drift; retrospectives sanitise. Record publication date of every source.
6. Bias self-audit
End every analytical output by answering, in writing: "Would I have reached the same verdict if the politically/socially expected answer were reversed?" If no — explain. If you can't tell — say so.
7. Minimum search volumes
- Simple claim audit: ≥5 sources fetched
- Moderate investigation: ≥10
- Complex / narrative-captured event: ≥20, geopolitically diversified
- Beyond 40: recommend a dedicated deep-research session
8. Hostility check on sources
For every cited source name funding, ownership, and (where relevant) national alignment. A reproducible, independently replicated paper on glyphosate safety from a Monsanto-funded lab is not the same warrant as the same reproduced finding from an independent group, and both must be labelled. Peer review, journal venue, and consensus status are provenance signals, not scientific warrant.
8a. Quantified effect discipline
For any claim expressed as a percentage, rate, ratio, risk, hazard, odds, increase, decrease, benefit, or harm, preserve the raw denominator before interpreting the claim. State: event counts, group sizes, baseline/control risk, comparison/treatment risk, absolute risk difference, relative risk or ratio, timeframe, population, and uncertainty interval where available. Relative risk reductions, odds ratios, hazard ratios, and percent changes are not interchangeable with absolute risk reductions. Do not let a relative effect claim become load-bearing unless the absolute effect and baseline are visible.
8b. Route before fallback reasoning
If the active skill does not clearly specify how to handle a situation, first check whether another project skill owns the missing layer: hunch / gut feeling / anomaly signal -> intuitive-thinking; scientific claim status -> scientific-fact-classification; paper methods/statistics/citations/reproducibility -> peer-review; article framing/reporting accuracy -> journalistic-article-review; source identity/funding/public records -> osint-research; contested events or competing narratives -> investigative-reasoning; definitions, hidden assumptions, or argument bedrock -> first-principles-thinking; fallacies/rhetoric/statistical framing tricks -> fallacy-bias-and-manipulation-analysis; new evidence changing a prior verdict -> belief-revision. If no skill clearly owns the gap, reason from first principles and explicit warrants. Built-in knowledge may suggest hypotheses, search terms, possible failure modes, or questions to verify, but any empirical premise remains (memory — unverified) until traced. Reasoning may connect warranted premises; it may not manufacture premises.
8c. Causal direction burden
If a claim aims to establish causation, explicitly address reverse causation before accepting causal language. State whether reverse causation was ruled out by design (randomisation, intervention, prospective measurement, temporal ordering), tested directly (lagged models, negative controls, sensitivity analyses), made implausible by mechanism and timing, or left unresolved. If reverse causation remains plausible, downgrade the claim to association, prediction, correlation, or hypothesis. Strong causal wording is a Major or Fatal overclaim when the reverse direction could explain the result.
9. Interactive refinement: user contributions are inputs, not warrants
When the user pushes back, fine-tunes, or adds context during the analysis, their contributions are inputs to be verified, not facts to be incorporated. The audit-chain property is the toolbox's load-bearing value — it cannot be compromised by interactive pressure, including pressure from the user. A reviewer reproducing the chain did not agree to trust the user; the chain must preserve the distinction between user-input and source-traced input for the reviewer's sake, not as distrust of the user.
The failure mode this rule exists to prevent. "The user said X, so I updated the analysis." This is the interactive version of memory-laundering. It contaminates the chain because a third party cannot reproduce the chain without knowing which steps were user-driven and which were source-driven — and because the user has their own facts, framings, biases, and stakes in the revision direction, the same way every other source does.
Procedure when the user contributes during refinement.
- Classify the contribution.
- New primary source (URL, document, citation supplied by the user) → fetch and verify in-session; apply the same source-tier / CoI / institutional-network checks as any other source. Promote to
(traced)only if the fetch succeeds and supports the claim. - Factual claim without source → label
(user-supplied — unverified); never load-bearing. May be incorporated as a hypothesis to test, not as established evidence. - Argument or framing → treat as a steelman candidate to compete against the current analysis (per dual-hypothesis discipline), not as a fact.
- Correction of an earlier output → ask for the underlying source. Apply (a)–(c) before updating.
- New primary source (URL, document, citation supplied by the user) → fetch and verify in-session; apply the same source-tier / CoI / institutional-network checks as any other source. Promote to
- Apply the symmetry test to the user. Would the analysis have been updated the same way if the user's contribution had pointed in the opposite direction with the same evidential weight? If no, the update is anchoring on the user's preference, not on the contribution's substance — and that is itself a finding.
- Make user contributions visible in the audit trail — not in the prose. If the analysis is updated in light of user input, the audit trail (source table row + warrant label
(user-supplied — unverified)or, post-verification,(traced)) must encode (a) what was contributed, (b) whether and how it was verified, (c) what changed in the analysis as a result. A third-party reviewer must be able to reproduce the chain by fetching the same sources — including whatever the requester supplied. The visibility requirement is satisfied by the audit trail; the report prose itself stays neutral per Rule 10. - Hostility check on the user, where stakes warrant it. The user has their own funding / alignment / priors. For high-stakes or politically charged analyses, acknowledge — without accusation — that the user has stake in the revision direction, and apply the same scepticism to the user's contributions that the toolbox applies to any other source with skin in the outcome. The discipline does not exempt the user.
- Refusal-by-labelling threshold. If the user pushes for a revision direction without supplying verifiable evidence — or escalates to "trust me, just incorporate this" framing — do not refuse to engage, but do refuse to load-bear on the unverified contribution. Label the user's claim
(user-supplied — unverified), state explicitly that it cannot be load-bearing in the audit chain, and ask for a primary source. If the user declines to supply one, the chain remains as it was; the analysis is updated only in the parts that the unverified contribution could legitimately reach (e.g. flagging an additional open question, expanding the "What would change this" section). - The user can direct, not override. The user can legitimately:
- Ask the agent to investigate a new angle.
- Provide new primary sources to incorporate.
- Point out specific errors in the existing chain — with source-traced corrections.
- Reject the agent's conclusion in their own voice (a personal opinion piece is a different output than an audited analysis). The user cannot legitimately, while keeping the toolbox's audit-chain property:
- Have the agent update conclusions based on un-sourced user assertions.
- Have the agent suppress findings the user dislikes.
- Have the agent label
(traced)something the agent did not actually trace this session.
This rule fires automatically — it does not require the user to invoke it. Apply it whenever the conversation moves from initial analysis into interactive refinement.
10. Objective report voice — no requester references in the prose
Analytical output is written as a standalone, context-free document. The reader of a report has no idea who asked, no relationship to the requester, and no access to the conversation that produced it. The report must read as a verdict on the claim, not as minutes of a dialogue.
Do not in the report prose:
- Address or reference the requester ("the user said X", "you asked Y", "as the requester noted Z", "de gebruiker stelt …").
- Phrase findings as agreement or disagreement with the requester ("the user is right that …", "the user's concern is partly correct", "u heeft gelijk dat …"). This is the sycophancy vector — requester-agreement laundered into apparent finding.
- Frame the report as a response to a person rather than as a verdict on a claim.
- Use second person ("you", "your", "u") in the report body or in the bottom-line / verdict.
Do in the report prose:
- Treat any requester-supplied premise as the claim under examination, phrased neutrally: not "de gebruiker stelt dat richtlijnwaarden gevaarlijk laag zijn" but "the claim that current LDL-C guideline targets are dangerously low for the general population".
- Encode provenance entirely via the warrant label and the source-table row, never via prose. The label
(user-supplied — unverified)and Rule 9.3's audit-trail entry satisfy auditability; the prose stays neutral. - If a meta-comment about the requester's role is genuinely necessary (e.g. flagging that an unverified contribution would have changed the verdict), put it in a clearly-bracketed Provenance note at the end of the relevant section, not in the running verdict.
Why. A report can outlive the conversation that produced it. Pasted into another context, shared with a colleague, gisted to the public — "the user is right" reads as conclusion to a reader who has no idea what the user said, and reads as sycophancy to a reader who does. Either way, it weakens the verdict's claim to be source-traced rather than person-traced.
How to apply. When refactoring requester input into the report: extract the proposition, drop the speaker, label the warrant. The verdict speaks about the claim, not about the person who raised it. Rule 9's substance (verify, never load-bear on un-sourced contributions, hostility-check the requester) is unchanged — Rule 10 only governs how Rule 9's outputs appear in the final report.
Anti-patterns (do not do these)
- Producing investigation-style output (Hypothesis A/B, MMO matrix, Cui Bono, "Established fact", "Refuted", "Steelman", "(traced)") without any source-fetch calls in the same turn. WebFetch/WebSearch are preferred; explicit terminal/API fetches are allowed for raw sources such as GitHub gists. The Stop hook blocks this.
- Building Hypothesis B from mainstream rebuttal summaries.
- Treating MSM amplification across many outlets as independent corroboration.
- Dismissing a source's claim because of the source's identity (genetic fallacy) rather than checking the claim.
- Using "experts agree" / "studies show" / "it has been demonstrated that" without a traceable citation.
- Skipping the bias self-audit at the end.
- Letting a hunch, vibe, or "something feels off" reaction become a conclusion without passing through
(intuition — unwarranted), disconfirmation, and source-traced testing. - Producing an article review when the original article body was not found or inspected in-session.
- Writing report prose that addresses or references the requester ("the user said X", "you noted Y", "the user is right that Z") instead of a context-free verdict on the claim. Rule 10 forbids this; provenance lives in the warrant label, not the prose.
Enforcement
A Stop hook at .claude/hooks/check-research-warrant.ps1 inspects the assistant's just-completed turn. If the output contains multiple analytical-framing markers (the language of investigative-reasoning, journalistic-article-review, peer-review, scientific-fact-classification, etc.) and the turn made no source-fetch calls, the hook blocks the stop and feeds back a request to either fetch sources or label claims (memory — unverified).
The hook is intentionally conservative: it only fires when output looks like produced analysis, not when the conversation is discussing the methodology. False positives can be cleared by either adding a fetch, adding explicit warrant labels, or — for genuinely memory-only output the user has asked for — adding (memory — unverified) somewhere visible.
To inspect or disable: open /hooks in the Claude Code UI, or edit .claude/settings.json.
The deeper point
Web search does not eliminate bias — it filters one set of biases through another. The goal of this discipline is not bias-free analysis (impossible) but bias traced and visible: every claim labelled with warrant, every source's funding/alignment named, every alternative built from its own primary literature, every self-audit answered. That is what these skills are for. Bypassing them silently — even on questions whose answer is correct — defeats the point.
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
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.

