claim-evidence
Skill0 starsChanged 3 months agoMITView on GitHub
Back every factual claim about a codebase with a file and line number, and refuse to state as fact anything you have not actually read. Use when the user asks "are you sure", "where is that", "prove it", or when you are about to describe code behavior, architecture, or a bug to the user.
In this repository11 other files it gives its agents.
- agent-verificationSkill
- crypto-news-digestSkill
- defi-risk-assessmentSkill
- evidence-firstSkill
- flaky-test-detectorSkill
- on-chain-analystSkill
- prompt-injection-hardeningSkill
- refactor-verifierSkill
- skill-security-auditSkill
- smart-contract-reviewSkill
- test-integrity-guardSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
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 25
Do not present a hypothesis as a fact.
- Distinctive: line 29
Before asserting, read: If the claim is about this repo, use your environment's file-reading and code-navigation tools (read a file, search …
- Distinctive: line 67
When the user asks a "where / why / prove it" question, answer with:
≈ “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.
Files
Claim Evidence
An agent that states confident falsehoods destroys trust faster than one that says "I don't know yet." The fix is mechanical: every claim about the code carries a citation, and a claim you cannot cite is not yet a claim — it is a hypothesis, and must be labeled as one.
The distinction that matters
- Fact — you have read it in a file this session, and can point at the
exact location. "The route is defined in
backend/routes/auth.js:42." - Hypothesis — you believe it from context, training, or a pattern, but have not verified it in this codebase. "I'd expect the auth middleware to run first — I haven't confirmed it yet."
Do not present a hypothesis as a fact. The label is the difference.
Procedure
- Before asserting, read. If the claim is about this repo, use your environment's file-reading and code-navigation tools (read a file, search its contents, jump to a definition, find its usages — whatever your client exposes) to locate the truth. Do not reason from a file you have only seen referenced.
- Attach the evidence. State it as
path:line, and quote the smallest slice that supports the claim. - Narrow claims that are actually narrower. "The API returns 401 on expired tokens" is a fact about one code path; "the whole app is secure" is a hypothesis built on it. Keep the claim at the level you verified.
- When unsure, say so. "I haven't verified X" is a correct and valuable answer. Prefer it to a guess dressed as knowledge.
Citation format
Use one of:
path/to/file.ts:42 — `return tokens;`
path/to/file.ts — function `resolveToken` (exact line not read)
unknown — hypothesis, not verified in this sessionThe unknown tag is allowed; omitting it and stating the thing anyway is not.
Why this matters beyond honesty
Two concrete failure modes this prevents:
- Stale belief. You remember
getPricing()from a previous session, but the route now bills through a different path. Only a fresh read catches the drift. - Hallucinated shape. Naming a function, file, or API shape you inferred rather than read sends the user (and your own later steps) down a wrong path. A citation forces the read that would have caught it.
Output format
When the user asks a "where / why / prove it" question, answer with:
Claim: <one sentence>
Evidence: <path:line> — <quoted slice>, or "unverified"
Confidence: fact | hypothesisDoes 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.

