deep-check
Skill0 starsChanged 44 days agoMITView on GitHub
Evidence first verification of anything: a claim, a fix, a config, a deployment, a dataset, a piece of content, another AI's output. Produces a verdict (CONFIRMED / PLAUSIBLE / UNVERIFIED / REFUTED) backed by primary evidence gathered with tools, plus an explicit list of what was NOT checked. Use whenever the user says "check this", "verify", "double-check", "is this actually working", "audit this", "make sure", "did it really", or before telling the user that something works, is fixed, is deployed, or is done. Also use to review work done by subagents, scripts, or earlier sessions before trusting it.
In this repository6 other files it gives its agents.
- duo-codexSkill
- duo-geminiSkill
- fableSkill
- orchestrateSkill
- teamSkill
- trioSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 43
Exercise the real flow: Run the command, hit the endpoint, open the file the user will open, run the app end to end.
- Distinctive: line 45
Inspect the actual artifact: Read the deployed file, the generated output, the database row, rather than the source that supposedly …
- Very specific: line 73
When the stakes are high, or when you produced the work yourself, spawn a verifier agent (or a general-purpose agent) to check it *without …
≈ “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
Deep Check
The core rule: a claim is unverified until you have seen primary evidence through your own tools in this session. Names, comments, docs, commit messages, agent reports, and your own memory of earlier turns are all testimony, not evidence.
Method
Work through these five steps for each claim being checked. For a batch of claims, triage first: verify the load bearing ones deeply and spot check the rest.
1. State the claim precisely
"The site works" is not checkable. "GET https://example.com/sitemap.xml returns 200 with at least 40 URLs" is. Rewrite vague claims into observable statements before checking anything. Often the rewrite alone exposes that nobody knows what was actually promised.
2. Enumerate how it could be false
Spend a moment listing the two to four most likely failure modes, because that decides what to observe. A fix can be false because the code path never executes, an edge case bypasses it, it works locally but not where it is deployed, it broke something adjacent, or the observed success was cached and stale.
3. Choose the cheapest discriminating observation
Pick the observation that would distinguish true from false, not the one that merely feels related. Reading the code that should produce behaviour X is weaker than triggering behaviour X and watching it happen. Prefer, in order:
- Exercise the real flow. Run the command, hit the endpoint, open the file the user will open, run the app end to end.
- Inspect the actual artifact. Read the deployed file, the generated output, the database row, rather than the source that supposedly produces it.
- Read the code or config. Weakest, and acceptable only when execution is impossible.
4. Observe, with the traps in mind
- A file existing is not the same as a file having the correct content. Read it.
- Exit code 0 is not success. Read the output, because many tools fail politely.
- Tests passing is not the same as the feature working. Tests check what tests check, so drive the real flow for anything user facing.
- "It worked earlier in this conversation" is not "it works now", if anything changed since.
- Docs and comments describe intent at the time of writing, not current reality.
- Memory files and CLAUDE.md describe what was true when written, so re-verify paths, flags, and URLs before relying on them.
- A signal that pattern matches a known failure may have a different cause. Check that the evidence supports the specific diagnosis before acting on it.
5. Assign a verdict
- CONFIRMED. Primary evidence observed in this session directly supports the claim.
- PLAUSIBLE. Code or config reading supports it, but the flow was not exercised. Say what would upgrade it.
- UNVERIFIED. Could not gather discriminating evidence, for example missing access, environment or credentials. Never round this up to "probably fine".
- REFUTED. Evidence contradicts the claim. Quote the evidence.
Independent verification
When the stakes are high, or when you produced the work yourself, spawn a verifier agent (or a
general-purpose agent) to check it without telling it the expected answer. Give it the claim and
the paths, not your reasoning. Confirmation from an agent that was told "confirm this works" is
worthless. Confirmation from one asked "determine whether X is true" is signal.
Report format
Lead with the verdict, then the evidence, then the gaps:
**Verdict:** CONFIRMED (2 of 3 claims) / REFUTED (1)
1. <claim> = CONFIRMED. <one line of evidence: command run plus relevant output, or file and line>
2. <claim> = REFUTED. <quoted evidence>
3. <claim> = PLAUSIBLE. <what was read, and what real flow test would confirm it>
**Not checked:** <anything in scope you did not verify, and why>The "Not checked" section is mandatory. Silent scope shrinking, meaning verifying the easy 80 percent and letting the reader assume 100 percent, is the most common way verification lies.
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.

