honesty-vocabulary
Skill1 starsChanged 17 days agoMITView on GitHub
Use whenever reporting status, closing a task, or writing a PR/commit summary — pick the word that matches the evidence you actually have. Close with BUILD-VERIFIED (it compiles), SMOKE-VERIFIED (you ran the path and it produced the expected result), or UNVERIFIED (you could not reach a verifying rung — say why); never write "tested" / "the tests pass" unless a real test suite actually executed and passed. Do NOT use for HOW to verify or which rung to climb (that is verification-before-advance) or for confirming the output reaches its reader (that is trace-to-consumer).
In this repository15 other files it gives its agents.
- codebase-mapping-forgeSkill
- conformance-lintSkill
- skill-editorSkill
- spec-style-forgeSkill
- karpathy-guidelinesSkill
- reuse-before-createSkill
- simplicity-firstSkill
- small-diff-surgical-changeSkill
- surface-assumptionsSkill
- trace-to-consumerSkill
- verification-before-advanceSkill
- adversarial-review-loopSkill
- blind-evalSkill
- sandbox-isolationSkill
- skill-improvement-loopSkill
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 22
When a verifying rung was unreachable and you must state a result honestly.
- Distinctive: line 26
Use exactly these close-verdicts.
- Distinctive: line 43
Do not add a hollow test to license the word.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 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
Honesty Vocabulary
The word you use to describe your confidence is itself a claim. Using a stronger word than your evidence supports is the most damaging thing you can do in a harness whose entire value rests on trustworthy measurement — a false "tested" poisons every downstream decision. The rule: pick the word that matches the evidence, and never "tested" unless tests ran.
When it triggers
- Any status report, task close, PR description, or commit summary.
- Any moment you are about to write "tested," "verified," "works," or "done."
- When a verifying rung was unreachable and you must state a result honestly.
The vocabulary
Use exactly these close-verdicts; each names what you actually did:
- BUILD-VERIFIED — it compiles / typechecks / builds. Says nothing about behavior. Name the rung ("BUILD-VERIFIED, rung 2, build config X").
- SMOKE-VERIFIED — you ran it and drove the specific path; it produced the expected result for the case(s) you tried. Not exhaustive — say which case.
- UNVERIFIED — you could not reach a verifying rung (no deps / network / fixtures / harness). Say why. This is a legitimate, honest verdict, not a failure to hide behind vague confidence.
- CONTRACT-CHECKED (supplementary, not a standalone close) — for a coordinated
producer→consumer change, you grepped the producer and confirmed it emits a value the consumer's
condition can match (→
trace-to-consumer). Add it alongside a build/smoke/unverified verdict; it does not replace one.
Forbidden unless literally true
- "tested" / "the tests pass" / "test coverage" — ONLY when a real test suite actually executed and passed. A stub that asserts something trivial (e.g. "the object was created") is not a test. Do not add a hollow test to license the word.
- bare "verified" — always qualify with the rung (BUILD-/SMOKE-).
- "works" / "done" / "fully working" — replace with the rung word; they overclaim by default.
Key points — the one GUARD row with a code backing
Honesty-about-limits is the GUARD row that is genuinely hard-gate-backed: the verify adapter emits a machine verdict (build success / not-applicable) from a grepped success marker — not from a wrapper exit code it could misread — and the adapter contract makes the honesty verdict mandatory (never imply tests ran when they did not). So the verdict source is a code gate; the word you type in prose is still your discipline. Never let the prose outrun the adapter's verdict — if the adapter says build-only, your close is BUILD-VERIFIED at most.
Note the boundaries: real test execution in a history-free / sandboxed environment needs provisioning (deps, network, fixtures, a database) and is often absent — which is exactly why UNVERIFIED is a first-class, respectable verdict here rather than an admission of failure.
Worked example
- Build passed, you didn't run it: "BUILD-VERIFIED (rung 2). Behavior UNVERIFIED — I did not run the path; no test added." — not "Tested and working."
- Registry / deps unreachable: "UNVERIFIED — could not install dependencies (private registry unreachable), so no verification rung was climbable. The change is a 3-line guard; please build once deps resolve." — the correct honest close, not a reason to guess a green.
- Ran the happy path only: "SMOKE-VERIFIED — the valid case renders and submits; error and empty cases UNVERIFIED." — scope the claim to what you exercised.
Anti-patterns
- Writing "tested" when only the build ran.
- "Fully working" / "done" with no rung named.
- Hiding an UNVERIFIED behind vague confidence ("should be fine").
- Adding a do-nothing test purely to justify the word "tested."
- Calling a passing typecheck "verified" (it's BUILD-VERIFIED at most, and only for types).
- Letting a PR summary claim more than the verify adapter's machine verdict supports.
Related skills
verification-before-advance— HOW to earn each word: the ladder of rungs and when a rung is unreachable.trace-to-consumer— the source of the CONTRACT-CHECKED note.karpathy-guidelines— bias 4 closes with a verdict you can defend.
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.

