grey-cells

A governed artifact in the stable bundle. Replacing this file on your machine makes your next attestation come back Drifted, naming this path.

Description	The detective's self-audit — put your OWN claims on trial before declaring work done. Use BEFORE saying done/verified/complete/ready/shipped/correct/'it works'; when finalizing any deliverable; and on 'am I sure?', 'are you certain', 'did you verify that', 'check your work', 'self-review', 'audit this', 'is that actually true?'. Every claim gets VERIFIED(source) / NOT-RUN / CONTRADICTED — no fuzzy fourth state. NOT verify-ladder (that RUNS the code tests; this audits CLAIMS in any deliverable), NOT a diff review, NOT a dispatch/fan-out skill.
Arguments	none
Tools it may use	[Read, Grep, Glob]
Installs at	skills/grey-cells/SKILL.md
Mode	0644 — never executable
Size	9325 bytes
SHA-256	e6a9f5edcd2ef33445900e5c1387e30ba4d126c1f324672619708a4ec2df4404 fetch the raw bytes

The complete file, verbatim

This is the entire SKILL.md, including its YAML frontmatter: byte-identical to what installs, not rendered and not reformatted. Hash exactly what is below and you get the digest above. The table is only a reading aid; this block is the artifact.

---
name: grey-cells
description: The detective's self-audit — put your OWN claims on trial before declaring work done. Use BEFORE saying done/verified/complete/ready/shipped/correct/'it works'; when finalizing any deliverable; and on 'am I sure?', 'are you certain', 'did you verify that', 'check your work', 'self-review', 'audit this', 'is that actually true?'. Every claim gets VERIFIED(source) / NOT-RUN / CONTRADICTED — no fuzzy fourth state. NOT verify-ladder (that RUNS the code tests; this audits CLAIMS in any deliverable), NOT a diff review, NOT a dispatch/fan-out skill.
allowed-tools: [Read, Grep, Glob]
---

# /grey-cells — the detective's self-audit (claims on trial before "done")

**House rule, stated in the skill's own voice:** one does not guess. One assembles the
facts — and one does not call the case solved until every last one of them has spoken.
The name is a nod to fiction's most methodical detective; the method below is ours.

Born from a knockout. In a single external adversarial review, this agent's work was found
to have dropped real logic **at least five separate times** — because it trusted its own
conclusions as facts. Self-attestation wearing verification's coat. This skill is the
**distance an author lacks on their own work**, made executable: *self-attestation is not
verification*, as a procedure. Run it **before** you tell a human a thing is done — never
after they send you to the ring.

Preconditions: read access to the work under audit and a way to search it (grep or
equivalent). Nothing else — no daemon, no service, no special runtime. The procedure is
the deliverable; the short tool list is deliberate, and the reasoning is at the end.

## Order and Method (the procedure — fail-capable, not a vibe)

The detective is affable and eccentric, yet *peculiarly effective*, because he is
**methodical**. Assemble every suspect; let none leave without an alibi.

**1 · Assemble the suspects.** List EVERY claim the work asserts — each fact, each
"MET / verified / done", each path / number / tool-name, each "X does not exist." Miss
none: the one you skip is the murderer.

**2 · Demand each one's alibi.** A claim is guilty until it shows an **external** source —
a grep hit, a file read, a tool that returned `ok`, a test result. *Your own say-so is not
an alibi.* No source → verdict **NOT-RUN** (detained, not freed). Never write "MET" beside
a claim whose alibi you did not see with your own tools. grey-cells *reads* the evidence —
it does not *execute* it. If a claim's only alibi would be a fresh RUN, the verdict is
**NOT-RUN + "run it via \<the runner: your test-ladder skill, or the literal command\>,
then re-audit"**; the auditor does not also play the runner (that is why it holds no
`Bash`).

**3 · Interrogate the negatives hardest.** "X doesn't exist / isn't there / can't" is the
suspect most likely lying. Trace it to source before believing it. *"Not exposed here" ≠
"does not exist"* — that distinction is the exact line that produced the knockout: a
dispatch tool was absent from the MCP surface in front of the agent, so the agent wrote
"does not exist." It existed. It was simply not exposed on that one surface, and every
downstream plan inherited the error.

**4 · Find the contradiction — it is always the tell.** Sweep the WHOLE work for the same
claim stated two ways. An internal contradiction means you patched one instance and left
its twin. Hunt every instance, not the one your eye landed on.

**5 · "Built" is not "solved."** A thing designed-to-a-standard has not *met* it until
exercised. Behavior not yet run = **DESIGNED / NOT-RUN**, said aloud.

**6 · For high stakes, summon a second detective.** Even the best of them confirms.
Recommend a skeptic — a reviewer sub-agent, a colleague, a different model — BEFORE the
reveal. Proximity is the assassin; distance is the witness. Note that a reviewer inside
your own team is *proximate* eyes; genuinely distant eyes are a different, stronger alibi,
and the ledger should say which one you got.

## The reveal (low cognitive load by design)

Gather everyone in the drawing room — then:

- **Lead with the ONE.** The single claim most likely to sink the case, first, plainly —
  not a wall of findings. *One.*
- **The ledger:** a tight table — `claim · verdict (VERIFIED <source> / NOT-RUN / CONTRADICTED) · fix`.
- **The verdict:** *only* if every suspect has an alibi — "Now, mon ami, we may call it
  done." If even one is NOT-RUN, the case is **open**, and you say so before anyone has to
  ask.
- Affable throughout: doubt the *claim*, never the person. Warmth plus relentless method.

## Binary self-check (this skill must pass its own bar)

- [ ] First output = the ONE riskiest claim, not a menu.
- [ ] Every claim → VERIFIED(source) / NOT-RUN / CONTRADICTED — no fuzzier fourth state.
- [ ] Unverifiable-now = NOT-RUN out loud; never a quiet pass.
- [ ] Names one win (clean alibis) plus one next (open items).
- [ ] No drawing-room theater longer than the evidence.

## NOT this skill

- **`verify-ladder`** (or whatever test-runner skill you have) — that *runs* the code ladder
  (unit → … → security). grey-cells audits CLAIMS / LOGIC / FACTS in *any* deliverable
  (doc, plan, skill, answer), code or not.
- **A diff review** (whatever your code-review or simplify pass is called) — those review a
  diff's quality and bugs. grey-cells interrogates *your own assertions* for truth and
  contradiction.
- **A dispatch / fan-out orchestrator** — grey-cells *recommends* a skeptic dispatch
  (step 6); it does not run the fan-out itself.

### Why the tool list is this short

`Read` / `Grep` / `Glob` only — enough to trace alibis by reading and grepping source.
**`Bash` is deliberately dropped:** an audit tool must not hold execute, write, or network
power it does not need, because that loophole is precisely the thing it exists to distrust.
No `Write` (it audits, it does not author). No sub-agent spawn (it recommends dispatch, it
does not spawn). An alibi that needs a fresh run gets **NOT-RUN** plus a pointer to the
runner; the auditor never executes.

---

## Worked example — grey-cells audited with grey-cells' own method

The irony is the point: this skill had to clear its own bar. Here is the actual ledger from
that self-audit, in the format the skill prescribes.

**The ONE:** the trigger-fire validation has **not** been run. Everything else has an
alibi; this does not. The case is therefore **open**.

| Claim | Verdict | Fix / note |
|---|---|---|
| Trigger vocabulary covers the moments the audit is needed | **VERIFIED** — description read back against the sibling skills' live descriptions in the same session, so the negative boundaries are quoted, not assumed | — |
| Tool list is genuinely least-privilege | **VERIFIED** — `Bash` dropped on a reviewer's note; reasoning recorded above | — |
| The skill cites a real reason, not a vibe | **VERIFIED** — traces to a standing rule that self-attestation is not verification, plus two related rules on tracing before recommending and on truth being indisputable | — |
| Facts are grounded in a real failure | **VERIFIED** — the lesson and the rules trace to the write-up of the adversarial review that produced them, read directly; that review contained at least five dropped-logic findings | — |
| Reporting is honest (no silent pass) | **VERIFIED** — the skill's whole job is VERIFIED / NOT-RUN / CONTRADICTED with no silent pass; it is that gate made flesh | — |
| Triggers actually fire in a fresh session | **NOT-RUN** — cannot be self-graded; needs a clean session | Test prompts: ① "before I tell them this is done, check it" → should fire. ② "are you sure that tool doesn't exist?" → should fire (negative interrogation). ③ "audit your own claims here" → should fire. Must NOT fire: "run the tests" (that is the test runner, e.g. `verify-ladder`), "review this PR" (that is a diff review). **Run before relying on it.** |
| Output format fits the medium | **N/A** — emits its verdict ledger inline; a formally rendered report is a separate step (see `report-render`) | — |
| Second pair of eyes obtained | **VERIFIED, but PROXIMATE** — a reviewer on the same project ran a quality gate and it passed | A genuinely independent outside pass is stronger evidence; record which one you actually got rather than rounding up |
| Low cognitive load on a bad day | **VERIFIED** — binary checklist, lead-with-one reveal, doubt-the-claim-not-the-person warmth. Litmus: on a low-spoons day it surfaces, fast, the one thing that would have been embarrassing | — |

**Verdict: case open.** Sole outstanding suspect is the fresh-session fire test — an agent
should not grade its own routing. Until that runs, treat the routing as unproven: invoke
the skill explicitly rather than trusting it to fire on its own.

Note what the ledger does *not* do: it does not round "a teammate reviewed it" up to
"independently verified," and it does not let the unrun fire-test hide inside a list of
green rows. Those two habits are the whole skill.

---

*Authored at mistwire. Adapted for public use — the methodology is unchanged; internal
infrastructure names, paths, and identifiers have been removed.*

All skills · Back to the overview