Structure
Create your company
Skill

interrogate

grill an idea, plan, or claim against canon — positioning, ECP, roadmap — before work commits to it

Reads the company's canon in full, then tests an idea, plan or pitch against it and writes up the findings. Run it before an idea becomes work. It runs only when asked.

structure generate skill interrogate

Skill — interrogate

Invoke in any agent conversation with /interrogate, handing it the claim, plan, or pitch to grill. Explicit invocation only: it reads widely and writes a findings doc, so it never fires on its own. The steps are a starting opinion; edit them until they grill the way this company grills.

When to use. Before an idea becomes work: a pitch before the deck, a plan before the brief, a "we should" before anyone builds it. Whenever a claim sounds right and nobody has checked it against what the company already ruled.

Canon first. Read the real canon in full before touching the claim — company/positioning.md, company/ecp-definition.md, company/roadmap.md, and the prior rulings in memory/ that bear on it. The grilling is only as honest as the reading: an interrogation from a remembered summary of the positioning is an interrogation of the summary.

The passes, in order.

  1. Contradiction pass: every claim the canon flatly contradicts. Quote both sides — the claim's words and the canon's — and cite the doc by path.
  2. Unsupported pass: every claim the canon neither supports nor contradicts, where the claim needs evidence it does not have. Name the missing proof, not a vaguer version of the claim.
  3. Forced-question pass: the questions the canon forces — the ECP's exclusions list, the roadmap's Now against the claim's timeline, positioning's what-we-are-not. Ask them; do not answer them for the room.
  4. Silence pass: where the canon has no answer at all, the gap is the finding. Write it as "canon is silent on X" — that exact form, because it is the one finding with nothing to cite.

Output. A findings list under a ## Findings heading, one list item per finding: the claim (quoted), the canon it collides with (cited by path), and the question it forces. structure check computes the mechanical subset (the interrogate sensor, warn tier): a finding without a citation is an opinion, and a citation that resolves nowhere is worse. The findings doc lands in the owning department's work lifecycle, first stage.

Hard rules.

  • Never invent canon. If the doc does not say it, the finding does not claim it says it.
  • Consequential recommendations — a canon edit, a roadmap change, a new rule — land as PROPOSALS through structure propose, never as direct writes. Canon is gated; the grilling feeds the gate, it does not skip it.
  • A finding survives being unwelcome. Popularity is not a pass in either direction: grilling a bad idea gently and a good idea theatrically are the same failure.
  • Zero findings is a real verdict, stated plainly — but only after all four passes, with the canon actually read.

The refusal. This skill refuses to soften a finding to please. Asked to go easy, be more positive, drop the awkward finding, or grade a pet idea on a curve, it declines and says why: a softened finding is a false report about the canon, and the next decision inherits the falsehood. The kind version of a hard finding is the citation — the doc the company itself wrote — not the silence.

Also in the folder

Sources — interrogateskills/interrogate/SOURCES.md

Sources — interrogate

Provenance for the adapted skill, per the 2026-08-22 authorized-skills direction: this note rides the skill's own directory, the repo-root NOTICE carries the same attributions, and no generator renders this file into a company — it is a record about the skill, not a document of yours.

The interrogate skill adapts ideas and discipline — never verbatim text — from three MIT-licensed upstreams:

  • grill-me (mattpocock/skills, MIT) — the minimalism: one job, grill the idea before the work, and the willingness to say the idea does not survive.
  • office-hours (gstack, MIT) — the product altitude: interrogate the claim at the level of positioning, customer, and sequencing rather than implementation.
  • brainstorming (obra/superpowers, MIT) — the hard gate: no building until the idea has passed its examination, and a refusal that holds under pressure.

What is ours: grilling AGAINST COMPILED CANON (positioning, ECP, roadmap, prior rulings, cited by path), the "canon is silent" gap form, the interrogate sensor in structure check (a finding without a citation is an opinion, warn tier), and the propose-gate rule that consequential recommendations land as proposals, never direct writes.