Skip to content

jrichlen/grill-me

v1.0.0MIT

Interview the user about a plan before work starts, single-session and no subagents required, walking its design tree and scaling question depth to each branch's stakes (reversibility x blast radius) while offering a recommendation at almost every step. Use on "grill me" or "stress-test this plan".

GitHubOpen Website

grill-me

Interview the user about a plan before work starts, single-session and no subagents required, walking its design tree and scaling question depth to each branch's stakes (reversibility x blast radius) while offering a recommendation at almost every step. Use on phrases like grill me, interview me about this plan, stress-test this plan, or before starting a nontrivial multi-step change whose design isn't yet settled.

Install

/plugin marketplace add JRichlen/agent-plugins
/plugin install grill-me@jrichlen

PORTABILITY: harness-agnostic. grill-me is a plain conversation with the user — no subagents, hooks, or Workflow tool required on any harness. (The Differentiation section below names those primitives only to describe two other plugins, not a dependency of this one.)

How it works

Full procedure lives in skills/grill-me/SKILL.md (kept lean, always resident). In short: model the plan as a design tree, walk it in numbered rounds over the current "frontier" of decisions whose prerequisites are already settled, and triage every branch on two axes — reversibility (two-way door vs. one-way door) and blast radius (internal/dev-only vs. user-facing/production/cross-team/financial/compliance) — before deciding how hard to press on it. LIGHT branches get one confirm-or-skip question; STANDARD branches get their natural 1-3 questions; DEEP branches (irreversible and wide blast radius) get the full sub-tree, a devil's-advocate pass, and a required explicit confirm. Every question, at every tier, ships with a stated recommendation the user can accept with a bare "yes." The session ends with a synthesized recap (Confirmed Decisions / Open Risks Accepted As-Is / Deferred-for-Later) and a request to confirm shared understanding before the plan is treated as ready to execute.

Deeper technique detail (devil's-advocate craft, pre-mortem prompts, worked triage examples, fact-finding rules) lives in skills/grill-me/references/ and loads only when a session actually needs it — that's the token-efficiency half of the design, not a missing feature.

Differentiation — why this isn't second-opinion or wave-research

Two other plugins in this marketplace sit near this one. Neither does what grill-me does:

  • second-opinion (plugins/voice/skills/second-opinion) is post-hoc, offer-only, subagent-required fact-checking of a verdict or recommendation that already exists — it batches fact-checks against sources, adds scoped advisor personas, and re-emits the response as Verified/Flagged/Conflict with a mandatory delta. It never runs unbidden.
  • wave-research (plugins/wave-research) is a staged research method with Workflow-tool templates for fanning research out across multiple subagents and adversarially verifying the claims that research surfaces.

grill-me is neither of those. It is a single-session, no-subagents-required, live conversational interrogation of a plan, conducted directly with the user before any work starts, to reach shared understanding — not to validate a finished answer, and not to fan research out across workers.

Design departures from the original grill-me

This skill is a deliberate rewrite of the technique, not a port. Three departures, by design:

  1. Dynamic, not fixed-depth. The canonical version exhausts every branch of the design tree uniformly. This version spends question budget in proportion to how expensive a branch would be to get wrong (reversibility x blast radius), so a naming choice gets one line and a public-API or security decision gets a full devil's-advocate pass.
  2. Token-efficient, not monolithic. SKILL.md carries only the invariant, the triage table, and the round loop — the material every session needs. Devil's-advocate technique, fact-finding mechanics, and worked triage examples live in references/ and load only for the sessions that actually need them (DEEP branches, contested calls, mid-round lookups, ambiguous triage).
  3. Assistive, not pure interrogation. Every question ships with a stated recommendation from round one — not as a later speed retrofit. DEEP branches additionally get a short tradeoff comparison before the recommendation, so the offered answer reads as an audited conclusion rather than an assertion.

Attribution

The interview technique — walking a design tree in rounds, offering a recommendation, reaching shared understanding before work starts — is adapted from Matt Pocock's grill-me / grilling skills, MIT-licensed at github.com/mattpocock/skills (see specifically skills/productivity/grill-me/SKILL.md), and his writeup at aihero.dev/my-grill-me-skill-has-gone-viral. His technique in one line: "Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree resolving dependencies between decisions one by one."

This SKILL.md does not copy his prose — the frontier/round mechanic, the stakes-tiered depth model, and the progressive-disclosure file layout are written in this repo's own words, and the dynamic/token-efficient/assistive departures above are a deliberate redesign, not an accident to fix. Per the MIT license's notice-retention term, no substantial portion of his text is reproduced here; this note is the credit that term calls for.

Status

Implemented. The cheap eval (evals/cheap/checks.sh) defends the invariant mechanically: the two-axis stakes table, the ❓/➡️ question-and-recommendation format, the "Not this" differentiation naming both adjacent plugins, and a line-count budget on SKILL.md so the lean-file design can't silently regress.

License

MIT