krishpunjabi.com ~/skills

← all commands

/grilling

/grilling plan and decide

Grill the user relentlessly about a plan, decision, or idea. Use when the user wants to stress-test their thinking, or uses any 'grill' trigger phrases.

  • author Matt Pocock
  • invocation model + user
  • prereqs None.
  • tested used daily
  • suggested model · effort opus · high

Install

Install just this, from a clone of the repo. Anything it needs comes with it.

./scripts/install-skills.sh --skill grilling   # bash, Git Bash, macOS, Linux
.\scripts\install-skills.ps1 -Skill grilling   # Windows PowerShell

Works on its own.

What it does

Interviews you, one round at a time, until a plan or decision has no unexamined branches left. Every question comes with Claude's recommended answer, so you can just say "yes" to most.

When to use

  • You have a plan, design or decision and want it stress-tested before acting.
  • You say "grill me", "grill this plan", "poke holes in this".
  • Other skills call it: consult-plan uses it to test an idea against a repo's plan.

How to invoke

  • /grilling <the plan or idea>, or just ask to be grilled; Claude auto-invokes it (model-invocable).
  • /grill-me is a user-only alias.

How it works

  1. Maps the topic as a design tree: each decision branches into the decisions that depend on it.
  2. Asks the whole frontier (every question answerable now, with prerequisites settled) in one numbered round: ❓ Q1 … then ➡️ recommended answer.
  3. Waits for your answers, recomputes the frontier, asks the next round. Questions that depend on an open question wait for a later round.
  4. Looks up facts itself (dispatching a sub-agent for filesystem/tool lookups) instead of asking you; only decisions go to you.
  5. Ends when the frontier is empty, and does nothing with the result until you confirm shared understanding.

Reads/writes no files by itself.

Design principles

  • Design tree, asked by frontier. A question is only asked once everything it depends on is settled, so you never answer a hypothetical that a later answer would undo.
  • Facts are Claude's job, decisions are yours. Anything the environment can answer gets looked up; only real choices reach you.
  • Recommend, don't just ask. Each question carries a recommended answer worded so "yes" accepts it, which keeps rounds fast.

Shared context for these principles: Matt Pocock's video.

Use cases

  • Stress-testing a design before writing code (a migration, a new service boundary, an API change).
  • Turning a vague idea into a list of settled decisions.
  • Testing a proposed change against a repo's existing plan, via consult-plan.

Tips

  • Answer by number ("1 yes, 2 no: keep sessions, 3 yes"); anything you skip stays on the frontier.
  • A fact lookup only holds back the questions that depend on it; the rest of the round is asked anyway.
  • Nothing is written down. Use grill-docs when the answers need a record.
  • It acts on nothing until you confirm a shared understanding, so say so explicitly when you're done.

Example

/grilling move auth from sessions to JWTs

Round 1 asks ~4 numbered questions (token lifetime, refresh strategy, revocation, migration of live sessions), each with a recommendation. You answer; round 2 follows up on revocation storage because you chose short-lived tokens. Stops when nothing is left open.

  • grill-me — alias.
  • consult-plan — calls this with the repo's plan as context.
  • verifier agent — may answer its questions on your behalf when authorized (see agents/README.md).

Prereqs

None.

Sample workflows

Source

Use cases & tips from the community