grill

Pre-capture interrogation. Pressure-tests a plan or decision before it's frozen into an artifact, then routes to the right capture skill. Writes nothing, touches nothing under .task/.

See the grill guide for the full walkthrough.

Usage

/task:grill [<context>]

Input — $ARGUMENTS, optional. A topic or free-form context to grill ("the retry design", "whether to shard the queue"). Empty → grills the plan being discussed in the current chat.

What it does

• Frames the target in a sentence or two.
• Resolves facts itself (from the codebase) and spends questions only on genuine decisions.
• Grills one AskUserQuestion fork at a time — each with 2–4 concrete options and an honest recommendation.
• Keeps a decision ledger: {the decision} — because {the reason}.
• Ends with a pre-mortem kill-shot question, then prints the full ledger.

Typical depth is 3–7 questions.

Config

No config gate, no setup. grill is the one skill that neither reads nor writes anything under .task/ — it runs in a fresh, unconfigured project. Dialog mirrors the language of your chat.

Output

A decision ledger printed as chat text (there is no file), followed by a routing footer naming exactly one next skill:

## Decision ledger
1. {decision at full specificity} — because {the load-bearing reason}
2. …

→ Next: /task:to-plan — one task with the approach nailed down; capture Description + Plan.

Does not

• Write any file, anywhere — serializing the ledger is the to-* skills' job.
• Batch questions — one fork per prompt, always.
• Implement, refactor, or "just fix it" mid-grill.
• Agree by reflex — if your leaning is the weaker call, the recommendation says so.