leadgrowgtm/dev-skills/grill

grill

Input: a written plan file. If the plan only exists in chat, write it to a file first.

Rounds 1-5, unattended

Each round works one frontier, not a list you thought of. Which questions belong in this round, and which must wait, is set by references/design-tree.md.

Each round:

1. Send references/adversary-prompt.md to an adversary that cannot edit:

a general-purpose subagent told it is read-only, or, when the codex CLI is installed, codex exec -s read-only "<the prompt>".

1. Settle every item yourself from source (read the file, run the command, check

the docs). Each ends in one state:

• Resolved: the answer, with file:line.
• Plan changed: the attack landed; edit the plan, note what changed.
• Needs user: intent, priority, spend, or outside knowledge only. A fact

you could look up never goes here.

1. Stop early when a round raises nothing new.

Round 6, one batch to the user

Send every "needs user" item in ONE message:

Q<n> - <title>: <question, with options> Why source could not settle it: <one line> Recommendation: <pick and why>

Nothing gets implemented while one is open. Round 7 runs only if the answers change the plan's structure; it is the ceiling.

Verdict, appended to the plan

## Grill verdict - YYYY-MM-DD

• Rounds: N (self) + N (user)
• Attacks raised: N - resolved: N, changed the plan: N, escalated: N
• Plan changes: one line each
• Accepted risks: what is knowingly unhandled, and why that is fine
• Unresolved: anything open, and what would settle it

Recount the numbers before writing them. An empty "accepted risks" on a real plan means the grill was theatre.

After the grill

A settled plan becomes a spec the next session can act on without re-interviewing anyone. That step is references/to-spec.md: pure synthesis, no new questions. If gaps remain, go back to Round 1 rather than writing around them.