grill-with-docs

Procedure

1. Run /grilling for the interview loop, second-brain sweep, and capture-at-close.
2. Run /domain-modeling for domain work — challenge against the glossary, sharpen fuzzy language, stress-test with scenarios, cross-reference against code, update CONTEXT.md inline, and offer ADRs.

Four-pass discipline — must complete before declaring the plan ready

Conversational grilling alone is not sufficient. Before declaring a plan ready (whether for /to-spec, /to-tickets, ready-for-human, or ready-for-agent), the grilling agent must complete all four passes below. Skipping any pass and falling back to "the user said it only touches X" is the failure mode this discipline exists to prevent.

1. Trace, don't list

For every flow the work touches, follow the request path through the code. Enumerate every branch encountered (if/else, switch, polymorphic dispatch). Each branch is a candidate surface for the change. In a shallow-module codebase there is no useful module boundary to list — you trace by behaviour, not by directory.

2. Consumer search

For every field, type, endpoint, or behavioural contract being changed, grep for reads. Do not trust "this is only consumed by X" from the user — verify by search. Record what you found in the brief's Key interfaces section.

3. CONTEXT.md sketch if missing