grill · diff
git:20260719.6892f8c to git:20260804.018ac9e
0 added, 2 removed. Audit A to A.
What this update changes in the checks
No check changed: every fixed check gives the same answer before and after this update.
A tripped check is a label, not a verdict: read the matched text and the lines below. This week's updates that newly trip a check.
---
name: grill
description: Structured questioning to reach shared understanding before action. Walks the decision tree one branch at a time, checks memory for already-answered questions, saves every answer as a lesson. Use when user wants to think through a plan, align on approach, poke holes in a design, or figure out decisions before committing. Triggers include "before we start", "think this through", "what am I missing", "poke holes", "help me think/decide", "let's align", "interview me", "grill me", "challenge this", "what should I consider", "what would you ask".
---
## Activation Behavior
**Explicit triggers** (user clearly wants structured interrogation):
- "grill me", "interview me", "challenge this"
- Activate immediately with a mode banner:
> 🔥 **Grill Mode** — I'll walk through one concern at a time.
- Then ask the first question.
**Ambiguous triggers** (user might want a dump OR structured questioning):
- "poke holes", "what am I missing", "think this through", "before we start", "help me think/decide", "let's align", "what should I consider", "what would you ask"
- Show a confirmation gate BEFORE activating:
> Want me to grill you on this one question at a time, or just give you the full critique?
- >
- > [OPTIONS: Grill me one at a time | Just give me the full critique]
- If user picks "full critique": respond with a comprehensive dump (no skill mode).
- If user picks "grill": activate with the mode banner and begin structured questioning.
## Structured Questioning Mode
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. For each question, provide your recommended answer.
Ask the questions one at a time.
If a question can be answered by exploring the codebase or memory, explore instead of asking.
Before asking ANY question, check memory for prior answers (`memory_recall` /
`memory_list`). If already decided, confirm rather than re-ask:
> "Previously you decided X. Still holds?"
After each answer, save it as a lesson:
```
memory_remember(rule="<decision>", category="knowledge", scope="workspace")
```
If the plan is simple and clear, say so and move on. If user says "enough" or "just do it", summarize decisions and proceed.