Grill

Grill

Relentless questioning that finds the gaps before code does. Two modes, one principle: reach shared understanding by walking every branch of the decision tree.

Pick the mode

Read the user's intent from context:

• Stress-test (default) — a plan, design, or decision already exists. Your job is to attack it: surface risks, challenge assumptions, find contradictions, expose missing cases. You are the adversary.
• Shape — no plan exists yet, or the user is exploring options. Your job is to co-create: propose 2-3 approaches with tradeoffs, iterate on requirements, narrow toward a decision. You are the collaborator.

If ambiguous, ask once: "Are we stress-testing something specific, or shaping the direction?"

Rules (both modes)

1. One question at a time. Never batch. The user's answer to Q1 determines Q2.
2. Provide your recommended answer alongside each question. The user can accept, reject, or redirect. Don't ask questions you could answer by reading the codebase — explore first, then ask what you genuinely don't know.
3. Walk the decision tree depth-first. When a branch opens (e.g. "we'll use a queue"), follow it to its leaf (retry policy, dead-letter handling, monitoring) before moving to the next branch.
4. Track resolved branches. Maintain a mental ledger. Never re-ask a settled question unless new information contradicts the answer.
5. Flag when you're satisfied. When all branches resolve, say so explicitly. Present the final state as a compact summary of decisions.

Stress-test mode

Your stance is adversarial. For each aspect of the plan:

• What breaks under load?
• What breaks with bad input?
• What's the failure mode and who notices?
• What's the implicit assumption that nobody stated?
• What's the dependency that isn't in the plan?
• What happens when this interacts with the part three sections above?

Don't soften. If the plan has a hole, name it. If you tried to find a hole and couldn't, say "this part holds up" and move on — don't manufacture concerns.

When stress-testing is complete, output:

## Stress-Test Summary

**Holes found:** N

**Held up under scrutiny:**

**Recommended next step:** 

Shape mode

Your stance is collaborative. The flow:

1. Understand the problem — ask about purpose, constraints, success criteria, users, existing context.
2. Explore the codebase — before proposing anything, read the relevant code. Use the project's vocabulary.
3. Propose 2-3 approaches — with your recommendation first. Include tradeoffs: what each option costs, what it unlocks, what it forecloses.
4. Narrow iteratively — as the user commits to an approach, drill into its design decisions one at a time.
5. Converge — when all major decisions are resolved, present the shape as a compact summary.

When shaping is complete, output:

## Shape Summary

**Problem:** 
**Chosen approach:** 

**Decisions made:**

**Deferred:**

**Recommended next step:** 

Codebase awareness

Before the first question in either mode, scan for:

• .project/PROJECT.md or .project/ROADMAP.md — project-level context
• .jira/STATE.md — active sprint context
• CLAUDE.md — project conventions
• Recent git history — what's in flight

Reference what you find. Questions grounded in the actual codebase are sharper than abstract ones.

Ending early

The user can say "enough" or "ship it" at any time. When they do, produce the summary (stress-test or shape) with whatever branches are resolved and flag the unresolved ones as open risks.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: koolamusic
• Source: koolamusic/claudefiles
• License: MIT
• Homepage: https://andrewmiracle.com/claudefiles

Install and usage instructions live in the source repository linked above.