Grilling

Grilling

Walk down the decision tree relentlessly — one branch, one question, one resolved dependency at a time.

When to Use

Use this skill when a plan, design, or proposal has open decisions that should be resolved before work starts: stress-testing a proposal, sharpening a design, choosing among tradeoffs, pinning down dependencies, or turning vague intent into a clear goal and next action.

When Not to Use

• A written plan, PR, or design doc needs findings rather than an interview — use scrutinize. Scrutinize reviews an artifact from the outside; grilling questions its author from the inside.
• The idea is still rough and has no shape yet — use brainstorming. Brainstorming generates the options; grilling resolves them. If there is nothing to interrogate, there is nothing to grill.
• The user asked for a direct change, a quick answer, or a small well-understood edit — make the change. Interviewing here is theater.

Artifacts

• Produces: goals doc at the goals key path (on request) — see references/artifact-paths.md (default docs/goals/.md)
• Consumes: .context/project.md

Core Rule

Resolve upstream decisions before downstream details. Ask one question at a time, and offer a recommended answer when evidence supports one.

Decision Frame

Build and maintain this frame internally during the interview:

• Goal: the outcome, user, and success signal.
• Context: current system, constraints, assumptions, prior decisions, and non-goals.
• Decision: the choice being made now and the options being rejected.
• Tradeoffs: cost, complexity, reversibility, time, quality, performance, security, and maintenance.
• Dependencies: people, systems, data, tools, approvals, and sequence constraints.
• Risks: failure modes, unknowns, blast radius, and ways the plan can be wrong.
• Validation: evidence that will prove the decision worked.
• Rollback: how to undo or limit damage.

Goal and context are upstream of everything else — resolve them first.

Workflow

1. Restate the current goal in one sentence. If the goal is unclear, make that the first question.
2. Inspect available artifacts before asking questions that local files, docs, tests, logs, or supplied material can answer.
3. Pick the highest-leverage unresolved branch in the decision tree.
4. Ask exactly one question.
5. Include why it matters and a provisional recommendation when there is enough evidence.
6. Wait for the user's answer, then update the frame and move to the next dependency.
7. When major branches are resolved, summarize agreed decisions, open questions, risks, validation plan, and next action.

Question Format

Question: 
Why it matters: 
Recommendation: 

Operating Rules

• Do not ask multiple independent questions in one turn.
• Do not perform interrogation theater — every question must unlock a decision, reduce risk, or expose a dependency.
• Prefer evidence over asking when local artifacts can settle the matter.
• Keep recommendations provisional until the user confirms or corrects them.
• Do not implement or write final artifacts until core decisions are resolved or the user redirects.
• Do not create files by default. If asked to save the result, use a clear user-provided or inferred path.

Next Step

This interview is done when every open decision branch is resolved.

• If approved (all branches resolved): hand off to whichever skill triggered the interview — typically write-a-prd, write-a-story, brainstorming, or the relevant implementation skill.
• If not approved (an open branch remains): continue interviewing that branch — do not hand off with an unresolved dependency.

Source & license

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

• Author: ntwrcht
• Source: ntwrcht/blvck-skills
• License: MIT

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