Grill Graph

grill-graph

Grill — foreground, the only thing the user feels

Interview the user relentlessly about every aspect of the plan until you reach a shared understanding. Walk down each branch of the design tree, resolving dependencies one at a time. For every question, give your own recommended answer — take a position, don't just ask.

Ask one question at a time and wait for the answer before continuing. Keep it short, pointed, specific. If a question can be answered by reading the codebase, read the codebase instead of asking.

This is exactly grill-me. Nothing below is allowed to slow the rhythm or soften a question.

Track — in context, one node per turn, nothing written yet

After each answer, before you ask the next question, mentally record the resolved decision as a node following resources/node-template.md: the adopted option, every rejected option with a one-line reason, and the decision it builds on. Capture the reasons now — they are only alive in the moment — but hold the node in context. Don't touch disk.

If an answer reopens an earlier decision, don't stop to chase the consequences — drop a one-line Reopens # flag on the new node and keep grilling. It gets resolved at convergence.

Do not create decision-graph.md, wire edges, propagate staleness, or render anything mid-session.

Converge — the single end point, where everything is written and drawn once

When the user calls for convergence:

1. Write every node held in context to decision-graph.md in one shot, in order, each following

resources/node-template.md (repo root, or docs/ if it exists). This is the first time the file touches disk.

1. Read references/graph-mechanics.md and follow it. That is the only moment the rest of the

heavy work runs — wiring dependencies, propagating the Reopens flags into stale premises, rendering the graph once, and emitting spec.md. Don't load that file before now.

Convergence produces three artifacts together: decision-graph.md, decision-graph.html, and spec.md. None alone is convergence — emit all three or you haven't converged.

Source & license

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

• Author: LeaYeh
• Source: LeaYeh/skills
• License: MIT

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