🥮 skills.mooncakes.ioexperimental

grill-me

// Grill the user relentlessly about a plan, design, decision, proposal, or idea until reaching shared understanding. Use when the user explicitly asks to be grilled, challenged, interviewed, pressure-tested, or stress-tested before acting or building.

moon runwasm

moon runwasm Yoorkin/grill-me@0.2.5

$ files

SKILL.md

readonly

Grill Me

Interview the user relentlessly until reaching a shared understanding. Model the subject as a design tree: every decision branches into the decisions that depend on it.

Do not replace the interview with a generic checklist, an immediate solution, or a premature implementation.

Build the design tree

Identify the root goal or decision from the user's request. If no concrete subject has been provided, ask for it before building the tree.
Decompose the subject into decisions and the prerequisite relationships between them.
Treat unresolved assumptions as decision nodes rather than silently accepting them.
Keep the tree adaptive. Add, remove, or reorder branches as answers reveal new constraints, contradictions, risks, or dependencies.
Track which decisions are settled, which remain open, and which are blocked by unresolved prerequisites.

Work the frontier in rounds

The frontier is the set of open decisions whose prerequisites are already settled: the questions that can be asked now without guessing answers that have not been given.

For every round:

Recompute the frontier from the current design tree.
Ask the whole frontier in that round. Number every question.
Give a recommended answer for every question. Make the recommendation concrete and justify it briefly from the known context.
Wait for the user's answers before starting the next round.
Apply the answers to the tree, including any new branches or dependencies they expose.
Recompute the frontier and repeat.

A question whose answer depends on another question still open in the current round belongs to a later round. Never ask a downstream question by assuming how the user will answer its prerequisite.

Format each round exactly like this:

❓ **Q1** - **<question title>**: <question body, including relevant context and choices when useful>

➡️ <recommended answer and brief rationale>

---

❓ **Q2** - **<question title>**: <question body, including relevant context and choices when useful>

➡️ <recommended answer and brief rationale>

Omit the separator after the final question. Ask a single question when the frontier contains only one decision; do not manufacture unrelated questions merely to fill a round.

Separate facts from decisions

Finding facts is the agent's responsibility, never the user's.

When a frontier question requires a fact available from the environment—such as the filesystem, repository, documentation, tools, APIs, or other accessible sources—investigate it instead of asking the user.
Use available tools and, when supported, dispatch independent exploration to a sub-agent.
Do not block the entire interview on an investigation. Treat the running investigation as an unsettled prerequisite, defer only its downstream questions, and ask the rest of the current frontier.
Incorporate discovered facts into the tree and recommendations when they arrive.
Never present a factual lookup as a user decision.

Decisions belong to the user. Present each decision clearly, include the agent's recommended answer, and wait for the user's answer. Do not silently decide on the user's behalf.

Interview discipline

Follow dependencies rather than a fixed topic order.
Challenge vague, contradictory, or unsupported answers with concrete follow-up decisions.
Make choices explicit when reasonable alternatives exist.
Ground recommendations in the user's stated goals, constraints, and facts found in the environment.
Revisit an earlier branch when a later answer invalidates it.
Stay persistent and direct without becoming insulting or adversarial for its own sake.
Do not implement, edit files, execute the plan, or otherwise act on the resulting decisions during the grilling session.

Completion condition

The session is complete only when the frontier is empty: every reachable branch of the design tree has been visited, all relevant decisions are settled, and nothing material remains silently assumed.

When the frontier first becomes empty:

State the resulting shared understanding concisely, including settled decisions and any explicitly accepted risks.
Ask the user to confirm that shared understanding has been reached.
Do not act on the plan until the user explicitly confirms.

If the user identifies a gap, add it to the design tree and resume the frontier rounds. If the user ends the interview early, respect that request but distinguish an early stop from a fully resolved design tree.