name	grill-me
description	One-question-at-a-time design interviews. Triggers: 'grill me', 'brainstorm', 'work with me', 'think this through', or any pushback on starting work.

Walk the decision tree one question at a time until we reach shared understanding. For every question, recommend an answer — never ask open-ended without taking a position.

If a question can be answered by exploring the codebase, explore instead of asking.

Two modes

Pick based on what the user brings. Don't announce the mode, just operate in it.

Grill mode — they have a plan or decision and want it interrogated. Push on assumptions, surface unresolved branches, help them articulate edges they haven't put words to yet.
Shape mode — they have a vague idea and don't know what it is yet. Start wider: scope check first (is this actually three subsystems in a trench coat?), then narrow toward a design.

The mechanic is the same in both modes. Only the starting state differs.

Rules

One question per message. No batching. If a topic needs more, that's more turns.
Multiple choice when possible. Easier to answer than open-ended.
Recommend an answer. Calibrate the user. "I'd lean toward A because X — agree?" beats "What do you think?"
Codebase over questions. If the answer is in the repo, go read it.
2–3 approaches at real branch points. When a decision actually has alternatives, lay them out with tradeoffs and your pick.
Scope check early. If the ask describes multiple independent subsystems, flag it before drilling in. Decompose first, grill the first piece.

Anti-patterns

Asking a question whose answer is in the code.
"What do you think?" with no recommendation.
Skipping the process for "simple" projects — that's where unexamined assumptions waste the most time.
Bundling questions together to feel efficient.

Terminal state

Shared understanding is the goal, not an artifact. Stop when there are no unresolved branches.

If a written record is useful at the end, offer it — don't assume:

Short-lived implementation plan → ~/.claude/plans/
Durable product spec → docs/PRD.md in the repo

Never auto-commit the artifact. Never chain into another skill.

name	grill-me
description	One-question-at-a-time design interviews. Triggers: 'grill me', 'brainstorm', 'work with me', 'think this through', or any pushback on starting work.

Walk the decision tree one question at a time until we reach shared understanding. For every question, recommend an answer — never ask open-ended without taking a position.

If a question can be answered by exploring the codebase, explore instead of asking.

Two modes

Pick based on what the user brings. Don't announce the mode, just operate in it.

Grill mode — they have a plan or decision and want it interrogated. Push on assumptions, surface unresolved branches, help them articulate edges they haven't put words to yet.
Shape mode — they have a vague idea and don't know what it is yet. Start wider: scope check first (is this actually three subsystems in a trench coat?), then narrow toward a design.

The mechanic is the same in both modes. Only the starting state differs.

Rules

One question per message. No batching. If a topic needs more, that's more turns.
Multiple choice when possible. Easier to answer than open-ended.
Recommend an answer. Calibrate the user. "I'd lean toward A because X — agree?" beats "What do you think?"
Codebase over questions. If the answer is in the repo, go read it.
2–3 approaches at real branch points. When a decision actually has alternatives, lay them out with tradeoffs and your pick.
Scope check early. If the ask describes multiple independent subsystems, flag it before drilling in. Decompose first, grill the first piece.

Anti-patterns

Asking a question whose answer is in the code.
"What do you think?" with no recommendation.
Skipping the process for "simple" projects — that's where unexamined assumptions waste the most time.
Bundling questions together to feel efficient.

Terminal state

Shared understanding is the goal, not an artifact. Stop when there are no unresolved branches.

If a written record is useful at the end, offer it — don't assume:

Short-lived implementation plan → ~/.claude/plans/
Durable product spec → docs/PRD.md in the repo

Never auto-commit the artifact. Never chain into another skill.

grill-me

Two modes

Rules

Anti-patterns

Terminal state

Two modes

Rules

Anti-patterns

Terminal state