Challenge me

One interview skill for a document, plan, idea, design, architecture, or anything else that needs shared understanding. Act like a manager: critical questions, then brainstorming, then a locked map of decisions. Stop there unless the user asks to write a plan artifact or tickets.

Research notes: research-synthesis.md. Question shape: round-format.md, templates/round.md.

Principles

1. Facts vs decisions. Look up what the repo or docs already answer. Only the user decides trade-offs. See facts-vs-decisions.md.
2. Numbered question + recommended answer, then wait. Never dump an interrogation without a rec. Default one-at-a-time; switch to a frontier round when the user asks or after calibrate.
3. Attack the work, not the person. Intensity is a dial. Thermonuclear is off unless they opt in (intensity.md).
4. Past the first three. Do not stop at the obvious answers; force at least one non-obvious option, then pick. Not a multi-agent loop.
5. Stop at shared understanding. Do not execute a refactor, invent a stack, or write docs/superpowers/plans/*.md unless the user asks. Then hand off (peers below).
6. Language-agnostic. The round format works in any language; do not load a second language pack.
7. Design gate (brainstorm). Nothing gets built until the user approves the design brief. See design-brief.md.

Step 0: pick a mode

Do this first; do not announce it. Load only that mode file plus the core files it names.

Precedence: explicit mode → artifact (docs present → grill-docs; "too big"/"hidden work" → challenge; "just an idea"/no design yet → brainstorm; resume later or too big for one session → decision-map; "make this prompt better" → sharpen) → default grill. Combine grill + grill-docs when they bring a plan and domain docs. Run calibrate whenever they say softer, harder, batch, or one-by-one; keep the setting for the session.

Loop mechanics: interview-loop.md. Frontier rounds: frontier.md. Brainstorm extras: approaches.md, play-back.md, design-brief.md.

Domain tracks (optional)

Subject is not a mode. After the mode, load at most one stub if the artifact is clearly in that world. Index and how to add a track: domains/README.md.

Peers (do not duplicate)

Do not absorb do / make-plan orchestrators. If they want subagent execution, point at write-plan then handoff.

Layout

challenge-me/
├── SKILL.md
├── references/core/      # loop, frontier, facts vs decisions, intensity, round rules, mindmap, research, approaches, play-back, design brief
├── references/modes/     # grill, grill-docs, challenge, brainstorm, decision-map, sharpen
├── references/domains/   # optional stubs + contribution README
└── templates/            # round.md, decision-map.md, context.md, design-brief.md

Templates: round.md, decision-map.md, context.md, design-brief.md.

Old skill name

Related skills

• codebase-learner: questions need facts about the current system.
• second-opinion: the artifact is code or a diff, not a plan.
• ready-spec: the settled design is a backend or API change that needs a full spec.
• super-tech-blueprint: the open question is which stack to pick.
• write-plan: the grilled design is settled and needs a plan.
• create-tickets: the plan is settled and needs tickets.
• safe-refactor: the plan is a refactor.
• design-system-kit: the design under review is a UI or a design system.
• handoff: save open questions for the next session.

Done when

• The matching mode ran; only needed references were loaded.
• Every asked question had a number and a recommended answer; the user made the decisions.
• Facts cited a file, doc, or "unknown — looked, not found."
• Shared understanding is written (answers, glossary/ADR updates if grill-docs, a decision map if they asked, or an approved design brief after brainstorm).
• No implementation-plan file, refactor, or invented stack unless they explicitly asked for that handoff.