skill-composer

Budget-weighted prompt execution. depth_probe frames thinking. plan_prompt budgets effort. Per-phase pauses enable midstream metathinking. audit_probe catches gaps.

Surfaces interpretations before answering. Use when user appends #? or when question has multiple valid readings. Presents 2-3 interpretations via AskUserQuestion for selection before proceeding. For deep investigation, use mode-think.

Side-by-side comparison of options. Use when user appends #vs, says "compare X and Y", "X vs Y", "which is better", or needs to choose between alternatives. For honest yes/no verdicts, use mode-judge instead.

Save comprehensive context to a handoff file for future sessions. Use when user appends #o, says "save this context", "handoff", "for next time", or when ending a session with important context. Creates a persistent record.

Gives honest evaluation without cheerleading. Use when user appends #j, asks "good idea?", "should we", "worth it?", or needs unbiased assessment. No validation theater. For challenging approaches specifically, use mode-challenge instead.

Triple-checks proposals and claims before execution. Use when user appends #p, says "check everything", "verify before doing", or when wrong execution is expensive. Three passes: facts, completeness, consequences. For post-hoc evidence proving, use mode-verify instead.

Ruthless clarification context. Use when #qh prefix detected. Interrogates assumptions and forces explicit answers before proceeding. For lighter touch, use mode-question-light.

Light clarification context. Use when #ql prefix detected. Adds thoughtful questions without aggressive interrogation. For ruthless clarification, use mode-question-heavy.

Two-phase research mode. Use when #research prefix detected. Phase 1: Exa search for landscape survey. Phase 2: Firecrawl skill for deep dives on promising results. For quick lookups, use firecrawl directly.

Socratic teaching mode that builds understanding incrementally. Use when user appends #l, says "teach me", "explain like", "help me understand", or wants to learn not just get an answer. For quick answers, skip this mode.

Trace code execution paths. Use when user appends #x, says "trace this", "follow the flow", "how does this get called", or needs to understand execution order. For architecture overview, use mapping-codebase instead.

Proves claims with evidence or retracts them. Use when user appends #v, says "are you sure", "prove it", "really?", or expresses skepticism. Shows evidence or admits uncertainty. For codebase-wide fact-checking, use verifying-claims instead.