agent-skills

This skill should be used when the user asks to "extract skill candidates", "find out what's worth extracting into skills", "find skill candidates in my AI history", "which of my workflows should become skills", "find the workflows I never encoded", "scan my AI conversations for repeatable routines", or wants to know which of their recurring routines are worth becoming reusable AI skills. Reads the local histories of Claude Code, Claude Cowork, Codex CLI, and Cursor, distills them into two small text files (nothing installs, nothing leaves the machine), asks what the user actually does, and answers in the chat, in the user's language, with the patterns found in their own requests: named, counted, quoted, plus proposed skill names. Falls back to plain conversation when history is thin. It proposes candidates only, never builds them. Not for code search (Grep/Glob) or grounding a claim in evidence — candidates are user-confirmed, never auto-clustered.

This skill should be used when the user asks to "edit my video", "edit my recording in Descript", "run the video edit pipeline", "descript edit my video", "remove retakes from my video", or wants to turn a raw Descript talking-head recording into the finished long-form YouTube edit. Drives the Descript MCP Underlord agent through the validated pipeline - duplicate-first safety, script-diffed retake removal with human cut approval, scoped filler cleanup, Studio Sound and Eye Contact, brand stat cards and lower-third from the brand package, 1.28x zoom punch-ins with Smart Transition glides, verification after every destructive pass. Not for audio podcasts - use podcast-edit for that. Not for generating vertical shorts - use personal-brand-youtube-shorts-generation for that. Not for writing the teleprompter script - use personal-brand-youtube-script-generation for that.

This skill should be used when the user asks to "check my company brain", "grade my company brain", "run the company brain completeness check", "how complete is our knowledge base", "assess our knowledge base structure", "audit our knowledge repository structure", "set up a company brain", or "recommend a starter structure for our knowledge base". Grades the structural completeness of a company or personal knowledge repository — the place where projects, concepts, and conventions live — across six dimensions plus additional checks (mode 1), or recommends a generalizable starter structure when no structured knowledge base exists yet (mode 2). Runs at a repository or workspace root in Claude Code or Claude Cowork. Not for reviewing source-code quality, not for writing or editing documentation content, and not for auditing a single document's prose.

This skill should be used at the start of a session when the user asks to "become Fable 5", "become-fable-5", "work like Fable", "prime the session", "set the operating discipline", or "do the work don't narrate it". Loads a small set of high-level working habits — build only what the task needs, do the work instead of announcing each step, plan once and finish the job, fix small in-scope problems as you go, and never trade verification for speed. Not for wrapping up a session — use meta-session-handover. Not for sharpening a task's objective — use meta-intent-hardening. Not for diagnosing a specific past mistake — use meta-learn-from-mistakes.