Voice Dump To Skill

Skills look easy until you write one. The YAML description has to match every plausible phrasing of a request without firing on the wrong ones. The instructions have to be specific enough that Claude can actually follow them — "write engagingly" is useless, "open every section with the most surprising fact, not the most logical starting point" is a rule that changes behaviour. And the expert's actual judgement — the "it depends on..." moments where they decide between options — has to be encoded as explicit criteria, not handed off to "use your judgement."

The extraction works in four phases. Phase one mines the dump for the four things that make a skill non-trivial: core process steps (often scattered, need reordering), decision points (the most valuable parts — where expertise actually lives), quality criteria (what good output looks like vs bad), and anti-patterns (what the expert explicitly says not to do, often the most instructive). Phase two maps that onto the SKILL.md structure. Phase three engineers the description for fuzzy matching across primary triggers, secondary phrasings, and disambiguation from similar skills. Phase four runs a quality check: every major instruction has a WHY annotation, no vague language, decision points have explicit criteria, anti-patterns are detectable.

It refuses to fabricate expertise. If you provide only a topic — "make a skill about negotiation" — with no process, examples, or anti-patterns, the skill does not invent a skill from general knowledge. That defeats the point of capturing your expertise. Instead it produces 5-7 targeted questions designed to trigger expert narration ("walk me through the last negotiation you won — what did you do differently?") rather than abstract prompts that get abstract answers. Memorable phrases from the expert — the vivid, specific one-liners — get preserved verbatim because they're often the most instructive parts of the source.

Output is the SKILL.md itself, in a fenced code block, ready to save. Alongside it: an extraction summary (depth, confidence, gaps), a self-audit checklist, and three suggested test cases (simple, complex, edge) to validate the skill before you ship it. If the dump covers three or more distinct processes, the skill refuses to create one mega-skill — focused skills outperform sprawling ones, the 80-200 line target exists for a reason, and you get back the dominant skill plus a list of remaining skill candidates with notes on what additional input each one needs.