firehose> #llmops

Skill Extraction

A skill is the crystallized residue of a workflow you already proved by hand — so you build the thing first and package it last. The rule of thumb: you don't start with a skill, you end with a skill. Have an idea, get the agent to build a first version, and give it round after round of feedback until the output hits your standard and taste; only then tell the agent "take the workflow we just did and turn it into a skill," and separately design the onboarding (what context the skill must ask a user for). Packaging cold — writing SKILL.md before the workflow works — inverts the dependency: you'd be freezing a process you haven't validated. This is why the move is accessible to non-technical authors: the hard part is doing the task well and noticing your own taste, which is human work; the freeze is a single instruction to the agent.

Skill extraction is the provenance rule the authoring cluster leaves implicit. Skill-Driven Loop Development insists you pre-validate skills before looping over them and Skill Authoring Checklist governs a single skill's internals — but neither says where a skill comes from. The answer here: from a proven, hand-run workflow, extracted at the end.

Claims


Linked from