Skills & subagents
Skills
Skills are the authoring entry points. Three produce documentation from a source; one reviews it.
| Skill | Purpose |
|---|---|
lore:figma-to-doc | Generate docs from Figma design files. |
lore:brief-to-doc | Generate docs from briefs, PRDs, or user stories. |
lore:site-to-doc | Document a live product by observing it in a browser. |
lore:doc-reviewer | Validate documentation against the Definition of Done. |
Subagents
Subagents are workers the skills delegate to. They keep the main session's context clean by doing heavy or noisy work in isolation and returning a compact summary.
| Subagent | Role |
|---|---|
lore:doc-validator | Read-only validator that audits a document against the Definition of Done and returns a pass/fail report. Producer skills run it before delivery. It treats the source census adversarially — as a set of claims to falsify, not a report to read — re-probing sources and testing receipts, because the census was written by the agent whose work it is checking. |
lore:figma-extractor | Heavy Figma extraction — comments, annotations, the frame inventory, and frame images — returned as a distilled summary. |
lore:site-explorer | Heavy live-site exploration — drives the browser, captures screenshots, and records exact UI text — returned as a compact summary. |
lore:doc-reviser | Applies the validator's findings. It receives the report as someone else's work order and may edit only the files and sections each finding names, all findings of a round in one batch, one exact occurrence at a time. It has no ability to fetch and cannot create files, so the one kind of defect it cannot introduce is a fabricated source. A fix it cannot apply exactly — because the evidence contradicts it, or because it would need to touch something the finding did not name — comes back unapplied rather than guessed at. |
Why the fixer is not the writer
Revising a page in the same session that wrote it is where documentation quietly breaks: the reviser exists so the agent that produced a claim is not the one deciding which part of it to change, and so a round's fixes go in together rather than one at a time. Findings are routed by kind — text fixes to the reviser, a missing receipt back to source collection (a fixer that cannot fetch must never "resolve" one), and anything needing a product decision to you.
If two consecutive rounds keep finding defects the previous round introduced, the run stops and asks you instead of opening a third.
How they fit together
A producer skill orchestrates a run: it records what you asked for, gathers sources (delegating heavy extraction or browsing to a worker subagent) into a receipted source census, writes the documentation, and then calls lore:doc-validator to self-verify against the Definition of Done before delivery. If that review blocks, the findings go to lore:doc-reviser as one batch and exactly the files it edited are re-reviewed. The enforcement hooks run independently on top — checking the shape of what was written as files are written, and, at the end of the turn, that the evidence behind it actually exists. Self-verification and deterministic enforcement are deliberately separate: the skill can be wrong about its own work, and the hooks read artifacts it did not author.