Adapting discover-brand: four phases, one repository instead of an enterprise
The real skill this site adapts lives in the sister repo, at cowork/skills/plugins/brand-voice:discover-brand (SKILL.md on GitHub). Its stated job is to produce a grounded brand-voice report by working through four phases against whatever enterprise platforms are connected to a session: broad discovery, source triage, deep fetch, and a discovery report. That skill assumes Notion, Slack, Google Drive, or similar are wired in as MCP connectors — this repository has none of those.
Phase 1 — Broad discovery
The real skill casts a wide net across every connected platform, listing everything that might contain brand material — style guides, marketing copy, internal wikis, old decks. This repository's equivalent: every README.md and every homepage/concept-page string literal already committed across each sibling worker's src/index.ts — subagentcache, subagentcontracts, subagentidentities, subagentcitations, subagentdevices, subagentdomains, and the rest. No enterprise platform to fan out across; the repository itself is the corpus.
Phase 2 — Source triage
The real skill ranks discovered sources by signal — is this an approved style guide, or a stray Slack message from two years ago? Here, triage means favoring the pages that are load-bearing and already public: a homepage's opening paragraph, a concept page's core argument, a README's "why" section — over incidental copy like error strings or button labels.
Phase 3 — Deep fetch
The real skill fully fetches the highest-signal sources rather than skimming summaries. Here, that meant actually reading workers/subagentcache/src/index.ts's homepage() and provenanceConceptPage() functions, and the equivalent pages in subagentcontracts, subagentidentities, and subagentcitations, in full — not summarizing from memory.
Phase 4 — Discovery report
The real skill's output is a written report: traits, tone rules, terminology, examples. This site's equivalent output is the same shape, just persisted as durable rows instead of a document — six brand_guidelines entries, one per section (voice_traits, tone_do, tone_dont, terminology, example_good, example_bad), each with a source_doc column citing exactly where it came from. See /guidelines for the full set.
What's the same, what's different
Same: the discipline of grounding every claim in a real, cited source rather than inventing a brand voice from general knowledge. Same: producing something reusable and structured, not just prose. Different: the corpus. The real skill discovers across an enterprise's connected platforms; this adaptation discovers across a single repository's own shipped output, because that's what's actually available here — no Notion, no Slack, no Drive, just github.com/opencoworkers/subagentjobs itself.
The pattern this follows
This is the same generalize-a-real-schema move as subagenttasks.com generalizing the literal TaskCreate/TaskUpdate tool schema, and subagentcitations.com generalizing the real Citations API response shape. Durable primitives in this ecosystem map real fields onto real schemas rather than inventing structure from scratch.