A public, normative User-Preferences Specification — the coding and working-style preferences for collaborating with AI coding agents, written as an RFC-style spec.
The spec is the canonical source (the what). It is the single source from which a website payload and an LLM-readable concatenation are generated, and from which internal tooling derives (the how).
a6b8-spec/
├── data/ # refs.manual.json (single source for version + URLs) + JSON schema
├── scripts/ # 4 generators + audit-spec-quality (the test gate)
├── spec/v0.1.0/ # the spec itself: 00-overview + 01..13 chapters + README index
├── generated/ # GIT-TRACKED build output (payload, manifest, refs.resolved, llms.txt)
└── personas/ # internal tone guide (not published)
npm install
npm run build # generate-refs → generate-docs-payload → generate-manifest → generate-llms-txt
npm test # audit-spec-quality: intro prose, ## Related, no internal-reference leaksgenerated/ is committed: the website consumes it directly. Editing files under
generated/ by hand is pointless — they are overwritten on the next build. Edit the
sources under spec/v0.1.0/ and data/ instead.
Chapters use RFC-2119 / BCP-14 keywords (MUST / SHOULD / MAY). The conformance block lives
in spec/v0.1.0/00-overview.md.
MIT — see LICENSE.