Brand Context Protocol
Writing great BCPs
A meta-skill for authoring, reviewing, and continuously improving a Brand Context Protocol package. Think of it like a skill that teaches other skills, applied to brand.md.
Simon Edstrom pointed Encoded at Matt Pocock’s writing-great-skills: a reference that improves existing skills rather than only creating new ones. This page is that idea, mapped onto BCP. Use it in onboarding, or as a prune pass on a live package. It is not the spec, and it is not a substitute for owner approval of claims.
Why it exists
Agents already have a prior. A great BCP beats it.
A BCP exists to wrangle determinism out of a stochastic system. The goal is not the same sentence every time. The goal is the same process: fetch the root, pick the daughter, keep the condition on the figure.
Before the first compile
Interview against this bar. Empty files that look complete fail. Eight files with one job each pass.
After the package is live
Revise is the product. Run the review procedure. Prune sediment. Internal agents should change behavior, not only ChatGPT.
The holy-shit test
Same fact. Different file.
Our trucks are the most efficient on the market, with industry-leading electric range and a five-star safety rating.
Approximately 2 percent lower total energy consumption, depending on specification and operating conditions. Up to 720 km on selected 6x2*4 tractor units. Five-star Euro NCAP Haulage result for the G- and R-series tested. A rating is a test outcome, not a guarantee.
Worked example: the unofficial Scania Dynamic demonstration at scania-dynamic.pages.dev, compiled from public press material. Scania did not author or approve it. That honesty is part of the bar.
Eight files, one job each
Mixing jobs is how sediment starts.
Wayfinding
Always-loaded pointer. Identity, provenance, fetch map. Not a brochure. Not a claims dump.
How it sounds
Distinctive positive targets. Leading words. Generic “be authentic” is a no-op.
Slop filter
Cadences this brand never uses. Copied generic ban-lists are sediment.
Decision tests
A value that cannot veto a draft is a poster.
Safety with an alternative
Hard nos. Soft nos with conditions. Pair every elephant with what to do instead.
Single source of truth
Every figure lives here, with its condition, forbidden shortening, and source. Nowhere else as a restated number.
Third-party rails
How others may describe you. If it restates voice.md, cut one.
Tokens, not adjectives
Hex, type, assets, forbidden modifications. “Modern and clean” is not a token.
Review procedure
Done is checkable.
- Fetch the live root. Confirm the handle.
- Fetch every declared daughter. A 404 is a broken pointer.
- Name each file’s job in one sentence. If you cannot, split or cut.
- Extract every number and superlative. Diff against claims.md.
- Replace no-op voice and elephant boundaries with positive targets and allowed alternatives.
- Kill duplication. One source of truth per meaning.
- Date or drop caches of live site facts.
- Decide every silence: official or demonstration, claimed or verified, whose photo.
- Generate three artifacts. Score them against voice, claims, and boundaries. Failures are bugs in the BCP.
Completion criterion: every remaining figure has a condition. The three artifacts pass. The root is shorter than when you started, or you can say why it could not be.
Install
Agents read the markdown. Humans read this page.
Point an agent at the skill file. Keep this HTML as the overview layer.
curl -s https://writing-great-bcps.pages.dev/SKILL.md curl -s https://writing-great-bcps.pages.dev/GLOSSARY.md
Related: the open spec, the public authoring interview, Registry MCP for the Scania demonstration handle scania-f4a6118b.