docs(designbot): source materials for AuthZed brand generation

[Corey-T1000]

Draft — parked for review, not ready to merge.

Source material for a designbot: what "on-brand" means in a form a generation model can consume. Bridged from Sandworm's canon (sandworm/DESIGN.md + design/*.md) and ground-truthed against Burrow's live brand_config. Every rule traces to shipped code or an explicit brand decision.

Scope is diagrams, blog/social imagery, and mascot. Product UI is deliberately out.

The organizing idea

A generation request carries two kinds of content, and mixing them is where the expensive bugs live:

Layer	Read by	Contains
Composer rules	the text model assembling the prompt	assembly order, what may never co-occur, reference budget
Renderer facts	the image model	hexes, identity anchors, style locks

Burrow shipped a payload telling the model to "ignore character and costume, blend freely" alongside "the character MUST MATCH EXACTLY." Both correct individually; together they produced the wrong character — and nobody files a bug against output that looks deliberate.

Two routing calls worth arguing with

• Diagrams should generate SVG, not images. Sandworm diagrams are hand-rolled SVG, dense with mono-caps labels, with exact 1.5px strokes and a 7 3 dash. Image models misspell text, approximate strokes, and produce output that drifts on every regeneration.
• Headline-bearing OG cards should be templated renders, same reasoning — the headline is the entire payload.

That leaves image generation for atmospheric and character work.

What ground-truthing turned up

Measured against the live config. Nothing here is hypothetical:

• colors, vocabulary, asset_profiles are all empty; identity, voice, color_scales are null. A single 2,073-char prose field carries the entire brand, while prompt assembly reads nine metadata fields per reference — against nulls.
• mascotView is never read by selection. It's set in the UI, shown in the UI, copied onto the payload, and nothing filters or ranks on it. Camera angle only prepends a text clause. Ask for a back view and you get front-facing images plus a sentence saying "back view" — the images win.
• All six Dibs references sit at weight: 70. With a 4-slot budget and ties broken by insertion order, two never reach the model.
• Both three-quarter views are mislabeled (front / back), and one reference carries mascotView: "left-side" — not a member of the union, and the cause of a prior VIEW_CONFIG[x].label crash.
• defaultDiagramRef, defaultSceneRef, defaultObjectRef are declared in BrandConfig with no DB columns — writes appear to succeed and vanish.
• The identity prose says the eyes are "devoid of pupils or highlights." Every reference image shows a violet highlight. The spec contradicts the art on a signature feature.

designbot/staged-sql/01-fix-dibs-references.sql addresses the reference set. It is staged and unapplied — runs in a transaction ending in ROLLBACK, dry-run verified against Sandtrap.

Known gaps — deliberate, tracked separately

This landed as documentation. It does not yet follow the conventions authzed/messaging established, and it should:

1. No citation discipline (messaging requires a file path per claim + a Sources: block)
2. No .claude/ skill, agent, or /help command — nothing enforces these rules
3. No integrity agent or PASS / PASS WITH EDITS / HOLD verdict vocabulary
4. No CONTRIBUTING.md, frontmatter (title / owner / last-reviewed), or <!-- FLAG: --> convention
5. No validator — messaging ships cliche-check.mjs; the analogue here is a token checker flagging non-Sandworm hexes

Open questions for review

• Placement. Top-level designbot/ was chosen so it can graduate to its own repo the way authzed/messaging did. A sandworm/design/ spoke is the alternative, but this consumes the design system rather than describing it.
• MascotView is too coarse — one three-quarter member, two three-quarter views. The staged SQL parks the rear one on other as a compromise. Extending the union is the real fix.
• None of it is validated against a real render yet. Roughly a thousand lines of rules that have never been tested against a generation. If they don't move output, they're prose.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
design-basin	Ready	Preview	Aug 14, 2026 3:47pm
design-sandworm	Ready	Preview	Aug 14, 2026 3:47pm