THE DS-BRAIN

A design system
that remembers why.

Most design systems ship components. Ours also ships an institutional memory — a versioned vault of decisions, rules, and patterns that humans and AI agents read before touching a single pixel. Everything on this page is wired to the real vault.

0documents 0connections 17sections 91rules

One brain, two renderings

Every decision is written once — then rendered twice. As prose for people. As law for machines. Neither copy drifts, because both are generated from the same moment of agreement.

OBSIDIAN VAULT · FOR HUMANS

Prose with a point of view. Wiki-links that let a designer wander from a token question to the decision that created it. Maps of content instead of folder archaeology.

prosewiki-linksmaps of content
same decisions,
two renderings
AI-CONTEXT · FOR MACHINES

Tables, numbered rules, severity flags. Deterministic filenames, a dispatcher, registries an agent can grep. No mood. No ambiguity. No "it depends".

tablesrule numbersseverity flags
engineers & designers
AI agents (Claude Code, verifiers)

Why a brain exists

DRIFT

Docs that rot get ignored

Stale documentation is worse than none — a new contributor following it re-introduces bugs the team already fixed. The documentation contract makes updating the docs mandatory at visual sign-off: four surfaces move together, or the diff is rejected.

AMNESIA

Decisions re-argued every quarter

Without memory, every hard call comes back — same debate, new wave. The append-only decisions log records what was decided, when, by whom, and why. Settled means settled; the log is the receipt.

ONBOARDING

Months → minutes

A newcomer — human or AI — doesn't read the corpus. The dispatcher tells them exactly which files their task requires, in order, and nothing more. Day one starts with a reading list, not a scavenger hunt.

Load less, know more

Tier 0 — bootstrap · ~50 lines, loaded every session
Tier 1 — primer · ~7K tokens, the core context
Tier 2 — topic files · ~5K tokens each, loaded on demand
Tier 3 — tool inputs · registries, specs, generated data

loading the full corpus is non-conformant — the dispatcher resolves the MINIMUM file set per task

How a task flows

designer asks dispatcher 00-README loads 3–5 docs build Compliance Report fresh AI verifier PASS
update brain

Two words at the end of any session. Claude diffs everything learned, decided, or built in the conversation against the vault — and edits the right files directly. What the session knew, the system now knows. Forever.

17 sections, one spine

From entry points to machine canon — every document lives in exactly one section. Doc counts below are computed live from the manifest, not typed by hand.

The vault, as a graph

Every document is a node. Every wiki-link is an edge. Drag nodes, scroll to zoom, hover to trace a neighborhood, click to read the actual document.

0 nodes · 0 edges shown · nodes = documents · size = how referenced · this is the real vault, not an illustration

The whole vault, readable

All 378 documents, organized exactly as the brain organizes them. Markdown is lazy-loaded per section the first time you open a document from it.

378 documents · click any to read

The system remembers

Decisions are appended, never rewritten. Rules are numbered, never vibes. A sample of what the log holds:

2026-06-05W55 — Canonical Brick Wrapper: one primitive ends six waves of spacing wars (Rules 82–84)
2026-05-09W38 — Documentation Contract locked: four surfaces update together or the diff is rejected
2026-06-17W72 — Tier-0 skills precedence: the skills win on conflict
2026-06-30W110 — OI Lite pivot: killed our own playground for Claude Code
2026-04-10Rule 32 — Ant wrappers = styling + config ONLY

The registries

DEVIATIONS
14

Approved escapes

Every exception to the rules is signed and dated. Breaking a rule without an entry here isn't a deviation — it's a defect.

BRIDGES
∞→0

Legacy aliases with removal criteria

Nothing is deleted until a consumer-grep returns zero. Each bridge carries the exact condition of its own death.

FAILURE MODES
23

Silent-regression patterns

Every way the system has quietly broken before — each one recorded with the grep that catches it next time.

PAGE SPECS

Outcome contracts, signed by the Design Lead

Rules verify how something is built; a page spec verifies what got built. A build that passes every rule but fails a single spec bullet still fails — SPEC-NOT-SIGNED-OFF is a verdict, not a warning.

Nothing ships unverified

builder Compliance Report fresh-context verifier PASS / FAIL fix bottom-up, never silent

Earn your way up

Prototype code isn't held to production law — but it doesn't get to pretend it's production either. Four tiers, explicit promotion triggers, no silent graduation.

T0

Story-local

Lives inside the .stories.tsx file. Inline styles OK. Deletable tomorrow with zero ceremony — velocity is the whole point.

T1

Promotion candidate

JSDoc-marked. Still story-local, but the brain knows it's escalating — a named candidate with a watcher on it.

T2

Molecule / organism

Full 8-file contract: types, vars.css, module.css, TOKEN-DOC, stories with TokenMapping, barrel export. A real citizen with a token surface and docs.

T3

Brain-documented pattern

A named P-entry in the pattern catalog, referenced by page specs. The design system now teaches it to every future build.

T0→T1 a 2nd consumer wants it T1→T2 rule of three: 3+ consumers · stable API · clear tokens T2→T3 usage involves non-trivial decisions worth teaching

The pattern catalog

P-001 KPI strip P-002 hero stat P-003 card-internal segmented control P-004 dual-aspect card P-005 disclosure reference data P-006 top-N preview P-007 threshold-aware viz P-008 retired by W55 P-009 canonical chrome decorator

Deletion is a feature

flag early authorize explicitly delete loudly record forever

Anti-velocity

  • Premature extraction. Promoting on first use. Wait for the second consumer — most one-offs stay one-offs.
  • Dead tokens. A token surface nobody themes is maintenance debt wearing a costume.
  • 10-layer prompts. If the build plan needs ten layers, the scope is wrong — not the plan.
← ALL WORK