striatum — a rule library for Claude Code

A single source of truth for how Claude should behave on recurring kinds of work: brand and logo design, technical writing, marketing copy, professional communication, and whatever groups get added next. Install once, use in every project.

How it works

Rules live as skills. A skill is a folder with a SKILL.md whose description tells Claude when to load it. Only the description sits in context permanently; the body loads when a matching task shows up. That keeps coding sessions cheap and still gives brand work the full guidance.

Three other layers sit around the skills:

• agents/ holds reviewer personas. Run one after the work is done to check it against the rules from a fresh context.
• AGENTS.md holds the universal rules that apply on every turn, with the coding rule blocks rendered inline so any tool that reads AGENTS.md gets the full text. CLAUDE.md is a one-line @AGENTS.md import for Claude Code. Projects can import either.
• rules/ holds tool-agnostic rule blocks for things that must be always-on and readable by any agent, such as the git workflow. They are synced into AGENTS.md files and the project-starter template, and imported by Claude. See rules/coding/README.md.

Rules are generic, facts are per project

No skill in this repo hardcodes a brand name, palette or tone. Every skill reads project facts from a brand profile in the project it is running in (brand.md at the project root, or .claude/brand.md). The template is in templates/brand-profile.md. One library therefore serves every client and product.

Layout

striatum/
  README.md
  package.json                     npm manifest for striatum; bin striatum
  LICENSE                          MIT
  bin/striatum.mjs                 CLI: sync, emit, install, each running a script below
  AGENTS.md                        universal rules + coding blocks, importable by projects
  CLAUDE.md                        one-line @AGENTS.md import for Claude Code
  .claude/rules/                   path-scoped rules emitted from blocks (Claude Code)
  CHANGELOG.md
  skills/
    rule-authoring/                how to write a rule in this repo (meta-rule)
    brand-logo-design/
    brand-logo-generation/         LLM image-model loop: brief, rounds, scoring, vector
    writing-technical/
    writing-marketing/
    writing-professional-comms/
    coding-commit-pr/              commits, PR descriptions, pre-push checks
    coding-branch-cleanup/         safe pruning incl. squash-merged branches
    coding-dependency-change/      add/upgrade/remove dependencies safely
      SKILL.md                     triggers, principles, checklist  (<300 lines)
      references/                  long material, loaded on demand
      examples/                    before/after pairs
      scripts/                     deterministic helpers (e.g. contact sheets)
  agents/
    brand-reviewer.md
    copy-editor.md
    logo-critic.md                 blind scorer for logo candidates
    code-reviewer.md               conventions review of a diff or PR
  rules/
    coding/                        portable rule blocks + rendered CODING.md
  templates/
    brand-profile.md               facts each project supplies
    logo-brief.md                  plan + iteration log for a logo project
    SKILL-template.md              house format for new rules
  scripts/
    install.sh                     symlinks skills and agents, imports coding rules
    sync-rules.mjs                 writes rule blocks into AGENTS.md / .jinja targets
    check.sh                       every CI check: script syntax, rules in sync, frontmatter
    check-rules.mjs                skill and agent frontmatter, SKILL.md line budget
    emit-rules.mjs                 writes path-scoped blocks as Claude, Cursor or Copilot rule files
    lib/blocks.mjs                 block loader shared by sync-rules and emit-rules
  .github/workflows/ci.yml         runs scripts/check.sh on push and pull request
  .github/workflows/publish.yml    publishes to npm with provenance on a v* tag

Install

Personal (every tool on this machine). scripts/install.sh installs for each tool it finds. It is safe to re-run: links it did not create are left alone and reported, and synced files are brought up to date, so run it again after editing a rule block.

| Tool | What it installs | | :- | :- | | Claude Code | skills and agents linked into ~/.claude/, ~/.claude/rules/coding linked to the emitted path-scoped rules, and an import of rules/coding/CODING.md in ~/.claude/CLAUDE.md | | Codex | skills linked into ~/.agents/skills/, and the always-on coding blocks appended once to ~/.codex/AGENTS.md between sync markers, then re-synced on every run | | Cursor | nothing extra; it reads ~/.claude/skills/ and ~/.agents/skills/ on its own |

./scripts/install.sh

Any project, from npm. The library is published as striatum. Install it as a dev dependency and run the striatum CLI against the project's own files; the blocks come from the installed package, so no clone is needed.

pnpm add -D striatum
pnpm exec striatum sync AGENTS.md
pnpm exec striatum emit claude=.claude/rules cursor=.cursor/rules

Add striatum sync --check AGENTS.md to the project's check script so CI fails when a dependency bump changes a rule and AGENTS.md was not re-synced. Skills and agents ship in the package too: symlink node_modules/striatum/skills/<name> into .claude/skills/. A global install (pnpm add -g striatum) can run striatum install, which is scripts/install.sh pointed at the installed package.

One project only. Symlink or copy the skills you want into that project's .claude/skills/, and the agents into .claude/agents/.

A team. Add a .claude-plugin/plugin.json at the root and publish the repo as a plugin marketplace. The folder layout above is already plugin-shaped; nothing moves.

Use in a project

1. Copy templates/brand-profile.md to the project as brand.md and fill it in.
2. Optionally add @~/dev/striatum/AGENTS.md to the project's CLAUDE.md to import the universal rules.
3. Ask for the work in plain language. The matching skill loads on its own.
4. For a second opinion, ask for the reviewer: "have the brand reviewer check this".

Add a rule

Ask Claude to add a rule in this repo and the rule-authoring skill takes over. Doing it by hand: copy templates/SKILL-template.md into skills/<domain>-<topic>/SKILL.md, write a pushy description, then principles, then a checklist, then one before/after example. Keep SKILL.md under 300 lines and push detail into references/. Add a line to CHANGELOG.md.

Checks

scripts/check.sh runs everything CI runs: bash -n and shellcheck on shell scripts, node --check on Node scripts, a parse of Python scripts, a check that AGENTS.md, rules/coding/CODING.md and .claude/rules/ match their blocks, and a check that every skill and agent has frontmatter with a name matching its path, a description, and a SKILL.md under 300 lines. The workflow in .github/workflows/ci.yml runs the same script on every push to main and every pull request.

pnpm check

Release

Bump version in package.json, add the CHANGELOG.md entry, merge, then tag the merge commit v<version> and push the tag. .github/workflows/publish.yml re-runs the checks, verifies the tag matches the version, and publishes to npm through trusted publishing with provenance. Never move an existing tag; cut a new version instead.

Planned rule groups

Brand system (type, color, imagery, voice matrix) · product naming · UX writing · accessibility · release notes · help center · transactional email · social · executive summaries · slide decks · engineering conventions · document types (README, ADR, runbook, PRD) · legal and claims · localization · research reports.