diff --git a/.claude/settings.json b/.claude/settings.json
new file mode 100644
index 0000000..09ebe53
--- /dev/null
+++ b/.claude/settings.json
@@ -0,0 +1,18 @@
+{
+  "env": {
+    "CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS": "1"
+  },
+  "extraKnownMarketplaces": {
+    "superpowers-dev": {
+      "source": {
+        "source": "github",
+        "repo": "obra/superpowers",
+        "ref": "v5.1.0"
+      }
+    }
+  },
+  "enabledPlugins": {
+    "superpowers@superpowers-dev": true,
+    "frontend-design@claude-plugins-official": true
+  }
+}
diff --git a/AGENTS.md b/AGENTS.md
index 8622be0..a7f509c 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -29,6 +29,7 @@ pnpm test:watch
 - Active plan → `docs/superpowers/plans/`
 - Architecture → `docs/architecture/` (later plans)
 - Per-package map → `packages/<name>/AGENTS.md`
+- Project skills → `.claude/settings.json` (`superpowers`, `frontend-design`)
 
 ## Rules of engagement
 
@@ -40,8 +41,15 @@ pnpm test:watch
 6. Conventional commits.
 7. No emojis in source / docs unless requested for UI.
 8. No half-broken features. Verification command required per feature.
-9. No LLM attribution in commits, PRs, or tags. Author is the human; no
-   `Co-Authored-By: <LLM>` lines.
+9. Never attribute work to an LLM. Commits, PRs, tags, and code comments list
+   humans only — never an LLM (Claude, GPT, Codex, Copilot, etc.) as author,
+   committer, `Co-Authored-By`, or in-code attribution. The work is yours.
+10. Superpowers-first. Reach for `superpowers:<skill>` (TDD, debugging,
+    brainstorming, dispatching parallel agents, requesting/receiving code
+    review, plan writing, git worktrees) before improvising a workflow.
+11. UI work pre-loads `/frontend-design:frontend-design`. Any agent on React,
+    shadcn, Tailwind, or the Electron renderer loads it on session start and
+    follows the design system it encodes. No bespoke styling that diverges.
 
 ## Progressive disclosure
 
diff --git a/CLAUDE.md b/CLAUDE.md
index b6a75c8..a827bfc 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -27,6 +27,7 @@ pnpm test:watch       # iterative dev loop
 - **Adapter SDK** → `docs/adapters/sdk.md` · package: `packages/adapter-sdk/CLAUDE.md`
 - **IPC protocol** → `docs/ipc-protocol/spec.md` · package: `packages/contracts/CLAUDE.md`
 - **Contracts package** → `packages/contracts/CLAUDE.md`
+- **Project skills** → `.claude/settings.json` (`superpowers`, `frontend-design`)
 
 ## Hard constraints
 
@@ -42,9 +43,17 @@ pnpm test:watch       # iterative dev loop
 7. **No emojis in code or docs** unless explicitly asked for a UI surface.
 8. **No half-broken features.** If you can't define a verification command, it
    isn't a feature yet — it's a question.
-9. **No LLM attribution in commits.** Commits must list a human as author and
-   contain no `Co-Authored-By:` lines naming an LLM (Claude, GPT, Codex, etc.).
-   The work is yours; tools assist. Same rule for PR descriptions and tags.
+9. **Never attribute work to an LLM.** Commits, PRs, tags, and code comments
+   must list humans only — never an LLM (Claude, GPT, Codex, Copilot, etc.) as
+   author, committer, `Co-Authored-By`, or in-code attribution ("Generated by
+   …", "Written with …", etc.). The work is yours; tools assist.
+10. **Superpowers-first.** Reach for `superpowers:<skill>` (TDD, debugging,
+    brainstorming, dispatching parallel agents, requesting/receiving code
+    review, plan writing, git worktrees) before improvising a workflow.
+11. **UI work pre-loads `/frontend-design:frontend-design`.** Any agent touching
+    React, shadcn, Tailwind, or the Electron renderer loads the skill on
+    session start and stays consistent with the design system it encodes. No
+    bespoke styling that diverges from it.
 
 ## Layout (current)