SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsarcanea-meta

arcanea-meta

Agents & models

View repository

Updated 10-15 21:47

About this skill

Arcanea Meta

The system describing itself with the same primitives it uses to build everything else. This skill is the load-bearing meta layer — the place to start when you don't know where something lives.

Mental model

Arcanea is not one repo. It is an ecosystem of ~90 repos at github.com/frankxai, plus four installed plugins, plus a hierarchy of skills/commands/agents/hooks at three scopes:

GLOBAL  (~/.claude/)               ← user-scope, every project sees this
PROJECT (<repo>/.claude/)           ← per-repo overrides + additions
PLUGIN  (~/.claude/plugins/cache/)  ← marketplace-installed (read-only)

Most failures come from forgetting which scope a thing lives in. Default rule: if it's reusable across repos, put it global. If it's Arcanea-specific, put it project. Never duplicate across both.

Canonical locations (verified 2026-04-15, see references/canonical-locations.md for full table)

Concern Canonical Why
Memory ~/.claude/projects/C--Users-frank-Arcanea/memory/ Auto-loaded via MEMORY.md every turn
ACOS source ~/agentic-creator-os/ Has FRANK_DNA.md + memory.db (irreplaceable)
ACOS runtime ~/.claude/acos/ What the harness actually invokes
Arcanea monorepo ~/Arcanea/ Web app + packages + lore + book
Validation contract ~/Arcanea/CLAUDE.md § Cached-Belief Validation Protocol Loads every turn
Hooks (project) ~/Arcanea/.claude/hooks/ 9 hooks, project-scope
Hooks (global) ~/.claude/hooks/ 3 hooks, user-scope
Skills scattered (see references/installed-stack.md) Migration ongoing
Source clones (for absorption) ~/sources/ Pre-integration staging

When in doubt, read references/canonical-locations.md.

What's installed (verified 2026-04-15)

  • 4 Claude plugins (typescript-lsp, vercel, superpowers v5.0.7, superpowers-lab v0.4.0)
  • 2 marketplaces (anthropics/claude-plugins-official, obra/superpowers-marketplace)
  • 42 global agents, 84 global commands (~/.claude/{agents,commands}/)
  • ~100+ skill directories (~/.claude/skills/ + ~/Arcanea/.claude/skills/)
  • 9 project hooks + 3 global hooks (no overlap by name yet)

Update by re-running the verification commands in references/installed-stack.md. Treat the numbers as stale after 7 days.

Meta-rules (the contract)

  1. Disk-first, memory-as-history. Any claim about current state (versions, ship status, file paths, architecture, deploy state) requires same-turn verification OR explicit "unverified, from memory:" prefix. Memory authoritative only for intent, strategy, preferences, decision history. (See references/validation-contract.md.)

  2. Canonical wins. When two copies exist, the canonical (per canonical-locations.md) is truth. Drift copies get DEPRECATED_DELETE_AFTER_VERIFY.md markers, never silent edits.

  3. Absorption is provenance. Claiming "absorbed pattern X from repo Y" requires entry in references/absorption-log.md with source URL, commit SHA, integrated-at path, date, verifier. No vague "we took inspiration from" allowed.

  4. Scope before content. Before adding a skill/command/agent/hook, decide scope (global vs project vs plugin). Wrong-scope additions create the duplication that broke ACOS.

  5. Progressive disclosure. SKILL.md stays under 3K words and contains the mental model + workflow skeleton. Depth lives in references/. Skills that violate this become unreadable monoliths (the original acos-meta did this — it teaches progressive disclosure but didn't use it).

How to add a skill

1. Decide scope:
   - Global (multi-repo use)  → ~/.claude/skills/<name>/
   - Project (Arcanea-only)   → ~/Arcanea/.claude/skills/<name>/
2. Create SKILL.md with YAML frontmatter (name + description with trigger phrases)
3. Keep SKILL.md under 3000 words. Mental model + workflow + key principles.
4. Create references/ subdirectory for depth content.
5. Test: invoke the trigger phrase, verify activation.

How to add a command

1. ~/.claude/commands/<name>.md (global) or project-scope equivalent
2. YAML frontmatter: description, optional argument-hint
3. Body is a markdown workflow — no code, pure instructions
4. Reference skills/agents by name; the harness wires them
5. Test: /<name> [args]

How to add an agent

1. ~/.claude/agents/<name>.md
2. Frontmatter: description with usage examples, model preference (optional)
3. Body = the agent's system prompt (role, capabilities, constraints, output format)
4. Invoke via Agent tool with subagent_type=<name>

How to add a hook

1. Decide event: SessionStart | UserPromptSubmit | PreToolUse | PostToolUse | Stop | SessionEnd | Notification
2. Write ~/.claude/hooks/<name>.sh (global) or ~/Arcanea/.claude/hooks/ (project)
3. Register in settings.json (or settings.local.json) under "hooks": { "<Event>": [...] }
4. Test: trigger the event, check the hook ran (logs at ~/.arcanea/sessions/current/)

Hook count discipline: stay under 12 total. When two hooks share >30% logic, extract a shared lib.

How to absorb a repo

The contract (no exceptions):

1. git clone --depth=1 <url> ~/sources/<name>/
2. Capture HEAD SHA, list 3-5 specific patterns/files you want
3. Add entry to references/absorption-log.md (source, SHA, patterns, target paths, date, verifier)
4. Integrate the patterns at the target paths (separate session/PR)
5. Update absorption-log.md status: cloned → integrated → verified
6. After integration verified, decide: keep clone for reference or rm -rf ~/sources/<name>

Vague "inspired by" or "we took ideas from" is not absorption. Either it's logged with provenance or it didn't happen.

Validation contract integration

This skill is the home for the validation contract designed in 2026-04-15 session:

  • L0 Authority Registry → .arcanea/AUTHORITY.md (deferred, Week 2)
  • L1 Memory frontmatter → migration plan in references/validation-contract.md (deferred, Week 2)
  • L2 CLAUDE.md protocol → root CLAUDE.md § Cached-Belief Validation Protocol (Week 1 ship target)
  • L3 /verify-status skill → deferred Week 3
  • L4 Stop-hook claim audit → deferred Week 4

Status of Week-1 ship: designed, Guardian-approved, not yet shipped to disk.

When this skill activates

Use when:

  • "What's installed?" / "What do we have?"
  • "Where does X live?" / "Where should I add Y?"
  • "How do I add a skill/command/agent/hook?"
  • "How does absorption work?" / "Did we absorb repo X?"
  • "What's the canonical version of Z?"
  • Onboarding a new tool, repo, or capability
  • Auditing for drift, duplication, or stale claims

Don't use for: domain work (code, content, design, lore). Route to domain skills/agents.

References

  • references/ecosystem-map.md — full repo inventory + spot-check freshness
  • references/installed-stack.md — verified plugin/skill/agent/command/hook counts + how to re-verify
  • references/canonical-locations.md — which copy wins for every concern, archival plan for drift
  • references/absorption-log.md — strict-format provenance table
  • references/validation-contract.md — full 5-layer contract spec + rollout plan