Acos Meta

ACOS Meta

ACOS describes itself using the same primitives it uses to build everything else. This skill is ACOS about ACOS.

System Architecture

acos-intelligence-system/
├── .claude-plugin/
│   └── plugin.json              ← Plugin manifest (v11.0.0)
├── .mcp.json                    ← MCP server registry
├── CONNECTORS.md                ← Connector category map
├── skills/                      ← Domain expertise (subdirs, progressive disclosure)
│   ├── [skill-name]/
│   │   ├── SKILL.md             ← Lean main file (<3K words)
│   │   └── references/          ← Deep content, fetched on demand
├── commands/                    ← Slash commands (one .md per command)
├── hooks/                       ← Lifecycle automation (SessionStart, Stop, etc.)
├── docs/                        ← Strategy documents
└── README.md                    ← Entry point

The 3 Design Principles (Drawn from knowledge-work-plugins)

1. Progressive Disclosure

SKILL.md contains the mental model and workflow skeleton. Details live in references/. Claude loads the lean summary and fetches references only when needed. This keeps context efficient without sacrificing depth.

2. Connector Agnosticism

Skills reference ~~categories, not vendor names. The .mcp.json maps categories to specific tools. Swap tools without touching skill content.

3. Commands as Workflows

Commands are fully-specified workflows in markdown — trigger, input gathering, decision logic, output structure, follow-up options. No code. Claude interprets and executes.

How ACOS Auto-Routing Works

The /acos command is the entry point. It routes based on keyword detection:

User request
    |
    ├── AI architecture keywords    → Technical Architect agent
    ├── Content/writing keywords    → Content Engine
    ├── Music keywords              → Music Producer
    ├── Visual/image keywords       → Visual Creation Council
    ├── Deploy/build keywords       → DevOps Pipeline
    ├── Research keywords           → Deep Research swarm
    └── Complex/multi-file          → Full swarm (5+ agents)

The hook system (hooks/skill-activation-prompt.sh) enhances routing with pattern matching before Claude processes the request.

Adding a New Skill

1. Create skills/[skill-name]/SKILL.md
2. Add YAML frontmatter: name, description (include trigger phrases)
3. Write lean main content (<3K words) covering: overview, core concepts, workflow, key principles
4. Create skills/[skill-name]/references/ for detailed content
5. Register in skills/skill-rules.json if using activation matching
6. Test: mention trigger phrases and verify activation in session

SKILL.md frontmatter template:

---
name: skill-name
description: "One-sentence description. Include trigger phrases like: what actions activate this skill, what topics it covers."
---

Adding a New Command

1. Create commands/[command-name].md
2. Use the standard command structure:
   • YAML frontmatter: description, argument-hint (optional)
   • > See CONNECTORS.md reference
   • ## Workflow with numbered steps
   • Input gathering, tool use, output format, follow-up options
3. No code logic — pure markdown workflow
4. Test: run /[command-name] and verify execution

Command frontmatter template:

---
description: What this command does in one sentence
argument-hint: "<optional argument description>"
---

Adding a New Agent

1. Create .claude/agents/[agent-name].md
2. Define: role, capabilities, tools, escalation path
3. Reference from orchestration commands or swarm topology
4. Test via Task tool: Task(subagent_type="[agent-name]", prompt="...")

Debugging the Hook System

ACOS has 15 hooks across 6 lifecycle events. When hooks behave unexpectedly:

1. Check audit trail: cat .claude-flow/audit.jsonl | tail -20
2. Check circuit breaker state: cat .claude-flow/circuit-breaker.json
3. View learning metrics: cat .claude-flow/metrics/learning-status.json
4. Run monitor: npm run monitor (real-time hook dashboard)

Hook event map:

SessionStart    → session-start.js + starlight-bridge + todo-continuation restore
UserPromptSubmit → skill-activation-prompt.sh
PreToolUse      → quality-gate + circuit-breaker
PostToolUse     → post-tool-track.js + audit-trail
Stop            → stop-finalize.js + todo-continuation save + learning-hooks
PreCompact      → context preservation

Intelligence Score System

ACOS tracks its own intelligence score across sessions:

View with /acos-score.

ACOS × knowledge-work-plugins

ACOS v11 integrates patterns from the knowledge-work-plugins ecosystem:

ACOS contributes back to knowledge-work-plugins:

• creator/ plugin — creator-specific domain (content, visual, music)
• Pattern: quality gates in visual and content creation
• Pattern: music prompt engineering pipeline

See references/v11-architecture-decisions.md for the full integration rationale.

Horizontal Substrates (composable into every pillar)

These are NOT pillars — they are reusable substrates that every pillar can consume. Each ships as its own OSS repo with its own contracts.

How to extend with a new substrate

1. Author master spec at docs/superpowers/specs/YYYY-MM-DD-<name>-design.md.
2. Pressure-test with /starlight-board (per the board-before-tag invariant).
3. Build agent team + skill + command in .claude/.
4. Extract as standalone repo frankxai/<name> under MIT (mirror Library OS / Prompt Hub pattern).
5. Register here (add row to the table above).

Prompt Hub composition specifics

@prompt-conductor (Opus composer, top-level entry) routes to 12 specialists via 8 canonical flows:

• flow-design / flow-optimize / flow-evaluate — builds, refines, scores prompts
• flow-harvest / flow-curate — imports OSS patterns, maintains library
• flow-introspect / flow-profile — IFS Cartographer + psychometric instruments (maps not unburdens; crisis-routing-as-code at lib/prompt-hub/crisis-routing.ts)
• flow-knowledge-base — designs RAG ingestion + retrieval prompt pairs

Red Team gates every publish flow. Voice gate (lib/voice/frankx-voice.ts) checks every output before return.

Adding a new lab specialist when a new model family drops: author .claude/agents/prompt-<lab>-specialist.md, add to AgentName + Lane types in lib/prompt-hub/types.ts, add to Conductor's lab-specialist selection table, register <lab> row in repos/prompt-library/taxonomy/lanes.yaml. No new repo, no new command — the Hub absorbs new labs through specialist addition only.

Master spec: docs/superpowers/specs/2026-05-13-prompt-hub-design.md.