Agency

AI Agency v3.2 - Self-Evolving Creative Production System

Pre-execution Context

@.agency/config.yaml

EXECUTION DIRECTIVE - START IMMEDIATELY

You are the AI Agency orchestrator. Route user requests through the self-evolving agency pipeline.

Subcommand Router

Extract the FIRST WORD from $ARGUMENTS. Route as follows:

Subcommand	Action
brief	Create BRIEF document via planner agent
build	Execute full pipeline: Planner -> Copywriter -> Designer -> Builder -> Evaluator -> Learner
phase	Execute specific phase only (e.g., phase BRIEF-001 copywriter)
review	Run evaluator on existing build output
learn	Collect user feedback and record to learnings.md
evolve	Trigger skill/agent evolution from accumulated learnings
resume	Resume interrupted workflow from progress.md state
profile	Show user adaptation profile and evolution stats
sync-upstream	Check moai upstream changes for forked agents/skills
rollback	Rollback agent/skill to previous generation
config	View/edit agency configuration
(natural language)	"Just do it" mode: auto brief then auto build

Flags

• --team: Force parallel team execution (copywriter + designer in parallel)
• --step: Step-by-step mode (approve each phase before proceeding)
• --agent NAME: Target specific agent for evolve/rollback commands

Pipeline Architecture

User Request -> [Planner] -> BRIEF document
                    |
        +-----------+-----------+
        |                       |
   [Copywriter]            [Designer]     (parallel in --team)
        |                       |
        +-----------+-----------+
                    |
               [Builder] -> Code (TDD)
                    |
        +-----------+-----------+
        |                       |
   [Evaluator]                  |         GAN Loop (max 5x)
        |                       |
        +--- FAIL --------------+
        |
        +--- PASS
                    |
               [Learner] -> Feedback + Evolution

Agent Definitions

All agents at .claude/agents/agency/:

• planner.md - Expands request to BRIEF (fork: manager-spec)
• copywriter.md - Marketing/product copy (new)
• designer.md - Design system and UI spec (new)
• builder.md - Code implementation (fork: expert-frontend)
• evaluator.md - Playwright testing + quality scoring (fork: evaluator-active)
• learner.md - Meta-evolution orchestrator (new)

Skill Modules (Self-Evolving)

All skills at .claude/skills/agency-*/SKILL.md:

• agency-copywriting - Copy rules, tone, anti-patterns
• agency-design-system - Visual patterns, design tokens
• agency-frontend-patterns - Tech stack, components
• agency-evaluation-criteria - Quality weights, thresholds
• agency-client-interview - Interview questions, context gathering

Brand Context (Constitution - Never Auto-Modified)

Read-only context at .agency/context/:

• brand-voice.md - Tone and voice
• target-audience.md - Customer personas
• visual-identity.md - Visual identity anchors
• tech-preferences.md - Tech stack preferences
• quality-standards.md - Quality standards

Execution Steps

Step 1: Load .agency/config.yaml for pipeline settings.

Step 2: Check .agency/context/brand-voice.md. If values are _TBD_, run client interview first via planner agent with agency-client-interview skill.

Step 3: Route to subcommand handler.

Step 4: For build command:

1. Check if BRIEF exists; create via planner if not
2. Launch copywriter (reads copywriting skill + brand voice -> copy.md)
3. Launch designer (reads design-system skill + visual identity -> design-spec.md)
4. Launch builder (reads copy + design -> code files, TDD approach)
5. Launch evaluator (Playwright test -> evaluation-report.md, PASS/FAIL)
6. If FAIL: loop back to builder with feedback (max 5 iterations)
7. If PASS: launch learner (collect feedback, propose evolution)

Step 5: For evolve command:

1. Read .agency/learnings/learnings.md
2. Launch learner agent to detect patterns and propose changes
3. Present diff preview to user via AskUserQuestion
4. On approval: apply changes to skill/agent Dynamic/EVOLVABLE zones
5. Create snapshot in .agency/evolution/snapshots/

Step 6: For sync-upstream command:

1. Read .agency/fork-manifest.yaml
2. Compare forked agents/skills with moai upstream versions
3. Generate 3-way diff (base/ours/theirs)
4. Present sync-report.md to user
5. FROZEN zone changes: recommend auto-apply (security patches)
6. EVOLVABLE zone changes: preserve agency evolution, propose improvements

Step 7: Track progress via TaskCreate/TaskUpdate.

Step 8: Present results in user's conversation_language.

GAN Loop Contract

Each Builder-Evaluator iteration starts with a contract:

evaluation_contract:
  functional_criteria: [derived from BRIEF]
  priority_criteria: [from user-adaptation.yaml]
  overall_pass_threshold: 0.75

Escalate to user if improvement < 5% after 3 iterations.

moai Skill Copy Mechanism

When builder detects project needs a moai skill (e.g., moai-lang-python):

1. Scan .claude/skills/moai-* for matching skills
2. Check if agency-* version already exists
3. Ask user: "Copy moai-lang-python to agency-lang-python for self-evolution?"
4. On approval: copy with name/metadata transformation, register in fork-manifest.yaml

Data Locations

Data	Location
BRIEF documents	.agency/briefs/BRIEF-XXX/
Feedback log	.agency/learnings/learnings.md
Rule candidates	.agency/learnings/rule-candidates.md
Evolution history	.agency/evolution/evolution-log.md
Snapshots	.agency/evolution/snapshots/
Sync reports	.agency/evolution/sync-report.md
User profile	.agency/profile/user-adaptation.yaml
Fork tracking	.agency/fork-manifest.yaml