SaaS Pipeline Coordinator

Coordinate, never execute. Read project state, route to the correct skill, track progress, delegate all work.

• Explicit invocation only - activate only when invoked
• Stack-agnostic - read tech stack from project's CLAUDE.md (## Tech Stack). If absent, ask before phase 4
• Progressive disclosure - load references/phase-N-*.md only when entering a phase
• No duplication - delegate to skills, never reimplement their work
• Desync safety - verify actual file existence, not just status.json claims
• Skill-aware delegation - MANDATORY: every subagent prompt must include a === READ FIRST === block with relevant skills (see Skill Injection Rule)

State Management

Initialize

When .claude/saas-project/ does not exist, create it with status.json:

{
  "project_name": "<ask user>",
  "created_at": "<ISO-8601>",
  "updated_at": "<ISO-8601>",
  "current_phase": 1,
  "stack": { "frontend": "<from CLAUDE.md or null>", "backend": "<from CLAUDE.md or null>" },
  "phases": {
    "1-idea": { "status": "pending", "started_at": null, "completed_at": null, "outputs": [] },
    "2-business": { "status": "pending", "started_at": null, "completed_at": null, "outputs": [] },
    "3-product": { "status": "pending", "started_at": null, "completed_at": null, "outputs": [] },
    "4-architecture": { "status": "pending", "started_at": null, "completed_at": null, "outputs": [] },
    "5-design": { "status": "pending", "started_at": null, "completed_at": null, "outputs": [] },
    "6-implementation": { "status": "pending", "started_at": null, "completed_at": null, "outputs": [] },
    "7-deployment": { "status": "pending", "started_at": null, "completed_at": null, "outputs": [] },
    "8-marketing": { "status": "pending", "started_at": null, "completed_at": null, "outputs": [] }
  }
}

Update

After each significant action, update updated_at, phase status/timestamps, outputs array, and current_phase.

Router

Execute on every invocation:

Step 1: Load State

Read .claude/saas-project/status.json. If missing, initialize first.

Step 2: Classify Intent

Step 3: Auto-Route

phase = phases[current_phase]
IF pending   → read references/phase-N-*.md, invoke first skill
IF in_progress → find next pending skill in phase, invoke it
IF completed → advance current_phase, start next

Skill Injection Rule

MANDATORY for every subagent dispatch. Always run the scan — never skip it.

Before dispatching any agent (Task tool or Agent Team):

1. Scan ~/.claude/skills/ — match skill names/descriptions to the agent's task technology and type
2. Select up to 4-5 relevant skills
3. For each selected skill, read its SKILL.md and check for a Reference Guide table or references/ directory. Include references whose topics match the agent's task
4. Prepend a === READ FIRST === block listing each skill's SKILL.md AND its matching references. If no skills match, dispatch without the block — not every task needs skills

=== READ FIRST ===
Read: ~/.claude/skills/{skill-a}/SKILL.md
Read: ~/.claude/skills/{skill-a}/references/{matching-ref}.md
Read: ~/.claude/skills/{skill-b}/SKILL.md
=== END READ ===

Example — Flutter frontend implementer (skill with references):

The flutter-expert SKILL.md has a Reference Guide table. Task involves Riverpod state and navigation → include those references:

Task("Implement Task 2: User dashboard screen

=== READ FIRST ===
Read: ~/.claude/skills/flutter-expert/SKILL.md
Read: ~/.claude/skills/flutter-expert/references/riverpod-state.md
Read: ~/.claude/skills/flutter-expert/references/gorouter-navigation.md
Read: ~/.claude/skills/flutter-expert/references/widget-patterns.md
Read: ~/.claude/skills/test-driven-development/SKILL.md
=== END READ ===

=== TASK ===
You are implementing Task 2: User dashboard screen
[... full task text, context, requirements ...]
=== END TASK ===
")

Example — backend implementer (skill without references):

nodejs-backend-patterns has no reference files, so only SKILL.md is needed:

Task("Implement Task 3: Auth module

=== READ FIRST ===
Read: ~/.claude/skills/nodejs-backend-patterns/SKILL.md
Read: ~/.claude/skills/test-driven-development/SKILL.md
=== END READ ===

=== TASK ===
You are implementing Task 3: Auth module
[... full task text, context, requirements ...]
=== END TASK ===
")

When using subagent-driven-development: its implementer-prompt.md does NOT include skill injection. You MUST prepend the === READ FIRST === block before the template content — the block goes at the very top, before "You are implementing Task N."

For the full scanning algorithm, see references/skill-injection-protocol.md.

Dashboard

Display on status requests:

╔══════════════════════════════════════════════╗
║          PROJECT: {project_name}             ║
║          Stack: {frontend} + {backend}       ║
╠══════════════════════════════════════════════╣
║  Phase 1 - Idea          ████████████ 100%   ║
║  Phase 2 - Business      ████████░░░░  67%   ║
║  Phase 3 - Product       ░░░░░░░░░░░░   0%   ║
║  Phase 4 - Architecture  ░░░░░░░░░░░░   0%   ║
║  Phase 5 - Design        ░░░░░░░░░░░░   0%   ║
║  Phase 6 - Implementation░░░░░░░░░░░░   0%   ║
║  Phase 7 - Deployment    ░░░░░░░░░░░░   0%   ║
║  Phase 8 - Marketing     ░░░░░░░░░░░░   0%   ║
║                                              ║
║  Current: Phase 2 → Next: startup-biz-models ║
╚══════════════════════════════════════════════╝

Calculate percentage: skills completed / total skills in phase.

The 8 Phases

When entering a phase, read its reference file for detailed workflow.

Gateway Skills

Always invoke these before their dependent work:

Utility Skills (always available)

find-skills (discover/install), skill-creator (create custom), dispatching-parallel-agents (orchestrate teams)

Agent Teams

Three parallelism opportunities via dispatching-parallel-agents. Always offer sequential as alternative.

Before dispatching agents, apply the Skill Injection Rule. Each agent gets skills matching its domain (frontend tech skills for frontend agent, backend tech skills for backend agent, plus methodology skills like TDD).

1. Phases 4+5: Architecture + Design

When project has separate frontend/backend stack:

Agent A: backend-architect → API design, data models
Agent B: ui-ux-pro-max → UI architecture + design system
Sync: align API contracts with UI data needs

2. Phase 6: Frontend + Backend

When API contracts are defined:

Agent A: Frontend → implement UI against contracts (mock responses)
Agent B: Backend → implement API endpoints matching contracts
Sync: integration testing

3. Phase 8: All Marketing

All 5 skills are fully independent:

Agent A: landing-page-copywriter
Agent B: email-sequence
Agent C: marketing-ideas
Agent D: conversion-optimization-expert
Agent E: remotion-best-practices

On team acceptance: read relevant references/, use dispatching-parallel-agents to set up, give each agent project context + phase requirements.

Auto-Install Missing Skills

On every invocation, verify skills for current and next phase are installed.

FOR each skill in current_phase:
  CHECK ~/.claude/skills/{skill-name}/ exists
  IF missing → read references/install-commands.md, inform user, offer install command, block phase if declined

Stack-Specific Skill Discovery

Discover stack-matching skills dynamically. Never hardcode stack-to-skill mappings.

Trigger at:

1. Project initialization (after reading stack from CLAUDE.md)
2. Phase 4 start (before architecture work)

READ stack from CLAUDE.md
FOR each technology in stack:
  CHECK ~/.claude/skills/ for matching skill
  IF not found → RUN: npx skills find {technology}
  PRESENT results, OFFER install with -g -y

Non-Linear Navigation

• Skip ahead: allowed if target phase's required inputs exist (phase 4 needs design doc, phase 6 needs architecture + contracts, phase 8 needs .claude/product-marketing-context.md)
• Go back: always allowed, never deletes previous outputs
• Partial phases: track individual skill completion; resume from last incomplete skill, not phase start

FOR each skill in phase.skills:
  IF output in phase.outputs → skip
  ELSE → next skill to execute