claude-code-ultimate-guide/examples/commands/plan-start.md
Florian BRUNIAUX 7bda706da2 feat(v3.32.0): Plan-Validate-Execute Pipeline — 3-command AI-first workflow
New workflow for production teams: dynamic agent teams, ADR learning loop,
automated execution from PRD to merged PR.

Added:
- guide/workflows/plan-pipeline.md — complete workflow guide (philosophy,
  non-prescriptive AI-first, No Bandaids first principles, ADR learning loop,
  CLAUDE.md 120-line discipline, /clear context reset, cost profile)
- examples/commands/plan-start.md — 5-phase planning with 12-agent dynamic
  pool (trigger-based selection, Tier 0 Solo → Tier 4 Full Spectrum,
  planning-coordinator synthesis, auto-transition to validate)
- examples/commands/plan-validate.md — 2-layer validation (structural inline +
  8 specialist agents), ADR-aware auto-fix (Bucket A ~95% auto-resolve,
  Bucket B human input → new rule), issue persistence in metrics JSON
- examples/commands/plan-execute.md — worktree → TDD scaffold → level-based
  parallel agents → drift detection → quality gate → smoke test → PR squash
  merge → post-merge metrics → cleanup
- examples/agents/planning-coordinator.md — Opus synthesis agent: merges
  multi-agent reports into coherent task graph, resolves conflicts via ADR
  precedence, verifies plan completeness before output
- examples/agents/integration-reviewer.md — Opus runtime validator: connection
  params, async/sync consistency, env var completeness, library API
  correctness (WebFetch), OTEL pipeline validation

Updated:
- machine-readable/reference.yaml — 16 new indexed keys
- CHANGELOG.md — v3.32.0 entry with 6 detailed items
- VERSION, README.md, guide/cheatsheet.md, guide/ultimate-guide.md — bumped to 3.32.0

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-06 17:24:26 +01:00

6.7 KiB

name description
plan-start 5-phase planning command: PRD analysis, design review, technical decisions, dynamic research team, metrics. Produces a complete implementation plan + ADRs before any code is written.

Plan Start — 5-Phase Planning

Analyze the request and produce a complete implementation plan through structured phases. No code is written. Every significant decision is recorded. Run /clear after this command before running /plan-validate.


Phase 1: PRD & Design Analysis

Step 1.1 — PRD Analysis

Skip if no PRD exists (refactor, infra change, bug fix).

Read all PRD files and docs/INFORMATION_ARCHITECTURE.md if present. Scan the codebase to understand current implementation status.

Surface findings in 3 buckets:

Missing requirements — acceptance criteria that are absent or incomplete Ambiguous requirements — items with multiple valid interpretations Compliance concerns — security, data privacy, API contract implications

For each finding: present options with concrete pros/cons. Discuss with user. Record every decision in the plan file under a ## Decisions section before moving on. Do not proceed past unresolved ambiguities.

Step 1.2 — Design Analysis

Skip if no UI changes are in scope.

Read: DESIGN_SYSTEM.md, existing UX ADRs, CLAUDE.md UX rules.

Produce specs for:

  • Screen inventory: new/modified screens, route placement, component reuse audit
  • State catalog: empty, loading, populated, error, and partial states for every interactive element
  • Interaction specs: user flows (happy path + alternates), focus/keyboard behavior
  • Animation specs: map each interaction to existing keyframes or specify new ones, include prefers-reduced-motion fallbacks
  • Responsive behavior: breakpoints, web/mobile divergence decisions
  • Accessibility: WAI-ARIA pattern selection, live regions, error visibility

Create Design ADRs for significant UX decisions (choice of interaction pattern, new animation convention, platform divergence). Record minor layout choices directly in the plan file.


Phase 2: Technical Analysis

Spawn 1-2 Explore agents for targeted codebase research. Run them in the background via Task tool.

While agents run, check:

  • Existing ADRs in docs/adr/ — if 3+ ADRs confirm a decision → auto-resolve without asking
  • PATTERNS.md — apply confirmed patterns directly

When agents return: present architecture decisions with 2-3 options each, concrete pros/cons, and a recommendation. Ask for user input on each unresolved decision.

For each significant decision:

  1. Create docs/adr/ADR-XXXX.md using standard Nygard format (Context / Decision / Status / Consequences)
  2. Update docs/adr/PATTERNS.md with the new observation

Phase 3: Scope Assessment

Apply trigger rules to determine which research agents are needed. Present the proposed team with justification for each inclusion.

Research agent pool:

Agent Trigger Model
code-explorer Always Sonnet
arch-researcher Changes touch 2+ architectural layers Sonnet
database-analyst Any DB schema change Sonnet
security-analyst Auth, payments, PII, RBAC, rate limiting Opus
test-analyzer Non-trivial feature (not just a bug fix) Sonnet
cross-platform-specialist Web + mobile parity required Sonnet
native-app-specialist Tasks touch mobile/native UI package Sonnet
design-system-researcher UI changes in scope Sonnet
dependency-researcher New packages being added Sonnet
devops-specialist Docker, env vars, CI/CD changes Sonnet
integration-researcher New services, libraries, OTEL config Opus
planning-coordinator Always, when 2+ agents selected Opus

Tier labels (descriptive, not prescriptive):

  • Tier 0 (0 agents): Solo — inline research, no spawning
  • Tier 1 (1-3 agents): Focused
  • Tier 2 (4-6 agents): Standard
  • Tier 3 (7-9 agents): Comprehensive
  • Tier 4 (10+ agents): Full Spectrum

Tell the user: "I recommend a [Tier N - Label] team: [agent list with one-line justification each]. Want to add or remove any agents?"

Wait for approval before Phase 4.


Phase 4: Research & Plan Creation

Tier 0: Conduct inline research. Write plan directly without spawning agents.

Tier 1+: Spawn approved agents in parallel using Task tool (run_in_background: true). For each agent, provide:

  • Its specific research scope
  • The relevant files/areas to investigate
  • The questions it needs to answer

Monitor agents via TaskOutput polling loop. Report progress: "3/6 agents complete..."

When all agents return: if planning-coordinator was spawned, send it all agent reports and have it synthesize the final plan. Otherwise, synthesize directly.

Plan file structure (docs/plans/plan-{name}.md):

# Plan: {feature-name}
Created: {date} | Branch: {branch-name} | Tier: {N}

## Summary
One paragraph: what this implements and why.

## Decisions
Decisions recorded during Phase 1 (PRD analysis).

## Architecture
ADRs created, patterns applied, architectural choices made.

## Tasks
Ordered task list with layers (1 = foundation, 2 = depends on 1, etc.)

### Layer 1
- [ ] Task A — description, files affected, acceptance criteria
- [ ] Task B — description, files affected, acceptance criteria

### Layer 2
- [ ] Task C — depends on A — description, files affected, acceptance criteria

## Test Plan
How each task will be verified. TDD tasks marked explicitly.

## Integration Verification
Smoke test commands to run post-execution (if backend/services in scope).

## Out of Scope
What this plan explicitly does not address.

Commit: plan file + ADR files + agent report manifests.


Phase 5: Finalize Metrics

Record timestamps, phase durations, agent counts, and cost estimates in docs/plans/metrics/{name}.json. Commit.


Auto-Transition

If Phase 1 produced no unresolved ambiguities and Phase 2 produced no unresolved decisions: auto-start /plan-validate without asking.

If any human discussion occurred: ask "Ready to validate this plan?" before proceeding.


Usage

/plan-start

Provide the feature description or point to a PRD file when prompted. The command handles the rest interactively.

When to Use

Use for any non-trivial feature: anything touching more than 2 files, involving architecture decisions, or where a planning mistake would be expensive to undo.

For simple changes (typos, trivial refactors): use /plan mode instead.

See Also