Claude-Code-Workflow-Guide

Real-World Team Workflows — How Production Teams Ship with Claude Code

These are documented workflows from teams that have shipped production software using Claude Code. Each is adapted from public community contributions.


1. The Superpowers Workflow (188k★)

Origin: The most-starred Claude Code workflow in the community.

Philosophy: Claude does the implementation. You do the architecture and review. The dividing line is clear.

┌─────────────────────────────────────────────────────┐
│              SUPERPOWERS WORKFLOW                    │
├─────────────────────────────────────────────────────┤
│                                                     │
│  1. BRAINSTORM (you + Claude, plan mode)            │
│     "What are 3 approaches to this feature?"        │
│     Choose one. Commit to it.                       │
│                                                     │
│  2. WORKTREES (isolate the work)                    │
│     git worktree add ../feature -b feature/name     │
│                                                     │
│  3. DETAILED PLAN (plan mode, 15 min)               │
│     File-level plan with checkpoints                │
│                                                     │
│  4. SUBAGENT IMPLEMENTATION                         │
│     Spawn impl agent with the approved plan         │
│     Agent implements + runs tests                   │
│                                                     │
│  5. HUMAN REVIEW (you)                              │
│     Review the diff. Not the process, the output.  │
│                                                     │
│  6. SUBAGENT FIX (if review has issues)             │
│     Spawn fix agent with specific issues list       │
│                                                     │
│  7. MERGE                                           │
│     PR → merge → clean up worktree                 │
│                                                     │
└─────────────────────────────────────────────────────┘

Key insight: Brainstorm sessions use expensive models + plan mode. Implementation uses Sonnet. You only spend Opus on decisions, not execution.

In practice:

# Step 1-3: Planning session
claude --model claude-opus-4-7
/plan "implement user notification system"
# ... approve the plan ...

# Step 4: Implementation subagent
# (done automatically by the /plan command's approved output)
# Agent runs in the worktree

# Step 5: Your review
git -C ../feature diff main..HEAD | claude "review this diff for issues"

# Step 7: Merge
gh pr merge --squash
git worktree remove ../feature

2. The BMAD Method (Product → Production)

Origin: A product-first methodology that treats Claude as a full development team.

Philosophy: Product clarity drives code quality. Vague specs produce vague code.

Phase 1: PRODUCT BRIEF
  User: "I need a feature for X"
  Claude (Product Mode): "Here's a product brief: [users, problem, solution, metrics]"
  You: Approve or iterate

Phase 2: PRD (Product Requirements Document)
  Claude: Generates full PRD from brief
  You: Review, add constraints

Phase 3: ARCHITECTURE
  Claude (Architect Mode): Proposes technical approach
  You: Approve design

Phase 4: EPIC BREAKDOWN
  Claude: Breaks PRD into 3-5 epics
  You: Prioritize, define MVP

Phase 5: SPRINT PLANNING
  Claude: Breaks epic into 1-3 day tasks
  You: Assign, adjust

Phase 6: IMPLEMENTATION
  Claude: Codes each task using TDD
  You: Review PR per task

Phase 7: CODE REVIEW
  Claude (Reviewer Mode): Comprehensive review
  You: Final approval

Phase 8: RETROSPECTIVE
  Claude: "What did we learn? What should change in CLAUDE.md?"
  You: Update standards

Commands for BMAD:

/brief "feature idea"      # → product brief
/prd                       # → PRD from brief
/architect                 # → technical approach
/epics                     # → break into epics
/sprint                    # → sprint tasks
/implement [task]          # → code the task
/retro                     # → retrospective

3. The gstack Workflow (14-Stage Enterprise Flow)

Origin: A startup team shipping production features with a structured 14-stage process.

Philosophy: Nothing ships without passing through quality gates. Automation handles the mechanical parts.

Stage 1:  OFFICE HOURS     — weekly sync to choose next feature
Stage 2:  CEO REVIEW       — 1-line feature description approved
Stage 3:  ENGINEERING REVIEW — technical feasibility check
Stage 4:  DESIGN REVIEW    — UX/UI spec approved
Stage 5:  SPEC             — detailed spec written
Stage 6:  PLAN             — implementation plan approved
Stage 7:  CODE             — implementation (TDD)
Stage 8:  SELF REVIEW      — /review command, fix issues
Stage 9:  QA               — manual QA against spec
Stage 10: SECURITY         — /audit command
Stage 11: PERFORMANCE      — load test for user-facing features
Stage 12: STAGING          — deploy to staging, verify
Stage 13: PRODUCTION       — deploy with monitoring
Stage 14: METRICS          — verify feature metrics after 1 week

What makes this powerful: Each stage is a command or checklist that Claude can run. The process is encoded in .claude/commands/, not just in a Google Doc.

# Most stages map to a command:
/spec "feature"      # Stage 5
/plan                # Stage 6
# (code)             # Stage 7
/review              # Stage 8
/audit               # Stage 10
/deploy-staging      # Stage 12
/deploy-prod         # Stage 13
/metrics-check       # Stage 14

4. The Spec Kit Workflow (97k★)

Origin: A specification-driven workflow that uses Claude’s reasoning to stress-test requirements.

Philosophy: Ambiguous specs are bugs. Fix them before they become code.

Step 1: CONSTITUTIONAL
  "Write a one-paragraph 'constitution' for this feature — 
  the principles it must satisfy above all else."

Step 2: SPECIFY  
  Claude generates: user stories, acceptance criteria, edge cases
  
Step 3: CLARIFY
  Claude plays devil's advocate:
  "What could a developer misinterpret about this spec?"
  "What user behavior isn't covered?"
  
Step 4: PLAN
  Implementation plan based on approved spec
  
Step 5: TASKS
  Spec broken into granular tasks with tests
  
Step 6: IMPLEMENT
  Code to the spec, not to assumptions
  
Step 7: VERIFY-SPEC
  Claude checks implementation against original spec
  Reports: satisfied / missing / extra behavior

The constitutional step is the differentiator:

# Example constitution:
"This checkout flow must be: 
(1) forgiving — never lose a user's cart data, 
(2) transparent — always show users what they're paying and why, 
(3) recoverable — every error state must have a clear next step."

These principles guide all implementation decisions when the spec is ambiguous.


5. The Solo Indie Hacker Workflow

Origin: Synthesized from multiple solo developers shipping fast with Claude Code.

Philosophy: Ship fast, iterate often. Quality matters but so does speed.

Morning kickoff (10 min):
  Fresh session
  "Read the last 5 commits. What should I work on today for [goal]?"
  Choose one task. Start it.

Pomodoro coding (25 min):
  One focused session per Pomodoro
  Compact between Pomodoros
  Commit at end of each Pomodoro

Ship daily:
  Every working day ends with a PR or deployed change
  /ship before ending session

Weekly retrospect:
  "Read my git log for the last week. What patterns do you see?"
  Update CLAUDE.md based on what worked/didn't

Key commands for solo devs:

/kickoff        # daily orientation
/pomodoro       # 25-min focused sprint  
/ship           # end of day shipping
/retrospect     # weekly learning review

6. The Debugging War Room

Origin: Pattern used by on-call engineers for production incident investigation.

Philosophy: Production incidents need speed AND precision. Structure prevents panic mistakes.

ALERT FIRES:
  1. /incident "description of alert"
     → Claude reads recent commits, relevant code, checks for obvious causes
     → Returns: LIKELY CAUSE + CONFIDENCE + IMMEDIATE ACTIONS

  2. TRIAGE (5 min):
     Is this P0 (system down) or P1 (degraded)?
     P0: page the team immediately, then investigate
     P1: investigate first, page if not resolved in 30 min

  3. INVESTIGATE:
     /debug "error details + stack trace"
     → Systematic hypothesis testing
     → One fix attempt per hypothesis

  4. FIX:
     Minimal change only
     No refactoring during incidents
     Run tests before deploying

  5. VERIFY (10 min post-fix):
     Monitor error rate, latency, key metrics
     Confirm resolved

  6. POST-MORTEM (24h later):
     /post-mortem "incident summary"
     → Root cause analysis
     → What test would have caught this?
     → Update runbooks

The /incident command:

<!-- .claude/commands/incident.md -->
# Incident Response: $ARGUMENTS

Immediate triage for: $ARGUMENTS

1. Check recent git commits (last 10):
   `git log --oneline -10`
   
2. Check for relevant recent changes to affected service

3. Read the relevant service files — focus on recently changed code

4. Check if any known issues match this pattern

Output:
- LIKELY CAUSE: [your best hypothesis]
- CONFIDENCE: [high/medium/low]
- IMMEDIATE CHECK: [one command to confirm/deny]
- IF CONFIRMED: [fix approach]
- BLAST RADIUS: [what else could be affected]

What All These Workflows Have in Common

Looking across all these teams’ approaches, the shared patterns are:

  1. Plan before code — every team has an explicit planning phase
  2. Verification gates — nothing moves forward without checking it
  3. Spec clarity — ambiguous requirements are resolved before implementation
  4. Incremental shipping — features ship as vertical slices, not horizontal layers
  5. Commands encode the process — the workflow is .claude/commands/, not in someone’s head
  6. CLAUDE.md evolves — every mistake that gets fixed gets codified

The teams that get the most from Claude Code treat their .claude/ directory as valuable infrastructure, not an afterthought.