Autonomous Workflow
Execute complete feature development cycles autonomously — from task intake through tested PR delivery — using isolated Git worktrees.
CRITICAL: Before Starting ANY Work
You MUST complete these steps IN ORDER before writing any code:
Step 1: Detect Workflow Mode (MANDATORY)
Analyze the task scope to determine the workflow mode:
| Mode | Criteria | Artifacts Required |
|---|---|---|
| Full | 4+ files OR complex/architectural | YES - MANDATORY |
| Lite | 1-3 files AND simple/straightforward | No |
When in doubt, choose Full Mode.
Step 2: Plan Artifact Content (Full Mode ONLY)
For Full Mode, you will need these artifacts. Do NOT create the files yet — they must be created inside the worktree after Phase 2, not on the main branch.
| File | Purpose | Created |
|---|---|---|
plan.md | Implementation strategy, decisions, requirements, progress log | After Phase 2 |
walkthrough.md | Final summary for PR delivery | Phase 6 |
plan.md is the single source of truth. It must be comprehensive enough that a new Claude session can execute from it alone without the original conversation.
DO NOT create artifact files on the main branch.
Step 3: Announce Mode Selection
State your mode selection explicitly:
"This is a Full Mode task (affects 5+ files). Creating .gw/{branch-name}/ artifacts after worktree setup."or
"This is a Lite Mode task (2 files, simple fix). Proceeding without artifacts."
Prerequisites: gw CLI Installation
Before Phase 2 (Worktree Setup), verify the gw CLI is installed:
which gwIf gw is NOT installed
STOP and prompt the user to install gw. The workflow cannot proceed without it.
Installation options (present to user):
# Via npm (recommended)
npm install -g @gw-tools/gw
# Via Homebrew (macOS)
brew install mthines/gw-tools/gw
# Via pnpm
pnpm add -g @gw-tools/gwAfter installation, set up shell integration:
# For zsh (add to ~/.zshrc)
echo 'eval "$(gw install-shell)"' >> ~/.zshrc
source ~/.zshrc
# For bash (add to ~/.bashrc)
echo 'eval "$(gw install-shell)"' >> ~/.bashrc
source ~/.bashrcVerify installation:
gw --version
gw --helpThen initialize gw in the repository (if not already done):
gw init
gw init --auto-copy-files .env,secrets/ --post-checkout "npm install"Once gw is installed and configured, resume the workflow from Phase 2.
Rules
| Rule | Description |
|---|---|
| overview | HIGH - Workflow phases, when to use, expected outcomes |
| smart-worktree-detection | CRITICAL - Fuzzy match task to current worktree before creating new |
| phase-0-validation | CRITICAL - MANDATORY - Validate requirements before any work |
| phase-1-planning | HIGH - Deep codebase analysis and implementation planning |
| phase-2-worktree | CRITICAL - MANDATORY - Create isolated worktree with gw add |
| phase-3-implementation | HIGH - Incremental implementation with verification after each change |
| phase-4-testing | CRITICAL - Fast iteration loop until tests pass (Ralph Wiggum pattern) |
| phase-5-documentation | MEDIUM - Update README, CHANGELOG, API docs |
| phase-6-pr-creation | HIGH - Create draft PR, deliver results |
| phase-7-cleanup | LOW - Optional worktree removal after merge |
| decision-framework | HIGH - Branch naming, test strategy, iteration decisions |
| error-recovery | HIGH - Recovery procedures for common errors |
| safety-guardrails | CRITICAL - Validation checkpoints, resource limits, rollback |
| parallel-coordination | HIGH - Multi-agent coordination, handoff protocol |
| artifacts-overview | HIGH - Two-artifact pattern (Plan, Walkthrough), file locations |
| walkthrough-generation | MEDIUM - Final summary generation at Phase 6 |
Templates
| Template | Purpose |
|---|---|
| plan.template.md | Implementation plan with progress log |
| walkthrough.template.md | Final summary for PR delivery |
| agent.template.md | Agent file (copy to ~/.claude/agents/) |
| routing-rule.template.md | Auto-trigger rule (copy to .claude/rules/) |
Auto-Trigger Setup (Recommended)
Install the agent and routing rule so Claude auto-triggers on phrases like *"independently"*, *"in isolation"*.
Option A: Global (personal use — works in all projects)
mkdir -p ~/.claude/agents && \
ln -sf ~/.claude/skills/autonomous-workflow/templates/agent.template.md \
~/.claude/agents/autonomous-workflow.mdThen add the routing rule per-project:
mkdir -p .claude/rules && \
ln -sf ~/.claude/skills/autonomous-workflow/templates/routing-rule.template.md \
.claude/rules/autonomous-workflow-routing.mdOption B: Project-level (team use — committable to git, customizable)
mkdir -p .claude/agents .claude/rules && \
ln -sf ~/.claude/skills/autonomous-workflow/templates/agent.template.md \
.claude/agents/autonomous-workflow.md && \
ln -sf ~/.claude/skills/autonomous-workflow/templates/routing-rule.template.md \
.claude/rules/autonomous-workflow-routing.mdTo customize the agent for a specific project, copy instead of symlink and edit directly. See routing-rule.template.md and agent.template.md for details.
Quick Reference
Full Mode (4+ files, complex changes)
| Phase | Command/Action |
|---|---|
| 0. Validation | Ask clarifying questions, get user confirmation, detect mode |
| 1. Planning | Analyze codebase, prepare plan content in conversation (verbose, all detail) |
| 2. Worktree | gw add feat/feature-name, then CREATE & POPULATE .gw/{branch}/plan.md |
| 3. Implementation | Code in worktree, verify after editing, update Progress Log at milestones |
| 4. Testing | npm test, iterate until passing, log results in Progress Log |
| 5. Documentation | Update README, CHANGELOG |
| 6. PR Creation | CREATE walkthrough.md, gh pr create --draft, SHOW walkthrough to user |
| 7. Cleanup | gw remove feat/feature-name (after merge) |
Lite Mode (1-3 files, simple changes)
| Phase | Command/Action |
|---|---|
| 0. Validation | Quick clarification if needed |
| 1. Planning | Brief mental plan (no artifact files) |
| 2. Worktree | gw add fix/bug-name |
| 3. Implementation | Code directly, commit when done |
| 4. Testing | npm test, fix any failures |
| 5. PR Creation | gh pr create --draft |
Key Principles
- Detect workflow mode FIRST: Determine Full vs Lite before any other action.
- plan.md is the single source of truth: Must be comprehensive enough for a new session to execute alone.
- Always validate first (Phase 0): Never skip directly to implementation.
- Always create worktree (Phase 2): Isolation is mandatory.
- Verify after editing: Run fast checks after each change, full suite before PR.
- Iterate until correct: No artificial iteration limits (Ralph Wiggum pattern).
- CREATE walkthrough.md AND SHOW IT at Phase 6: MANDATORY for Full Mode.
- Stop and ask when blocked: Don't guess on ambiguity.
Artifact System
The workflow produces two artifacts in .gw/{branch-name}/:
| Artifact | File | Created | Purpose |
|---|---|---|---|
| Plan | plan.md | Phase 2 (end) | Implementation strategy, decisions, progress log |
| Walkthrough | walkthrough.md | Phase 6 | Final summary for PR delivery |
Files are gitignored and grouped by branch for easy browsing.
All timestamps in artifact frontmatter MUST use full ISO 8601 with time: YYYY-MM-DDTHH:MM:SSZ (e.g. 2026-03-07T14:30:00Z).
Workflow Flow
Phase 0: Validation <- MANDATORY
| (user confirms)
Phase 1: Planning (prepare content IN CONVERSATION, no files yet)
| (plan validated)
Phase 2: Worktree Setup <- MANDATORY
| Full Mode: CREATE & POPULATE plan.md INSIDE worktree
| (worktree created, plan.md populated)
Phase 3: Implementation
| Verify after editing. Update Progress Log at milestones.
| (code complete)
Phase 4: Testing <- iterate until passing
| (all tests pass)
Phase 5: Documentation
| (docs complete)
Phase 6: PR Creation
| Full Mode: CREATE walkthrough.md + SHOW to user
| (draft PR delivered)
Phase 7: Cleanup (optional)Smart Worktree Detection
Before creating a new worktree, the workflow checks if the current context matches the task:
| Scenario | Action |
|---|---|
| On main/master | Always create new worktree |
| Worktree name matches task keywords | Prompt user to continue or create new |
| No keyword match | Create new worktree |
Fast Iteration Loop (Phase 4)
Based on the Ralph Wiggum pattern:
while not all_tests_pass:
1. Run tests
2. If pass: done
3. If fail: analyze -> fix -> commit -> continue
4. Safety: warn at 10 iterations, stop at 20Troubleshooting Quick Reference
| Issue | Check | Recovery |
|---|---|---|
| Wrong worktree | gw list, pwd | gw cd <correct-branch> |
| gw command not found | which gw | npm install -g @gw-tools/gw |
| Secrets missing | cat.gw/config.json | gw sync <branch>.env |
| Tests keep failing | plan.md Progress Log | Focus on ONE failure, escalate at 7+ |
| Agent hallucinated cmd | Error message | See error-recovery |
| plan.md empty | cat.gw/{branch}/plan.md | STOP, populate plan.md before proceeding |
| walkthrough.md missing | ls.gw/{branch}/ | Create before announcing completion |
Related Skills
- git-worktree-workflows - Worktree fundamentals
- gw-config-management - Configure auto-copy and hooks
References
Detailed examples and scenarios (loaded on-demand):
Research Sources
- Google Antigravity Artifacts - Artifact pattern
- Ralph Wiggum AI Coding Loops - Iteration pattern
- Addy Osmani's LLM Workflow - Fast feedback loops
- Claude Code Worktree Support - Best practices