Autonomous Long-Task Development

This skill optimizes Claude Code for long, autonomous development tasks—from multi-hour feature implementations to multi-session refactorings.

Core Principles

1. Loop Closing (Test-Driven)

Every code change must close a feedback loop:

1. Write test (define expected behavior)
2. Implement code
3. Run test → loop closed
4. Refactor if needed
5. Re-run test → confidence

Anti-pattern (Open Loop):

"Implement feature X"
→ Code written, no verification
→ Bugs discovered later

Best Practice (Closed Loop):

"Implement feature X with tests.
Run tests after implementation."
→ Immediate verification

2. Incremental Progress with Checkpoints

Never attempt to "one-shot" complex features:

❌ Implement entire feature at once
   → Context runs out mid-implementation
   → Next session inherits chaos

✅ Small, tested increments
   → Each increment works standalone
   → Clean handoff between sessions

Checkpoint patterns:

• Use /rewind or Esc Esc for rollbacks
• Commit after each working increment
• Document state in claude-progress.txt for session handoffs

3. Context Management

Claude has 200K tokens, but:

• Subagents have isolated context windows
• Long sessions fragment context
• Use claude-progress.txt for session handoffs
• /compact <focus> for manual compaction

---

Task System

Use TaskCreate/TaskUpdate for persistent progress tracking (replaces deprecated TodoWrite).

Quick Start

1. TaskCreate: Create task with subject, description, activeForm
2. TaskUpdate: Set dependencies with addBlockedBy
3. TaskUpdate: Mark in_progress when starting
4. TaskUpdate: Mark completed when done
5. TaskList: Find next available work

Task Workflow

pending → in_progress → completed
         ↑
    (start work)

Tasks persist across sessions and auto-unblock when dependencies complete.

For detailed patterns: See references/task-patterns.md

---

Subagents & Parallelization

Subagents are lightweight Claude instances with isolated context. Only relevant results return to the orchestrator.

Available Agent Types

Agent	Purpose	Tools
general-purpose	Complex multi-step tasks	All
Explore	Codebase exploration, pattern search	Glob, Grep, Read (no Edit/Write)
Plan	Implementation planning	Glob, Grep, Read (no Edit/Write)
Bash	Git operations, command execution	Bash only

When to Use Subagents

DO:

• Explore unfamiliar codebase areas
• Search patterns across many files
• Parallelize independent tasks (max 10 parallel)
• Isolate high-volume operations (tests, logs)

DON'T:

• Read specific known file → Use Read directly
• Search in 2-3 files → Use Read directly
• Find class definition → Use Glob directly

Parallel Patterns

Background subagents:

"Implement Stripe integration parallel:

Subagent 1 (Backend): Create API endpoint
Subagent 2 (Frontend): Payment form component
Subagent 3 (Tests): Integration tests

Start all three with Task tool (run_in_background: true).
Wait for completion, then integrate."

Git worktrees for true parallelism:

git worktree add ../feature-a feature/a
git worktree add ../feature-b feature/b
# Separate Claude sessions in each worktree

Custom Subagents

Create custom agents in .claude/agents/:

• long-task-coordinator.md - orchestrates multi-step work
• test-runner.md - runs and fixes tests
• code-reviewer.md - reviews before commit

---

Session Management

Naming Sessions

Use /rename to give sessions descriptive names:

/rename stripe-integration

Resuming Work

claude --continue    # Resume most recent session
claude --resume      # Session picker
claude --from-pr 123 # Resume PR-linked session

Session Picker Shortcuts

Key	Action
↑/↓	Navigate sessions
P	Preview session
R	Rename session
B	Filter by branch
/	Search

Context Management

• Auto-compaction triggers at ~95% capacity
• Configure earlier: CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50
• Manual: /compact Focus on API changes
• Clear between unrelated tasks: /clear

For detailed patterns: See references/session-management.md

---

Template Prompts

Long-Running Feature Implementation

"Implement [FEATURE] with complete loop-closing.

Requirements:
- [Req 1]
- [Req 2]

Approach:
1. Create tasks with TaskCreate for each major step
2. Set dependencies with TaskUpdate (addBlockedBy)
3. Work through tasks: test → implement → verify

Constraints:
- Each increment must be testable
- Commit after each working step
- Update claude-progress.txt if context fills

Token-Budget: Use full 200K, do NOT stop early.
At context limit: Update claude-progress.txt for handoff.
Parallelize independent tasks with subagents."

Multi-Session Refactoring

"Refactoring Part [N] of [TOTAL].

Previous Sessions:
- See claude-progress.txt
- Git log for prior changes

This Session Goals:
- [Goal 1]
- [Goal 2]

At session end:
- Update claude-progress.txt
- All tests must pass
- Document clear next steps

If context runs low:
- Finish cleanly (no half implementations)
- Write handoff document
- Next session continues seamlessly"

Bug Investigation

"Investigate and fix bug: [DESCRIPTION]

Observed behavior:
- [What happens]
- [When it happens]
- [Error messages]

Investigation Steps:
1. Reproduce locally
2. Check logs
3. Identify root cause
4. Implement fix
5. Add regression test (loop closing!)
6. Verify

Document findings in claude-progress.txt if complex."

---

Autonomy Levels

High Autonomy (Default for Long Tasks)

"Implement feature X.
Make all implementation decisions based on best practices.
Do NOT stop due to token budget.
Create tests and docs as you work."

Guided Autonomy (For Critical Decisions)

"Implement feature X.
Ask me BEFORE decisions about:
- Database schema changes
- External API selection
- Breaking changes

For everything else: Proceed autonomously."

---

Common Failure Modes

Problem	Solution
Claude stops too early	Say "Do NOT stop due to token budget" explicitly
Too many clarifying questions	More context upfront, set autonomy level
Code style mismatch	"Follow pattern in [FILE]"
Task too complex	Break into increments, use TaskCreate
Context runs out	claude-progress.txt + clean commits
One-shotting fails	Explicitly require incremental approach
Lost progress between sessions	Use Tasks system (persists), not mental tracking

---

Verification Patterns

Test Pyramid

          /\
         /E2E\ ← Playwright (critical flows)
        /-----\
       / API   \ ← Integration tests
      /---------\
     /   Unit    \ ← Fastest feedback
    /--------------\

Loop-Closing Checklist

1. Test exists for the change
2. Test ran after implementation
3. Test passed (or failure was addressed)
4. Commit includes both code and test

For detailed patterns: See references/verification-patterns.md

---

Token & Cost Considerations

• Single agent: ~50K-100K tokens per session
• Parallel agents: 3-4x higher token consumption
• Trade-off: Higher velocity vs. higher cost

Recommendations:

• Quick tasks: Single agent
• Complex/long tasks: Subagents justify cost
• Claude Max Plan for heavy-duty usage

---

Quick Reference

START:
1. TaskCreate with task breakdown
2. Define autonomy level
3. Establish loop-closing pattern

DURING:
- Test each increment
- Commit after working steps
- Subagents for parallel work
- TaskUpdate status as you progress

END / HANDOFF:
- Update claude-progress.txt
- All tests green
- Clear next steps documented
- TaskUpdate: mark completed

---

Sources

Based on:

• Claude Code Best Practices
• Common Workflows
• Create Custom Subagents
• Enabling Claude Code to Work Autonomously