Long-Form Autonomous Planning (1A2A Workflow)

Task: $ARGUMENTS

1A Phase: Ask & Plan

You are in the Ask Phase - gather information and create a detailed plan before autonomous execution.

Spawn parallel sub-tasks using specialized agents:

Research Tasks:
- codebase-locator: Find all files related to the feature area
- codebase-analyzer: Understand existing patterns and architecture
- Explore: Investigate integration points and dependencies

For each research task, provide:

• Specific directories to examine
• Exact patterns or code to find
• Required output: file:line references

Step 1: Understand the Task

• Use Glob/Grep/Read to explore relevant codebase areas
• Identify files that will be modified
• Understand existing patterns

Step 2: Ask Clarifying Questions

Use AskUserQuestion to cover:

• Implementation scope and boundaries
• File modification permissions
• Testing requirements
• Phase priorities
• Any constraints or preferences

Step 3: Create TodoWrite Plan

TodoWrite(todos=[
  {content: "Task 1", status: "pending", activeForm: "Doing Task 1"},
  {content: "Task 2", status: "pending", activeForm: "Doing Task 2"},
  // ... more tasks
])

Step 4: Sanity Check with MiniMax Reviewer

Before finalizing your plan, get external perspective:

# Invoke MiniMax reviewer as "devil's advocate"
powershell -File .claude/skills/minimax-mcp/scripts/review-work.ps1 -Context "Plan summary" -Question "What risks or blind spots should I consider?"

What to ask MiniMax:

• "What's missing from this plan?"
• "What could go wrong that I haven't considered?"
• "Are there better approaches for [specific aspect]?"
• "What documentation should I reference?"

Remember: Use the feedback autonomously to strengthen your plan. MiniMax is advisory, not authoritative.

Step 5: Create Self-Contained Plan File

For long autonomous sessions (30+ minutes), create a persistent .md file with:

REQUIRED SECTIONS:

1. 🚫 Common Pitfalls - Anti-patterns to avoid

   • Table format: "Don't do X → Instead do Y"
   • Task-specific mistakes agents commonly make

2. 📚 Quick Reference Links - Essential docs for this task

   • List: roadmaps, instructions, troubleshooting guides
   • Brief one-line description of each

3. ⚠️ Troubleshooting - When things don't work

   • Table: Symptom → Check → Fix
   • Common failures and their solutions

4. 🔄 Reminders - Keep these in mind throughout

   • "Remember to reference [doc] after every N tasks"
   • "Use [skill] when [situation] occurs"
   • Gentle guidance for autonomous execution

Key phrasing for 2A autonomous work:

• "Remember to reference X" (not "check X")
• "Keep in mind to use Y" (not "verify Y")
• "Use Z for W" (clear autonomous action)
• Avoid: "ask", "check", "verify" (can trigger stops)

Step 6: Present Plan for Approval

Before proceeding, summarize:

• What will be done
• What files will be modified
• What the user should expect
• Any concerns raised by MiniMax review (and how you addressed them)

WAIT for user to say "START 2A" before proceeding to autonomous phase.

Scope Guardrails (Finish Game Requests)

When the user says "finish the game" or "finish the roadmap":

• Default to full local beta scope, not a partial pass.
• Include explicit success criteria (intro stability, quest flow to ending, basic HPV, key visuals, tests).
• Do not narrow scope unless the user explicitly approves the reduction.
• Plan should cover multiple phases and be marked as multi-session if needed.
• If you must break into blocks, note that the overall goal remains unchanged and continue into the next block without waiting for re-approval unless a HARD STOP applies.

2A Phase: Autonomous Execution

User has approved. You are now in Autonomous Phase - work through the todo list independently without stopping to ask questions.

CRITICAL: During 2A, work continuously. Do NOT stop for:

• Clarification questions (make reasonable assumptions)
• "Should I do X?" (use your best judgment)
• "Is this correct?" (proceed with confidence)

HARD STOPS still apply: Creating .md files, git push, editing CONSTITUTION.md, actions outside scope.

Work Guidelines

DO autonomously:

• Complete all todo items systematically
• Handle blockers by continuing with other items
• Use skills before manual implementation
• Run tests when appropriate
• Commit changes when work blocks complete

HARD STOPS (always ask, even in 2A):

• Creating NEW .md files (not edits)
• Editing .cursor/ directory
• Git push, force push, branch operations
• Editing CONSTITUTION.md
• Actions outside approved scope

When Blocked

1. Note the blocker
2. Continue with other todo items
3. Summarize blockers at end
4. Don't get stuck - move forward

Tracking with TodoWrite

All /longplan sessions use TodoWrite for progress tracking.

Why TodoWrite:

• Integrated with CLI /todos command for visibility
• Progress updated in real-time as you work
• No extra files cluttering the repo
• Automatically cleaned up after session ends

Session Persistence (.md file):

• /longplan creates an .md file ONLY when session persistence is needed
• Use when: terminal may close, work spans multiple sessions, user is away
• Format: temp/autonomous-work-[task-name].md or similar
• Shorter planning (via create-plan skill) uses TodoWrite only (no .md file needed)

Tracking Best Practices:

• Update todo status immediately when starting/finishing tasks
• Use activeForm to describe current action
• Mark blockers in notes, don't stop working
• Create .md file only if session may be interrupted

Dual-Reviewer System

Use BOTH reviewers during 2A phase for quality control:

GLM Devil's Advocate (Self-Review)

When to use:

• Before making significant changes
• After completing multi-step tasks
• When uncertain about approach

How to invoke: Read temp/glm-devils-advocate-prompt.md and respond as the critical reviewer.

Output format:

### 🔍 Critical Review
**Concerns:** [list]
**Edge Cases Missed:** [list]
**Alternatives Not Considered:** [list]
**Recommendation:** PROCEED / REVISE / RECONSIDER

MiniMax Reviewer (External Standards)

When to use:

• Checking against documentation standards
• Looking for precedents in trusted docs
• Verifying alignment with best practices

How to invoke:

powershell -File .claude/skills/minimax-mcp/scripts/review-work.ps1 -Context "what I did" -Question "question?"

Output: Retrieves relevant documentation from trusted domains (docs.anthropic.com, godotengine.org, etc.)

Reviewer Strategy

Remember: Reviewers are ADVISORS. Use judgment in applying feedback. Don't blindly follow - they're your "devil on your shoulder."

Completion

When all todos complete:

1. Summarize work done
2. List any blockers encountered
3. Show files modified
4. Ask if any follow-up needed

Example Flow

User: /longplan Fix the quest 4 dialogue bug

Agent: [1A Phase]
- Reads quest 4 dialogue files
- Asks: "Should I also check related quest files?"
- Creates TodoWrite plan
- Presents plan for approval
- WAITS...

User: START 2A

Agent: [2A Phase]
- Works through todos
- Uses GLM reviewer before major changes
- Uses MiniMax reviewer to check standards
- Completes work
- Summarizes results

Signature: [GLM-4.7 - 2026-01-20]