Write Plan Consult

Create an implementation plan by consulting external LLMs throughout the process.

User request: $ARGUMENTS

Phase 1: Understand the Task

Start by understanding what exists and what the user wants.

1. If relevant, explore the codebase to understand current state
2. Use AskUserQuestion to ask clarifying questions one at a time
3. Keep asking until you have enough clarity to write a plan

Rules for questions:

• ONE question per message (never batch multiple questions)
• Use AskUserQuestion with 2-4 options whenever possible
• Keep option labels concise (1-5 words), use descriptions for details
• If you realize you misunderstood something, acknowledge it and course-correct

Phase 2: Brainstorm with External LLMs

Once you understand the task, consult Gemini and Codex in parallel for approaches and ideas.

Spawn TWO parallel subagents (Agent tool, subagent_type: "general-purpose", model: "sonnet"). Each subagent makes the MCP call and returns the full response.

Gemini subagent:

Call mcp__consult-llm__consult_llm with:

• model: "gemini"
• prompt: The brainstorm prompt below
• files: Array of relevant source files for context

Codex subagent:

Call mcp__consult-llm__consult_llm with:

• model: "openai"
• prompt: The brainstorm prompt below
• files: Array of relevant source files for context

Brainstorm prompt:

I'm planning the following task:

[Task description with full context]

Relevant files and their roles:
[List the key files and what they do]

Propose 2-3 approaches for implementing this. For each approach:
- Describe the strategy and trade-offs
- List the files to create/modify with exact paths
- Include concrete code examples showing the key parts (not pseudocode)
- Note any edge cases or gotchas

Be specific and opinionated. Recommend your preferred approach and explain why.

Phase 3: Synthesize and Write the Plan

Review both LLM responses. Pick the best ideas from each and combine them with your own analysis into a single implementation plan.

Plan Structure

# [Feature Name] Implementation Plan

**Goal:** [One sentence describing what this builds]

**Approach:** [2-3 sentences about the chosen approach and why]

**Sources:** [Brief note on which ideas came from Gemini vs Codex vs your own analysis]

---

### Task 1: [Short description]

**Files:**

- Create: `exact/path/to/file.py`
- Modify: `exact/path/to/existing.py` (lines 123-145)

**Steps:**

1. [Specific action]
2. [Specific action]

**Code:**

```language
// Include actual code, not placeholders like "add validation"
```

---

### Task 2: [Short description]

...

Guidelines

• Exact file paths - never "somewhere in src/"
• Complete code - show the actual code, not "implement the logic"
• Small tasks - each task should be 2-5 minutes of work
• Assume no context - write as if the implementer knows nothing about this codebase
• DRY, YAGNI - only what's needed, no speculative features
• Credit sources - note when an idea came from a specific LLM's suggestion

Phase 4: Save

Save the plan:

• Write to a markdown file at history/<date>-plan-<feature-name>.md (e.g. history/2026-02-15-plan-user-auth.md)
• Include context, decisions made, and rationale

Phase 5: Review

Get Gemini and Codex to review the synthesized plan, again in parallel.

Spawn TWO parallel subagents (Agent tool, subagent_type: "general-purpose", model: "sonnet"). Each subagent makes the MCP call and returns the full response.

Gemini subagent:

Call mcp__consult-llm__consult_llm with:

• model: "gemini"
• prompt: The review prompt below
• files: Array including the plan file and relevant source files

Codex subagent:

Call mcp__consult-llm__consult_llm with:

• model: "openai"
• prompt: The review prompt below
• files: Array including the plan file and relevant source files

Review prompt:

Review this implementation plan. Consider:
- Are the tasks correctly ordered and sized?
- Are there any missing steps or edge cases?
- Are the file paths and code snippets accurate?
- Any architectural concerns or better approaches?

Provide specific, actionable feedback. Be concise.

After receiving feedback, present it to the user and ask if they want to revise the plan.

Principles

• One question at a time - never batch multiple questions
• Use AskUserQuestion - clickable options are faster for the user
• YAGNI - ruthlessly cut unnecessary features
• Validate incrementally - check understanding at each step
• Concrete over abstract - exact paths, actual code, specific commands
• Best of all worlds - synthesize the strongest ideas from each LLM