gods-plan

God's Plan

"I'm using God's Plan to create a multi-model ultra-plan for: $ARGUMENTS"

Phases are numbered 0 through 7 (8 total). God's Plan works WITH plan mode — the plan file is managed by plan mode, God's Plan fills it incrementally. Use ultrathink reasoning throughout for maximum depth.

Iron Laws

• No plan without confession — minimum 8 questions. The plan must not exist before confession is complete.
• No skipping external review — always send to Gemini + Codex, even if the plan looks perfect.
• Evidence before adding — do not add a suggestion from Gemini/Codex without subagent verification.
• User decides — controversial suggestions always go to user via AskUserQuestion.

Phase Restart

If a later phase reveals fundamental problems, restart the appropriate phase with user confirmation:

• Approach is wrong → restart Phase 0 (Brainstorming)
• Missing requirements → restart Phase 2 (Confession)
• Plan structure is wrong → restart Phase 3 (Draft)

Red Flags

Thought pattern	What to do instead
"This task is simple, no confession needed"	User invoked God's Plan — they want the full process. Do the confession.
"I already know what to do, skip questions"	Confession reveals what you don't know. Always ask.
"Gemini/Codex feedback isn't useful"	Present to user — they decide what's useful.
"Replan isn't needed, plan is good"	Replan is mandatory. It consistently finds improvements.
"I'll use Glob/Grep for analysis"	Use Serena semantic tools first — deeper analysis, fewer tokens.

Integration

• Skill("gods-plan:brainstorming") — Phase 0, always first
• Skill("gods-plan:writing-plans") — Phase 3 (does not present to user)
• Skill("gods-plan:replan") — Phase 4, external plan review
• Skill("gemini:rescue") — Phase 5, Gemini review
• Skill("codex:rescue") — Phase 5, Codex review
• Skill("gods-plan:executing-plans") — handoff after user approves plan in Phase 7

---

Phase 0: BRAINSTORMING

Invoke Skill("gods-plan:brainstorming") first. This explores user intent, proposes 2-3 approaches, and gets design approval before any code analysis begins.

Once brainstorming is complete and user approved the design direction, continue to Phase 1.

---

Phase 1: ANALYSIS (via Serena + subagents)

Launch subagents IN PARALLEL (Agent tool, subagent_type: "Explore") that use Serena semantic tools for token-efficient deep analysis:

Agent 1 — Architecture overview:

• get_symbols_overview on key files relevant to $ARGUMENTS
• Understand classes, interfaces, handlers without reading full files
• Read CLAUDE.md (if exists) for project conventions

Agent 2 — Impact analysis:

• find_symbol to locate specific handlers/components related to the task
• find_referencing_symbols to map what depends on what
• Identify what would be affected by changes

Agent 3 — Pattern discovery:

• search_for_pattern to find existing patterns that can be reused
• Identify similar features already implemented in the codebase

Each subagent returns a summary — main context stays clean. Fallback to Glob/Grep/Read only if Serena is unavailable.

Result: internal notes about code state, architecture, dependencies, and reusable patterns. User does nothing.

---

Phase 2: CONFESSION

Do not proceed until confession is complete.

Read ${CLAUDE_SKILL_DIR}/references/confession-questions.md and systematically ask via AskUserQuestion.

Rules:

• Ask 2-4 questions at a time via AskUserQuestion — not one by one, that's too slow
• Ask SPECIFIC questions, not vague ones
• Suggest options (A/B/C) where appropriate
• After each batch of answers, follow up with the next batch
• Minimum 8 questions total (Iron Law)
• Confession ends when you say "I have enough information for the plan"

---

Phase 3: DRAFT PLAN (internal — user does NOT see this)

1. Invoke Skill("gods-plan:writing-plans") without announcing to user — the draft is internal. Only announce phases that involve user interaction (0, 2, 7).
2. Create comprehensive draft plan based on:
   • Code analysis (Phase 1)
   • User context (Phase 2)
   • CLAUDE.md rules and project conventions
3. Plan includes: tasks, bite-sized steps, exact file paths, code snippets, testing strategy
4. Use Edit tool to write draft to the active plan file. Plan mode manages the file path. <topic> = key design direction from Phase 0 (e.g., notification-async-queue, auth-multi-provider).
5. Do not present to user. Continue to Phase 4.

---

Phase 4: REPLAN

1. Invoke Skill("gods-plan:replan") on the draft from Phase 3
2. Replan adaptively selects review perspectives based on plan type
3. Dispatches parallel subagents (3-8 based on complexity)
4. Result: findings categorized as Critical / Important / Minor
5. Incorporate Critical and Important findings into the draft
6. Continue to Phase 5 — do not present to user.

If 2+ of 3+ agents fail: warn user and present partial findings. Ask: (a) retry, (b) continue with partial replan, (c) abort.

---

Phase 5: FRIENDS (Gemini + Codex review)

Read ${CLAUDE_SKILL_DIR}/references/review-plan-prompt.md. Replace:

• {PROJECT_CONTEXT} — relevant CLAUDE.md content + tech stack from Phase 1
• {PLAN_CONTENT} — current draft plan after Phase 4

Send to BOTH models IN PARALLEL via their rescue skills:

Skill("gemini:rescue", args: "<full review prompt with plan>")
Skill("codex:rescue", args: "<full review prompt with plan>")

If a rescue skill fails: warn user and continue without that model. Both models failing is unlikely but not fatal — Phase 4 (Replan) already validated the plan.

If BOTH rescue skills fail: log the failure, inform the user, and proceed directly to Phase 7 (Presentation). Phase 6 (Verification) is skipped since there are no external claims to verify. Note in the Pipeline Report that external review was unavailable.

---

Phase 6: VERIFICATION (via Serena + subagents)

Claude does not blindly accept feedback from friends. Verify claims before incorporating them.

Step 1 — Categorize claims:

• (a) Factual/verifiable (file exists, pattern is used, architecture constraint) → spawn verification agent
• (b) Opinion/suggestion (caching, refactoring, alternative approach) → mark as CONTROVERSIAL and present to user in Phase 7

Step 2 — Verify factual claims: Group claims by source (Gemini, Codex). Spawn 1 verification agent per 2-3 claims (Agent tool, subagent_type: "Explore") using Serena:

• "Gemini claims file X doesn't exist" → find_file + find_symbol verification
• "Codex suggests using pattern Y" → find_symbol + find_referencing_symbols — does Y exist? is it compatible?
• "Gemini says edge case Z is missing" → get_symbols_overview + search_for_pattern — is Z actually possible?

Each agent reports VERIFIED or UNVERIFIED status per claim.

Step 3 — Categorize after verification:

• VERIFIED + ADD — claim confirmed and improves plan → add, mark "[Gemini]" / "[Codex]"
• VERIFIED + SKIP — claim is true but irrelevant / over-engineered
• UNVERIFIED — claim could not be verified → discard with reason
• CONTROVERSIAL — claim is valid but unclear if user wants it → note for presentation

---

Phase 7: PRESENTATION

Only now does user see the result. Present:

1. Final plan (after all revisions) — update draft in the plan file
2. Replan summary — what changed after replan skill
3. Friends summary:
   • What Gemini/Codex suggested
   • What Claude verified and added (and why)
   • What Claude rejected (and why)
   • Controversial suggestions → AskUserQuestion
4. Plan for approval by user

After approval: handoff to Skill("gods-plan:executing-plans").

---

Output format

Plan in .claude/plans/YYYY-MM-DD-gods-plan-<topic>.md:

# {Feature} Implementation Plan (God's Plan)

> **For Claude:** REQUIRED SUB-SKILL: Use gods-plan:executing-plans to implement this plan task-by-task.

**Goal:** {one sentence}
**Architecture:** {2-3 sentences}
**Tech Stack:** {key technologies/libraries}

**Pipeline Report:**
- Replan findings: {critical} critical, {important} important, {minor} minor
- Gemini suggestions: {count} (verified+accepted: {n}, rejected: {m})
- Codex suggestions: {count} (verified+accepted: {n}, rejected: {m})

## Replan Summary
| Finding | Severity | Incorporated? |
|---------|----------|--------------|

## External Review Summary
| Suggestion | Source | Verified? | Accepted? | Reason |
|------------|--------|-----------|-----------|--------|

## Task 1: ...
- [ ] Step 1: ...