From gods-plan
8-phase ultra-planning with Claude, Gemini, Codex and Serena. Brainstorm, analyze, confess, draft, replan, friend-review, verify, present.
How this skill is triggered — by the user, by Claude, or both
Slash command
/gods-plan:gods-planThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
"I'm using God's Plan to create a multi-model ultra-plan for: $ARGUMENTS"
"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.
If a later phase reveals fundamental problems, restart the appropriate phase with user confirmation:
| 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. |
Skill("gods-plan:brainstorming") — Phase 0, always firstSkill("gods-plan:writing-plans") — Phase 3 (does not present to user)Skill("gods-plan:replan") — Phase 4, external plan reviewSkill("gemini:rescue") — Phase 5, Gemini reviewSkill("codex:rescue") — Phase 5, Codex reviewSkill("gods-plan:executing-plans") — handoff after user approves plan in Phase 7Invoke 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.
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 $ARGUMENTSAgent 2 — Impact analysis:
find_symbol to locate specific handlers/components related to the taskfind_referencing_symbols to map what depends on whatAgent 3 — Pattern discovery:
search_for_pattern to find existing patterns that can be reusedEach 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.
Do not proceed until confession is complete.
Read ${CLAUDE_SKILL_DIR}/references/confession-questions.md and systematically ask via AskUserQuestion.
Rules:
Skill("gods-plan:writing-plans") without announcing to user — the draft is internal. Only announce phases that involve user interaction (0, 2, 7).<topic> = key design direction from Phase 0 (e.g., notification-async-queue, auth-multi-provider).Skill("gods-plan:replan") on the draft from Phase 3If 2+ of 3+ agents fail: warn user and present partial findings. Ask: (a) retry, (b) continue with partial replan, (c) abort.
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 4Send 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.
Claude does not blindly accept feedback from friends. Verify claims before incorporating them.
Step 1 — Categorize claims:
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:
find_file + find_symbol verificationfind_symbol + find_referencing_symbols — does Y exist? is it compatible?get_symbols_overview + search_for_pattern — is Z actually possible?Each agent reports VERIFIED or UNVERIFIED status per claim.
Step 3 — Categorize after verification:
Only now does user see the result. Present:
After approval: handoff to Skill("gods-plan:executing-plans").
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: ...
Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub marionitzke/claude-gods-plan --plugin gods-plan