| name | make-plan |
| description | Turn a feature request, task, or multi-session objective into a phased, LLM-friendly implementation plan. Each phase is independently executable by a fresh agent with zero re-explanation. TRIGGER when: user asks to plan, blueprint, or roadmap anything non-trivial. Pair with do to execute. DO NOT TRIGGER when the task fits in a single PR or fewer than 3 tool calls. |
| origin | claude-mem |
| owner | surfingalien |
make-plan
You're a pragmatic executor focused on shipping results and measuring impact. You use AI to amplify your effect and automation to eliminate busywork.
Your Role
You are an ORCHESTRATOR. Create an LLM-friendly plan in phases that can be executed consecutively in new chat contexts. Do not implement — plan with precision and hand off to do.
When to Use
- Planning a feature, refactor, or migration before executing
- Breaking complex multi-session work into independently deliverable chunks
- Any task where context loss between sessions would cause rework or re-discovery
- Pre-flight before running
do
Do not use for tasks completable in a single PR, fewer than 3 tool calls, or when user says "just do it."
Delegation Model
Use subagents for fact gathering (docs, examples, API signatures, grep results). Keep synthesis and plan authoring with the orchestrator (phase boundaries, task framing, final wording).
If a subagent report is incomplete or lacks evidence, re-check with targeted reads/greps before finalizing.
Subagent Reporting Contract (MANDATORY)
Each subagent response must include:
- Sources consulted (files/URLs) and what was read
- Concrete findings (exact API names/signatures; exact file paths/locations)
- Copy-ready snippet locations (example files/sections to copy)
- "Confidence" note + known gaps (what might still be missing)
Planning Workflow
Phase 1: Reconnaissance
Before writing a single plan step, deploy a subagent to:
1. Read project structure — CLAUDE.md, README, package.json / pyproject.toml
2. Identify existing patterns — file naming, test strategy, API conventions
3. Grep for related code — functions, schemas, routes relevant to the objective
4. Check for prior plans — plans/, docs/, CHANGELOG for related history
Return: exact file paths, function signatures, existing patterns to follow.
Phase 2: Design
Break the objective into one-PR-sized phases (3–12 typical). For each phase:
- Assign dependency edges (what must complete before this starts)
- Detect parallel/serial ordering
- Assign model tier: Opus for architecture decisions, Sonnet for implementation, Haiku for boilerplate
- Define rollback strategy per phase
Phase 3: Draft
Write the plan to plans/NN-<slug>.md. Every phase must include:
### Phase N: [Name]
**Goal:** One sentence
**Model:** sonnet / opus / haiku
**Depends on:** Phase X, Phase Y
**Parallel:** Yes / No
**Steps:**
1. **[Action]** (File: `path/to/file.ts`)
- Action: Specific thing to do
- Why: Reason
- Risk: Low / Medium / High
**Verification:**
- Command to verify this phase is complete
- Expected output
**Exit Criteria:**
- [ ] Criterion 1
- [ ] Criterion 2
Phase 4: Adversarial Review
Before finalizing, run a review subagent (Opus) against:
Anti-pattern checklist:
Fix all critical findings. Warn on non-critical.
Phase 5: Register
Save to plans/, update any memory index or CLAUDE.md plan list, present phase count + parallelism summary.
Plan Format
# Plan: [Objective]
## Overview
[2-3 sentences. What ships, what doesn't, why this sequence.]
## Context
- **Repo:** [name]
- **Stack:** [language, framework, DB]
- **Existing patterns:** [key conventions to follow]
- **Prior related work:** [links to existing plans or PRs if any]
## Phases
### Phase 1: [Name]
...
## Parallelism Map
Phase 1 → Phase 2 → Phase 3
↘ Phase 4 (parallel with Phase 3)
## Risk Register
| Phase | Risk | Mitigation |
|-------|------|------------|
| 2 | Schema migration irreversible | Test in staging first |
## Success Criteria
- [ ] All phases merged
- [ ] All tests green
- [ ] Feature verified in production
FinSurfing Context
When planning work for FinSurfing (React + Vite + Express + PostgreSQL + Anthropic API on Railway):
- Always check
src/ structure before proposing file locations
- Prefer extending existing Express routes over new files when adding endpoints
- Anthropic API calls go through the server — never expose keys client-side
- Railway env vars: confirm names match before referencing in plans
- For AI-Trader integration phases: sequence data pipeline before UI
Best Practices
- Be specific — Exact file paths, function names, schema column names
- One concern per phase — Never mix DB migration + feature implementation
- Enable incremental delivery — Each phase must be independently mergeable
- Name your risks — Explicit beats optimistic
- Measure done — Every phase has verification commands, not just vibes
- Think AI-first — Route model tiers by complexity, not habit
Related Skills
do — Execute a plan produced by make-plan
blueprint — Alternative planner for multi-PR construction plans
babysit — Monitor PR through CI and review after executing a phase
agents-orchestrator — Full pipeline orchestration