| name | plan-create |
| description | Create an agent-executable implementation plan for a feature or project. Generates plans/.context.md and individual plan files. |
| argument-hint | <feature description or project goal> |
| allowed-tools | Bash, Read, Write, Edit, Glob, Grep |
Create a set of implementation plans for $ARGUMENTS. Each plan must be self-contained and completable by a single agent in one context window.
Steps
1. Understand the project
- Read
CLAUDE.md, .context.md, and any existing plans/ directory
- Identify the tech stack, conventions, and existing architecture
- If
plans/.context.md already exists, read it and build on existing plans
2. Decompose the work
Break the feature into the smallest plans that:
- Can each be implemented in a single agent session
- Have clear input (what exists) and output (what to create/modify)
- Are ordered by dependency (no circular deps)
Use numbered plans: 01-short-name.md, 02-short-name.md, etc.
For complex features (5+ sub-tasks), create a meta-plan first (e.g., 15-frontend-breakdown.md) that decomposes into sub-plans with letter suffixes (e.g., 15a-setup.md, 15b-api-clients.md).
3. Create plans/.context.md
# Plans
## Purpose
[One sentence: what these plans implement]
## Functionality
- [Capability 1 these plans deliver]
- [Capability 2 these plans deliver]
## Dependency Order
\`\`\`
Phase 1: [Category]
01-plan-name ──────────────┐
│
Phase 2: [Category] │
02-plan-name ──────────────┤ (depends on 01)
03-plan-name ──────────────┘ (standalone)
\`\`\`
## Files
| File | Description |
|------|-------------|
| `01-plan-name.md` | Brief description of what this plan implements |
| `02-plan-name.md` | Brief description of what this plan implements |
## TODO
| Priority | Task |
|----------|------|
| P0 | Execute plan 01 (critical path) |
| P1 | Execute plans 02–03 (next phase) |
Conventions:
- Use ASCII art for the dependency graph, showing phases and arrows
- Mark completed plans with
~~DONE~~ in the Files table description
- Strike through completed TODO rows with
~~text~~
- Use priorities P0 (critical), P1 (important), P2 (nice-to-have)
4. Create individual plan files
Each plan file follows this structure:
# Plan NN: [Feature Name]
## Context
[What exists now and why this plan is needed. 2-3 sentences.]
**Depends on:** [Plan numbers or "Nothing"]
**Blocks:** [Plan numbers or "Nothing"]
## Current State
- [What relevant code/infrastructure already exists]
- [What's missing that this plan addresses]
## Scope
[One paragraph: exactly what this plan adds or changes]
## Files to Create/Modify
### `path/to/file.ext`
[What to do in this file]
\`\`\`language
// Key code snippets — enough for an agent to implement
\`\`\`
### `path/to/another-file.ext`
[Repeat for each file]
## Verification
\`\`\`bash
[Commands the agent should run to verify the implementation]
\`\`\`
## Acceptance Criteria
- [Specific, testable condition 1]
- [Specific, testable condition 2]
- [All verification commands pass]
5. Plan file conventions
Code snippets: Provide enough code for the agent to implement without guessing. Include:
- Function signatures with types
- Key logic (algorithms, error handling)
- Interface definitions
- SQL queries, proto definitions, config structures
Don't include:
- Boilerplate imports the agent can infer
- Every line of implementation — describe non-obvious parts, let the agent fill in the rest
Verification section should always include project-specific commands:
- Build/compile commands
- Lint commands
- Test commands (unit, integration, e2e as applicable)
Acceptance criteria should be specific enough that pass/fail is unambiguous.
6. Meta-plan structure (for complex features)
When a feature has 5+ sub-tasks, create a meta-plan:
# Plan NN: [Feature] Breakdown
## Context
[Feature overview and why it needs decomposition]
**Depends on:** [Prerequisites]
**Blocks:** Nothing
## Sub-Plan Dependency Graph
\`\`\`
NNa-sub-plan ────────────────────┐
│
NNb-sub-plan ────────────────────┤ (standalone)
│
NNc-sub-plan ──── (depends on NNb)
│
NNd-sub-plan ──── (depends on NNa, NNc)
\`\`\`
## Sub-Plans
| Plan | Title | Depends On |
|------|-------|------------|
| NNa | Sub-feature A | Nothing |
| NNb | Sub-feature B | Nothing |
| NNc | Sub-feature C | NNb |
| NNd | Sub-feature D | NNa, NNc |
## What's NOT in Scope
- [Explicit exclusion 1 — why]
- [Explicit exclusion 2 — why]
7. Review and validate
Before finishing:
- Verify every plan has a dependency path that starts from either "Nothing" or an already-completed plan
- Verify no circular dependencies exist
- Verify the
.context.md dependency graph matches individual plan headers
- Verify each plan's scope is achievable in one agent session (rule of thumb: 1-5 files to create/modify)
Notes
- Plans are designed for
/plan-loop (automated ralph-loop) and /plan-status (progress checking)
- Completion is tracked by adding
✅ COMPLETE to the plan title and ~~DONE~~ in .context.md
- Prefer many small plans over fewer large ones — agents work best with focused scope
- Keep plans in a flat
plans/ directory, using number prefixes for ordering