Plan Directory

Plan Directory Skill

When to Use This Skill

• Scaffolding a new project, feature, or migration that benefits from a structured, step-by-step plan.
• Creating a repeatable plan that an LLM or engineer can execute with explicit, verifiable checkboxes.
• Maintaining an existing plan directory: adding tasks, updating progress, or archiving completed plans.
• When the user says "create a plan", "scaffold a plan", "plan this feature", or similar intent implying structured task breakdown.

Do not use this skill for ad-hoc todo lists, single-file notes, or when the user explicitly wants a different format.

LLM Intake (Required Inputs)

Before writing any files, gather these inputs. If any are missing, ask the user for them explicitly.

When updating an existing plan, read the current PLAN.md first to understand context before modifying.

Invariants (Do Not Violate)

1. PLAN.md is the index, not the content. It contains only the purpose, usage instructions, locked decisions, and a task index with checkboxes. Detailed steps live in individual task files.

2. Task files use 3-digit numbering. Format: NNN-<slug>.md where NNN is zero-padded (001, 002, ..., 999) and <slug> is hyphenated lowercase.

3. Every task file has six sections. Goal, Scope, Checklist, Tests, Completion Criteria, and Dependencies. All are required (use "None" for Dependencies if truly independent, "N/A" for Tests only with justification).

4. Checkboxes are the only status markers. Do not add "Status: Done" fields, emoji indicators, or separate progress sections. Check - [x] when complete, leave - [ ] when pending.

5. Mirror completion in PLAN.md. When all checklist items in a task file are checked, check the corresponding task in PLAN.md.

6. Do not renumber existing tasks. Once a task number is assigned and work has started, it is permanent. Append new tasks at the end.

7. Keep task files self-contained. A reader should understand the task from the file alone without reading other task files.

8. Track blockers explicitly. When a task is blocked, add a ## Blockers section describing what's blocking and link to the blocking task or issue.

Workflow

1. Determine Mode: Create or Update

• Create: No plan directory exists. Scaffold everything from scratch.
• Update: PLAN.md already exists. Read it, understand the current state, and make targeted modifications.

2. Validate Inputs

• Confirm all required inputs are present.
• If the user provides a vague task list, ask for clarification before proceeding.
• For updates, verify the plan slug and location match the existing plan.

3. Create or Update the Plan Directory

For new plans:

1. Create the directory at the target location.
2. Write PLAN.md with the master index.
3. Write each task file (001-*.md, 002-*.md, etc.).

For updates:

1. Read the existing PLAN.md and relevant task files.
2. Apply changes: add new tasks, update checklists, check completed items.
3. Ensure PLAN.md index stays in sync with task files.

4. Fill PLAN.md (Master Index)

Keep it minimal. Include only:

• Purpose: 1-3 bullets on why the plan exists.
• How to use: Brief instructions for working the plan.
• Decisions (locked): Key constraints that should not change.
• Task Index: Checkbox list linking to task files.
• Completion: Definition of when the entire plan is done.

5. Fill Task Files

Each task file must include:

• Goal: Single sentence describing the outcome.
• Dependencies: List of task numbers that must complete first, or "None".
• Scope: Bulleted list of what's in scope (and optionally out of scope).
• Checklist: 3-8 concrete, actionable steps with checkboxes.
• Tests: Specific test commands, files, or manual QA steps.
• Completion Criteria: Measurable definitions of done.
• Notes (optional): Constraints, references, warnings.
• Blockers (optional): Added when work is blocked; removed when unblocked.

6. Maintain Progress

As work completes:

1. Check items in the task file's Checklist and Tests sections.
2. When all items are checked, check the Completion Criteria items.
3. Check the corresponding task in PLAN.md's Task Index.

Task Sizing Rules

• Target 3-8 checklist items per task. This keeps tasks focused and completable in a reasonable session.
• Split if exceeding 10 items. If a task grows beyond ~10 checklist items, break it into subtasks.
• Tests must be explicit. Include runnable commands or specific file paths. Avoid vague "test that it works" items.
• Completion criteria must be measurable. Use "X is true" or "Y passes" rather than narrative descriptions.

Update Rules (Existing Plans)

Templates

PLAN.md

# <Plan Title> - Master Plan

## Purpose

- <Why this plan exists>
- <What it delivers when complete>

## How to Use

1. Work tasks in order unless dependencies indicate otherwise.
2. Check items in task files as they are completed.
3. This file is the index only; details live in task files.

## Decisions (Locked)

- <Key constraint or choice that should not change>
- <Another locked decision>

## Task Index

- [ ] 001 - <Task Name> (`001-<slug>.md`)
- [ ] 002 - <Task Name> (`002-<slug>.md`)
- [ ] 003 - <Task Name> (`003-<slug>.md`)

## Completion

- [ ] All tasks in the index are checked.
- [ ] All tests listed in task files pass.
- [ ] <Any additional project-specific completion criteria>

Task File (NNN-.md)

# NNN - <Task Name>

## Goal

<Single sentence describing the outcome of this task.>

## Dependencies

- Requires: 001, 002 (or "None" if independent)
- Blocks: 004, 005 (tasks waiting on this one)

## Scope

**In scope:**
- <What this task covers>
- <Another in-scope item>

**Out of scope:**
- <What this task explicitly does not cover>

## Checklist

- [ ] <Concrete implementation step with verifiable outcome>
- [ ] <Another concrete step>
- [ ] <Another concrete step>

## Tests

- [ ] Run `<test command>` and verify all pass
- [ ] Manual QA: <specific verification step>

## Completion Criteria

- [ ] <Measurable definition of done>
- [ ] <Another measurable definition>

## Notes

- <Optional: constraints, references, links, or warnings>

Blocked Task (when work cannot proceed)

When a task is blocked, add a ## Blockers section immediately after Dependencies:

## Blockers

- **Blocked by:** External API not yet available (ETA: 2024-02-15)
- **Blocked by:** Waiting on 003 to complete first
- **Action needed:** Contact platform team for API access

Remove the Blockers section when the task is unblocked.

Quality Checklist (LLM Preflight)

Before delivering a plan, verify:

• All required inputs were gathered or reasonable defaults applied.
• PLAN.md is short and contains only the index, not detailed steps.
• Task filenames match the pattern NNN-<slug>.md with 3-digit padding.
• Every task file has all six required sections (Goal, Dependencies, Scope, Checklist, Tests, Completion Criteria).
• Dependencies are explicitly stated (even if "None").
• Checklist items are concrete and verifiable, not vague.
• Tests section contains runnable commands or specific QA steps.
• Completion criteria are measurable, not narrative.
• PLAN.md Task Index matches the task files exactly.
• No task exceeds ~10 checklist items (split if needed).
• Dependency graph has no cycles (task A can't depend on B if B depends on A).
• Parallel-executable tasks are identified where applicable.

Parallel Task Execution

Some tasks can run concurrently if they have no dependencies on each other. Indicate this in PLAN.md using a parallel block notation:

## Task Index

- [ ] 001 - Database setup (`001-db-setup.md`)
- [ ] 002 - Frontend scaffolding (`002-frontend-scaffold.md`) [parallel: 001]
- [ ] 003 - API design (`003-api-design.md`) [parallel: 001, 002]
- [ ] 004 - Integration (`004-integration.md`) [after: 001, 002, 003]

The [parallel: NNN] tag means "can run at the same time as task NNN". The [after: NNN] tag means "must wait for NNN to complete".

When working a plan:

• Start all tasks marked parallel that have no unmet dependencies.
• Track progress on parallel tasks independently.
• Only start [after: ...] tasks when all listed dependencies are checked.

Examples

Example 1 - Creating a New Plan

User prompt:

"Create a plan for adding user authentication to my app. Use JWT tokens, store users in PostgreSQL, and include password reset flow."

Expected behavior:

1. Ask for any missing inputs (e.g., preferred directory, additional tasks).
2. Create docs/plans/user-auth/ (or user-specified location).
3. Write PLAN.md with the task index.
4. Write task files like:
   • 001-user-model.md - Create User model and migrations
   • 002-jwt-setup.md - Configure JWT authentication
   • 003-login-endpoint.md - Implement login/logout endpoints
   • 004-password-reset.md - Implement password reset flow
   • 005-integration-tests.md - Write integration tests

Output structure:

docs/plans/user-auth/
  PLAN.md
  001-user-model.md
  002-jwt-setup.md
  003-login-endpoint.md
  004-password-reset.md
  005-integration-tests.md

Example 2 - Updating an Existing Plan

User prompt:

"Add a new task to the user-auth plan for implementing OAuth with Google."

Expected behavior:

1. Read the existing PLAN.md to understand current state.
2. Determine the next task number (e.g., 006).
3. Create 006-google-oauth.md with the standard sections.
4. Append the new task to PLAN.md's Task Index.
5. Do not modify existing task files unless explicitly requested.

Example 3 - Marking Progress

User prompt:

"I've completed the JWT setup task in the user-auth plan."

Expected behavior:

1. Read 002-jwt-setup.md.
2. Check all checklist items, tests, and completion criteria as done.
3. Update PLAN.md to check the 002 - JWT Setup entry.
4. Confirm the update to the user.

Example 4 - Complete Task File (Reference)

Here's a fully filled-out task file for reference:

# 003 - Login Endpoint

## Goal

Implement JWT-based login and logout endpoints that authenticate users and
return tokens.

## Dependencies

- Requires: 001, 002
- Blocks: 004, 005

## Scope

**In scope:**
- POST /api/auth/login endpoint accepting email/password
- POST /api/auth/logout endpoint invalidating tokens
- Token refresh endpoint
- Rate limiting on login attempts

**Out of scope:**
- OAuth/social login (separate task 006)
- Password reset flow (task 004)
- User registration (task 001 handles this)

## Checklist

- [ ] Create LoginSerializer with email and password fields
- [ ] Implement login view returning access and refresh tokens
- [ ] Implement logout view that blacklists the refresh token
- [ ] Add rate limiting: max 5 failed attempts per 15 minutes
- [ ] Return appropriate error codes (401 for bad credentials, 429 for rate limit)
- [ ] Log authentication attempts with user ID and IP (no passwords)

## Tests

- [ ] Run `pytest apps/auth/tests/test_login.py -v` - all pass
- [ ] Run `pytest apps/auth/tests/test_logout.py -v` - all pass
- [ ] Manual QA: Verify login with valid credentials returns token
- [ ] Manual QA: Verify 5 failed attempts triggers rate limit

## Completion Criteria

- [ ] Login endpoint returns valid JWT on correct credentials
- [ ] Logout endpoint invalidates the refresh token
- [ ] Rate limiting activates after 5 failed attempts
- [ ] All automated tests pass
- [ ] No security warnings from `bandit` scan

## Notes

- Use `djangorestframework-simplejwt` for token handling
- Token expiry: access = 15 min, refresh = 7 days
- See RFC 6749 for OAuth2 token response format reference

Anti-Patterns to Avoid

• Putting task details in PLAN.md. The master plan is an index only.
• Using arbitrary numbering. Always use 3-digit zero-padded numbers.
• Vague checklist items. "Set up database" is bad; "Create users table with email, password_hash, created_at columns" is good.
• Missing tests. Every task should have at least one test or explicit "N/A - no automated tests applicable" with justification.
• Renumbering tasks mid-project. This breaks references and history.
• Narrative completion criteria. "The feature works well" is bad; "Users can log in and receive a valid JWT token" is good.
• Ignoring dependencies. Always specify what tasks depend on each other to prevent wasted work on blocked tasks.
• Overly large tasks. Tasks with 15+ checklist items are hard to track; split them into focused subtasks.

Plan Lifecycle

When a Plan Completes

When all tasks are checked:

1. Verify the PLAN.md Completion section is fully checked.
2. Add a ## Completed section at the top of PLAN.md with the date:

   ## Completed
   
   **Date:** 2024-02-15
   **Duration:** 3 weeks
   **Notes:** All tasks completed successfully. OAuth added as follow-up plan.
   

3. Optionally move the plan directory to docs/plans/archive/<slug>/ or add an [ARCHIVED] prefix to the directory name.

Plan Retrospective (Optional)

For significant plans, add a RETROSPECTIVE.md in the plan directory:

# <Plan Title> - Retrospective

## What Went Well
- <Positive outcomes and practices>

## What Could Be Improved
- <Issues encountered and lessons learned>

## Follow-up Actions
- <New tasks or plans spawned from this work>

Git Integration Best Practices

• Commit plans early. Commit the initial PLAN.md and task files before starting work so the plan is version-controlled.
• Commit progress atomically. When checking off items, commit the file changes together with any code changes they represent.
• Use meaningful commit messages. Reference the task number: Complete 003-login-endpoint: implement JWT login/logout
• Don't commit partial checkbox states. Either the item is done or it isn't. Commit when work is complete, not in-progress.
• Branch per task (optional). For larger plans, consider a branch per task: plan/user-auth/003-login-endpoint.

Spawning Sub-Plans

When a task is too complex to fit in a single task file (even after splitting), spawn a nested sub-plan:

1. In the parent task file, add a note:

   ## Notes
   - **Sub-plan:** See `../user-auth-oauth/PLAN.md` for detailed OAuth implementation
   

2. Create the sub-plan in a sibling directory:

   docs/plans/
     user-auth/
       PLAN.md
       003-oauth-integration.md  # References sub-plan
     user-auth-oauth/            # Sub-plan for complex OAuth work
       PLAN.md
       001-google-oauth.md
       002-github-oauth.md
   

3. Link the sub-plan completion to the parent task:

   • The parent task's completion criteria should include: - [ ] Sub-plan user-auth-oauth is fully complete

4. When the sub-plan completes:

   • Archive the sub-plan as normal.
   • Check off the parent task that spawned it.

This keeps individual task files focused while allowing complex work to be properly structured.

Compatibility Notes

This skill is designed to work with both Claude Code and OpenAI Codex. The plan directory format is agent-agnostic markdown that any LLM can read, update, and execute.

For Codex users:

• Install via skill-installer with --repo DiversioTeam/agent-skills-marketplace --path plugins/plan-directory/skills/plan-directory.
• Use $skill plan-directory to invoke.

For Claude Code users:

• Install via /plugin install plan-directory@diversiotech.
• Use /plan-directory:plan to invoke.