Claude Code Plugins

Community-maintained marketplace

Feedback

writing-plans

@ArcadeAI/dexa
0
0

Use when spec is complete and you need detailed implementation tasks for LLM agents. Creates execution plans with exact file paths, complete code examples, and verification steps. Triggers: 'write plan', 'execution plan', 'implementation plan', 'break down into tasks', 'detailed steps'.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md


name: writing-plans description: Use when spec is complete and you need detailed implementation tasks for LLM agents. Creates execution plans with exact file paths, complete code examples, and verification steps. Triggers: 'write plan', 'execution plan', 'implementation plan', 'break down into tasks', 'detailed steps'. allowed-tools: '*'

Writing Execution Plans

Convert specs into detailed implementation plans for LLM agents.

Iron Law: EXACT PATHS. COMPLETE CODE. VERIFICATION GATES.

When to Use

Answer IN ORDER. Stop at first match:

  1. Spec exists and is approved? → Use this skill
  2. No spec yet? → Use brainstorming skill first
  3. Simple task, obvious steps? → Skip (use enforcing-tdd directly)

Phase 1: CONTEXT

Purpose: Understand the spec and identify all affected files.

Protocol:

  1. Read the spec from .safeword/planning/specs/
  2. Identify all files that will be created or modified
  3. Check existing patterns in the codebase
  4. Note test file locations and naming conventions

Exit Criteria:

  • Spec read and understood
  • All affected files identified
  • Existing patterns noted

Phase 2: DECOMPOSE

Purpose: Break the spec into atomic tasks.

Protocol:

  1. Each task = one logical unit (one component, one function, one integration point)
  2. Order tasks by dependency (foundations first)
  3. Each task must be independently verifiable

Task sizing guide:

Too Big Right Size Too Small
"Add auth system" "Add password validation function" "Import bcrypt"
"Build API" "Add POST /users endpoint" "Add route file"

Exit Criteria:

  • Tasks are atomic and ordered
  • No task depends on uncommitted work from another
  • Each task has clear boundaries

Phase 3: DETAIL

Purpose: Write exact implementation details for each task.

Protocol:

For each task, provide:

  1. Files - Exact paths (create/modify/test)
  2. Test code - Complete, runnable test
  3. Implementation code - Complete, minimal code to pass
  4. Verification command - Exact command with expected output
  5. Commit message - Ready to use

Task Format (adapt syntax to project's language/test framework):

### Task N: [Name]

**Files:**

- Create: `src/path/to/file.ext`
- Test: `src/path/to/file.test.ext`

**Step 1: Write failing test**

```
[Complete test code using project's test framework]
```

**Step 2: Verify test fails**

Run: `[exact test command]`
Expected: FAIL - "[specific error message]"

**Step 3: Implement**

```
[Complete implementation code]
```

**Step 4: Verify test passes**

Run: `[exact test command]`
Expected: PASS

**Step 5: Commit**

```bash
git add [files]
git commit -m "[type]: [description]"
```

Exit Criteria:

  • Every task has complete code (no "add validation here")
  • Every task has exact verification command
  • Every task has commit message

Phase 4: SAVE

Purpose: Write the plan to disk.

Protocol:

  1. Create plan file at .safeword/planning/plans/{slug}.md
  2. Include header with metadata
  3. Include all tasks in order

Plan Header:

# [Feature Name] Execution Plan

**Source Spec:** `.safeword/planning/specs/{spec-file}.md`
**Created:** YYYY-MM-DD
**Status:** Ready

**Goal:** [One sentence]

**Tasks:** N tasks

---

Exit Criteria:

  • Plan saved to .safeword/planning/plans/
  • Header includes source spec reference
  • All tasks included

Phase 5: HANDOFF

Purpose: Transition to execution.

Protocol:

  1. Summarize the plan (task count, scope)
  2. Offer execution options:
Plan saved to `.safeword/planning/plans/{slug}.md`

**Execution options:**

1. **Continue here** - I'll execute tasks one-by-one using enforcing-tdd
2. **Fresh session** - Open new session, reference the plan file

Which approach?
  1. If continuing: Start with Task 1, use enforcing-tdd skill

Exit Criteria:

  • User chose execution approach
  • Handed off appropriately

Key Principles

Principle Why
Exact file paths LLMs hallucinate paths without specifics
Complete code "Add X here" leads to inconsistent implementations
Verification gates LLMs need explicit pass/fail criteria
Atomic tasks Fresh context per task = less drift
TDD per task Catches errors immediately

Anti-Patterns

Don't Do
"Add error handling" Show exact try/catch code
"Create test file" Show complete test with assertions
"Run tests" npm test -- path/to/file.test.ts
Combine 3 features in one task One feature per task
Skip verification step Always: Run X, expect Y

Related

  • @./.safeword/guides/planning-guide.md
  • @./.safeword/guides/testing-guide.md