---

name: skill-plan description: Creates structured implementation plans before building skills or features. Use when starting a new skill, planning an improvement, or formalizing a build approach. Trigger with "plan this skill", "create a plan for", "skill-plan", "I want to plan before building". allowed-tools: Read, Write, Edit, Glob, Grep, WebSearch, WebFetch

Skill Plan

Creates structured implementation plans before building - the missing link between research and new-skill-wizard.

When This Activates

• "plan this skill"
• "create a plan for [name]"
• "skill-plan for [description]"
• "I want to plan before building"
• "let's plan [feature/skill] first"

The Planning Workflow

Phase 1: Context Loading (Automatic)

Always load these first:

1. CLAUDE.md - Read from project root for universal workflow rules
2. mission-context skill - Invoke for Ed's terminology, frameworks, current focus

Then run semantic search:

Use mcp__smart-connections__semantic_search with the skill/feature name and description as the query. This surfaces:

• Related research swarms
• Prior work on similar skills
• Relevant Zettelkasten notes

Present what you found: "I found [N] related notes including [titles]. Want me to pull any of these in?"

Phase 2: Gather Input (Ask Ed)

Ask these questions ONE AT A TIME:

1. What are you building?

   • Name (kebab-case)
   • One-line description

2. What problem does this solve?

   • The pain point
   • Why existing approaches don't work

3. What does done look like?

   • Paint the picture, not just criteria
   • "Ed triggers X, sees Y, ends up with Z"

If Ed provides a research swarm document or screenshots, incorporate those into "What We Learned."

Phase 3: Generate the Plan

Use this template exactly:

Copy# Build Plan: [Skill Name]

**Date:** YYYY-MM-DD
**Goal:** [One sentence from Ed's description]
**Research:** [[Wikilinks to related notes from semantic search]]
**Status:** Draft

---

## The Problem

[Why this exists - the pain point Ed described]

---

## What We Learned

[From research swarms, semantic search results, screenshots, exploration]
[If no prior research: "Starting fresh - no prior research found."]

---

## Current State

[What exists now - include if improving an existing skill]
[If net-new: Skip this section entirely]

---

## What Done Looks Like

[Paint the picture from Ed's answer. Not just "it works" but what the experience IS when complete.]

---

## The Build (In Order)

### Phase 1: [Name]

**Why:** [Why this comes first]

[Details of what gets built]

**Deliverable:** [What's done when this phase completes]

---

### Phase 2: [Name]

**Why:** [Why this follows Phase 1]

[Details]

**Deliverable:** [What's done]

---

[Continue phases as needed - typically 2-4 phases]

---

## Success Criteria

- [ ] [Specific, testable criterion that proves "What Done Looks Like" is real]
- [ ] [Another criterion]
- [ ] [Keep to 3-5 checkboxes]

---

## Kill Rules

- If [condition], then [pivot action]
- If [condition], then [stop action]

---

Phase 4: Save the Plan

Routing logic:

1. New skill (no existing folder in skills/):

   • Save to Zettelkasten: ~/Documents/COPYobsidian/MAGI/Zettelkasten/Skill Plan - [Name] - YYYY-MM-DD.md

2. Existing skill improvement (folder exists in skills/):

   • Save to: skills/[name]/docs/plans/[feature-name].md

After saving, tell Ed:

Plan saved to [location]. You can now:

• Hand this to new-skill-wizard to create the skill
• Start building directly using this as your guide
• Share it for review before committing

Integration Points

Connects to:

• research-swarm - Prior research feeds "What We Learned"
• new-skill-wizard - Plan becomes the blueprint for skill creation
• lets-wrap - Can compare outcomes to original plan (future enhancement)

Guidelines

1. Load context automatically - Ed shouldn't have to tell you to check CLAUDE.md or run semantic search
2. One question at a time - Don't overwhelm with a form
3. Paint pictures, not checkboxes - "What Done Looks Like" is a vision, not a list
4. Keep phases minimal - 2-4 phases max. More phases = more complexity = more failure points
5. Kill rules are real - Every plan needs explicit "when to stop" criteria
6. Save in the right place - New skills to Zettelkasten, improvements to skill docs

Example Interaction

Ed: "skill-plan for inbox-improver"

Claude:

1. Loads CLAUDE.md, invokes mission-context
2. Runs semantic search for "inbox-improver" - finds 2 related research swarms
3. "I found related notes: 'Research Swarm - Email Triage Tools' and 'Inbox Zero Process Notes'. Want me to pull these in?"
4. Asks: "What problem does inbox-improver solve?"
5. Asks: "What does done look like when it's working?"
6. Generates complete plan following template
7. Checks if skills/inbox-improver/ exists
8. Saves to appropriate location
9. "Plan saved. Ready to hand to new-skill-wizard or start building directly."