Plan and execute

Plan and Execute

Write the plan before writing the code. Execute it in small batches. Review between batches. This is how you avoid rewriting the same feature three times.

Announce: "Using plan-and-execute to structure this implementation."

Quick Start

For tasks under 30 minutes:

1. List the steps (3-7 concrete actions)
2. Do them in order (one at a time, verify each)
3. Commit after each working state

If the task is bigger than 30 minutes, use the full system below.

The Full System

Phase 1: Write the Plan

Every plan starts with a header:

# [Feature Name] Implementation Plan

**Goal:** [One sentence - what this builds]
**Architecture:** [2-3 sentences - how it fits together]
**Files involved:** [List of files to create or modify]

---

Then break it into tasks. Each task follows this structure:

### Task 1: [Component Name]

**Files:**
- Create: `src/components/widget.tsx`
- Modify: `src/app/page.tsx:45-60`
- Test: `tests/widget.test.ts`

**Step 1: Write the failing test**
[Exact test code]

**Step 2: Run it to confirm it fails**
Run: `npm test -- widget.test.ts`
Expected: FAIL - "widget is not defined"

**Step 3: Write the implementation**
[Exact implementation code]

**Step 4: Run the test to confirm it passes**
Run: `npm test -- widget.test.ts`
Expected: PASS

**Step 5: Commit**
git add src/components/widget.tsx tests/widget.test.ts
git commit -m "feat: add widget component"

Plan Rules

• Exact file paths. Not "the component file" - the actual path.
• Complete code. Not "add validation" - the actual code to write.
• Exact commands. Not "run the tests" - the actual command with expected output.
• One thing per task. If a task touches 5 files, it's 3 tasks.
• 2-5 minutes per step. If a step takes longer, break it down further.

Phase 2: Review Before Executing

Before writing any code, review your plan:

Review Checklist:
- [ ] Every task has exact file paths
- [ ] Every task has a test (or explains why not)
- [ ] No task depends on something not yet built
- [ ] Tasks are ordered so each builds on the last
- [ ] The plan covers the full feature (nothing missing)
- [ ] Nothing in the plan contradicts existing code

If you find a gap, fix the plan first. Never start executing a plan you're not confident in.

Phase 3: Execute in Batches

Default batch size: 3 tasks.

For each task in the batch:

1. Mark it as in_progress
2. Follow each step exactly as written
3. Run verifications at each step
4. Mark as completed when all steps pass

After each batch, stop and report:

• What was implemented
• What verification output showed
• Any deviations from the plan (and why)

Then: "Ready for feedback before continuing."

Phase 4: Adjust and Continue

After feedback:

• Apply requested changes
• Update the plan if scope changed
• Execute the next batch
• Repeat until all tasks are done

Phase 5: Final Verification

After all tasks:

• Run the full test suite
• Check that all planned files exist
• Verify no unplanned changes were made
• Confirm the goal from the plan header is met

When to Use

• Feature with 3+ files to create or modify
• Refactoring that spans multiple modules
• Any task where "just start coding" has bitten you before
• When you need to delegate subtasks (the plan becomes the delegation brief)
• When you want review checkpoints, not a surprise PR

When NOT to Use

• Single-file changes with obvious implementation
• Bug fixes where you already know the cause and fix
• Tasks under 10 minutes
• Exploratory work where you don't know what you're building yet (brainstorm first)

Quick DSV (Decision Check)

Before executing any plan, run 3 questions in 30 seconds:

1. What are the 2-3 key claims in this plan? (What does it assume is true?)
2. What alternative approach haven't I considered? (Is there a simpler way?)
3. Which assumption am I least sure about? (Validate that first.)

If question 2 reveals a genuinely simpler approach, update the plan before starting.

Examples

Example 1: Small Feature Plan

# Add Dark Mode Toggle

**Goal:** Users can switch between light and dark themes.
**Architecture:** CSS variables for theming, React context for state, localStorage for persistence.
**Files involved:** src/styles/theme.css, src/context/theme.tsx, src/components/theme-toggle.tsx

---

### Task 1: Theme CSS Variables
**Files:** Create: src/styles/theme.css
**Steps:** Define light/dark variable sets, add to global styles.

### Task 2: Theme Context
**Files:** Create: src/context/theme.tsx, Test: tests/theme.test.tsx
**Steps:** TDD - write test for toggle behavior, implement context with localStorage.

### Task 3: Toggle Component
**Files:** Create: src/components/theme-toggle.tsx, Modify: src/app/layout.tsx
**Steps:** TDD - write test for rendering, implement button, wire into layout.

Example 2: Handling Plan Deviations

During execution, you discover the CSS variables approach conflicts with the existing Tailwind setup.

Don't: Silently switch approaches and keep going.

Do: Stop the batch. Report:

Deviation: CSS variables conflict with Tailwind's dark mode utility classes.
Options:
A) Use Tailwind's built-in dark: prefix (simpler, less CSS)
B) Override Tailwind's config to use CSS variables (more control, more setup)

Recommendation: Option A - aligns with existing patterns.
Waiting for feedback before continuing.

Delegation Integration

Plans make delegation natural. Each task in your plan can become a delegation prompt:

• Independent tasks? Run them as parallel subagents.
• Complex task? Delegate investigation, keep the implementation.
• Review needed? Delegate to a review agent between batches.

See the smart-delegation skill for prompt templates and model routing.

Anti-Patterns

Pattern	Problem	Fix
"Just start coding"	No plan = rework 3 times	Write the plan first, even a 5-line one
Giant tasks	"Implement auth system" is 10 tasks, not 1	Break down until each task is 2-5 minutes
Vague steps	"Add validation" without specifying what	Exact code, exact file paths, exact commands
No verification	Assuming code works without running it	Every task ends with a test or verification command
Silent deviations	Changing the approach without reporting	Stop, report deviation, wait for feedback
Planning everything	2-day plan for a 10-minute task	Use Quick Start for small tasks
No tests in plan	"We'll test later"	TDD: write the test first, then the code

Blocker Protocol

When you hit a blocker during execution:

1. Stop. Don't try to work around it silently.
2. Document what you attempted and what failed.
3. Report the blocker with context: what's blocked, why, what you need.
4. Wait for guidance. Don't guess past blockers.

If the blocker is in your own code: use the systematic-debugging skill.

With Starter System

If you have the Starter System installed:

• Use /handoff to save plan progress between sessions
• Plans survive session boundaries through handoff files
• Use /remember to log plan patterns that work well for your project

With Kairn (Course)

When Kairn is configured, planning gains:

• AI-reviewed plans - Kairn checks plans against past project patterns before execution
• Blocker detection - "This task structure failed 3 times before - try this instead"
• Plan templates - learns which plan structures work for your codebase

Without Kairn: plans are manual and each session starts fresh. The plan-and-execute cycle still works, you just build institutional knowledge manually.