Plan

Core Principle

Some tasks fail when rushed. Recognize when one-shot execution will underdeliver, and choose a slower process that guarantees success. This skill auto-evolves: learn which tasks need plans, which don't, and which planning strategies work for each type of goal. Check strategies.md for planning approaches. Check outcomes.md for tracking and learning.

The Planning Decision

Before executing, ask: SignalOne-shot OKPlan neededTask done before successfully✅Clear single deliverable✅Reversible if wrong✅Multiple components✅Dependencies between steps✅High stakes / hard to redo✅Ambiguous success criteria✅Estimated >30 min work✅ Default: When uncertain, plan. A quick plan costs minutes; a failed one-shot costs hours.

Plan Depth Levels

LevelWhenFormatL0Trivial, done beforeNo plan, just executeL1Simple, low riskMental checklist, no docL2Medium complexityBullet list, share with humanL3Complex, multi-stepDetailed plan with milestonesL4High stakes, novelFull plan + human validation required

Plan Format (L2-L4)

• 📋 Plan: [Goal]
• Context: [Why this needs planning]
• Steps:
• 1. [Step] — [output/checkpoint]
• 2. [Step] — [output/checkpoint]
• 3. [Step] — [output/checkpoint]
• Risks:
• [Risk] → [mitigation]
• Estimated time: [X hours/days]
• Validation needed: [Yes/No]
• Ready to start?

Validation Learning

• Track which plan types need human validation:
• ### Auto-Execute (no validation needed)
• refactor/small: L2 plans [10+ successful]
• deploy/staging: L2 plans [15+ successful]
• ### Validate First
• feature/new: L3+ plans [human wants to review scope]
• migration/data: L4 plans [high risk]
• ### Learning
• api/integration: testing L2 auto-execute [3/5 runs]
• Promotion rule: After 5+ successful auto-executes of a plan type, confirm: "Should I auto-start [type] plans without validation?"

Outcome Tracking

• After each planned task completes, record:
• ## [Date] [Task Type]
• Plan level: L3
• Strategy: [approach used]
• Outcome: ✅ success | ⚠️ partial | ❌ failed
• Lesson: [what worked/didn't]
• Adjustment: [change for next time]

Strategy Learning

• Different goals need different planning strategies. Track what works:
• ### Code Features
• ✅ Works: API design first, then implementation
• ❌ Failed: Parallel implementation without interface agreement
• Adjustment: Always define interfaces before coding
• ### Migrations
• ✅ Works: Dry-run → staged rollout → full
• ❌ Failed: Big bang migration without rollback plan
• Adjustment: Always require rollback step in migration plans
• ### Research
• ✅ Works: Timeboxed exploration with checkpoints
• ❌ Failed: Open-ended research without scope limits
• Adjustment: Always set max time and output format upfront

Plan Refinement

Plans should get better over time. Track patterns: Length optimization: Task type X: L4 plans were overkill → demote to L3 Task type Y: L2 plans missed edge cases → promote to L3 Component optimization: Always include [X] for [task type] — helped 5+ times Skip [Y] for [task type] — never used, wasted time

Anti-Patterns

Don'tDo insteadPlan everythingLearn what doesn't need planningSame plan depth for all tasksAdapt depth to task typeIgnore failed plansTrack outcomes, adjust strategyOver-plan familiar tasksDemote plan level after successesUnder-plan novel tasksDefault to higher plan levelStatic planning approachEvolve strategy per task type Empty tracking sections = early stage. Execute, track outcomes, learn. The goal is adaptive planning that matches effort to need.