implementation-plans

This skill guides creation of implementation plans for solo implementers working with AI agents (not teams requiring approval).

Core Principles

Context

• Solo implementer: Bobby + AI agents, not requiring team approvals or stakeholder sign-offs
• Small team: 3 engineers with different focus areas
• Fast execution: Plans completed in days/hours, not weeks
• AI execution model: All implementation steps involve AI agents

Non-Negotiables

1. No timelines or estimates - Provide sequence and dependencies only
2. Validate assumptions first - Ask clarifying questions or flag for verification
3. Brutally concise - If team ignores it, it's too long
4. AI-ready steps - Frame as "Have AI do X" vs "Review/verify Y manually"
5. No stakeholder theater - Skip approval phases, sign-off steps, team alignment meetings

Process

1. Start with Clarifying Questions

Before writing any plan, ask questions to validate assumptions:

• What does the actual data look like?
• Do we have confirmed access/permissions?
• Are there known constraints or requirements?
• Which tools/libraries/versions are we using?

2. Build the Plan

Structure:

## [Task Name]

### Verification Steps (if assumptions can't be validated upfront)
1. Verify [assumption] by [method]

### Implementation
1. Have AI [specific action]
   - Context: [relevant details]
   - Expected output: [what to verify]

2. Review [output] for [specific concerns]

3. Manually verify [critical check]

### Dependencies
- [What must complete first]

### 🚨 Unvalidated Assumptions
- [List assumptions that need verification during execution]

Keep total plan under 500 words. If you need more detail, split into main plan + technical appendix.

3. Focus on Reality

• If you can't validate something with web search or available context, ASK or FLAG it
• Never confidently declare solutions built on unverified assumptions
• When debugging, check actual data before analyzing code
• Remember: obvious data issues > complex code analysis

Common Failure Patterns

❌ The Assumption Cascade

Building multi-step plans on unverified assumptions that collapse when reality doesn't match.

Fix: Front-load verification or flag assumptions explicitly

❌ The Confident Wrong Answer

Declaring root cause without seeing actual data or system state.

Fix: Include data inspection steps before solution steps

❌ The Enterprise Theater

Including approval gates, week-based timelines, team alignment meetings.

Fix: Assume work is approved, sequence steps by technical dependency only

Progressive Disclosure

For detailed guidance, reference these files:

Validation practices: validation-checklist.md

• Comprehensive assumption checklist
• Common technology gotchas
• Data validation patterns

AI delegation: ai-delegation-patterns.md

• How to frame steps for AI execution
• When human review is critical
• Context requirements for AI tasks

Examples: examples/

• Good plan: database migration
• Bad plan: stakeholder-heavy approach
• Complex plan: RAG pipeline implementation

Load these only when additional context would help create a better plan.

Quick Reference

Good step: "Have AI generate migration script adding preferences JSONB column with rollback"

Bad step: "Update the database" (too vague)

Good verification: "Check current schema: SELECT column_name..."

Bad verification: "Ensure database is ready" (unclear how)

Good flag: "🚨 Verify: Azure AI Search tier supports semantic ranking"

Bad flag: Assuming it works without checking