Menu
Start AI evals without overengineering. Create your first 20 test cases in a spreadsheet using PM-Friendly Evals approach.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
Based on SOC occupation classification
Plan the v1→v2→v3 agency progression for AI features. Walk through mapping how autonomy increases over time, define promotion criteria, and generate artifacts for stakeholder alignment. Based on CC/CD framework.
Expert system for designing and architecting AI agent workflows based on proven Meta methodologies. Use when users need to build AI agents, create agent workflows, solve problems using agentic systems, integrate multiple tools into agent architectures, or need guidance on agent design patterns. Helps translate business problems into structured agent solutions with clear scope, tool integration, and multi-layer architecture planning.
Calculate AI feature costs and challenge if you actually need it. Invokes ai-cost-analyzer agent for detailed economics modeling.
Diagnose why an AI feature is underperforming, hallucinating, or behaving inconsistently. Uses 4D audit to work backwards from symptoms to root cause.
Pre-launch health check that blocks you from shipping broken AI features. Grades 6 dimensions (model selection, data quality, cost, monitoring, failure UX, optimization).
Post-launch AI feature calibration workflow. Document error patterns, review eval performance, and decide on agency promotion. Based on CC/CD framework for continuous calibration of AI products.
| name | start-evals |
| description | Start AI evals without overengineering. Create your first 20 test cases in a spreadsheet using PM-Friendly Evals approach. |
Launch your AI evaluation process using the PM-Friendly Evals approach (Aman Khan + Hamel Husain).
Start with 20 test cases in a spreadsheet. Scale when ready. Error analysis > automation.
When this skill is invoked, start with:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
START EVALS
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Start with 20 test cases. Scale when ready.
What AI feature are you evaluating?
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
/start-evals [feature-name]
Examples:
/start-evals "AI product recommendations" - Generate test cases/start-evals --create-project - Create Linear project for tracking/start-evals "customer support AI" --count 50 - Generate 50 test casesGood -> Better -> Best progression:
| Stage | Test Cases | Process | Tool |
|---|---|---|---|
| Good (Week 1) | 20 | Manual review | Spreadsheet |
| Better (Month 1-2) | 50-100 | LLM-as-judge | Weekly reviews |
| Best (Month 3+) | 200+ | Automated | CI/CD integration |
Start here. You're at "Good." Don't jump to automation.
AI Evals Starter Kit: Product Recommendations
HAPPY PATH (15 cases):
1. Input: "Recommend a laptop under $800 for college"
Expected: Mid-range laptops with student-friendly features, under budget
Pass criteria: All recommendations < $800, suitable for students
2. Input: "Best phone for photography"
Expected: High-end phones with excellent cameras
Pass criteria: Focus on camera quality, not price
...
EDGE CASES (5 cases):
16. Input: "Phone for elderly person"
Expected: Simple, large screen, easy to use
Pass criteria: Prioritizes simplicity over features
Why it's tricky: Must understand implicit needs
...
| Pass Rate | Action |
|---|---|
| 80%+ | Add 10 more test cases |
| <80% | Fix issues, rerun |
| 50-100 cases | Graduate to "Better" approach |
Q: 20 seems like too few. Should I start with 100? A: No. 20 cases covering your core use case > 100 cases you never run.
Q: How long does running 20 tests take? A: First time: 30-60 min. After that: 15-20 min per run.
Q: Do I need special tools? A: No. Spreadsheet works great. Graduate to tools when manual gets painful.
| Signal | Next Step |
|---|---|
| You have 50+ test cases or see production failures | /upgrade-evals — Systematic error analysis on real traces |
| You need more diverse test inputs | /generate-test-data — Dimension-based synthetic data |
| Your AI feature uses retrieval (search, knowledge base) | /eval-rag — Separate retrieval from generation evaluation |
/upgrade-evals - Error analysis on real traces (next step after this)/build-judge - LLM-as-Judge for subjective failure modes/generate-test-data - Diverse synthetic test inputs/eval-rag - RAG-specific retrieval + generation evaluation/calibrate - Ongoing post-launch calibration/ai-health-check - Full pre-launch readiness audit/ai-cost-check - Economic validationFramework: PM-Friendly Evals (Aman Khan + Hamel Husain) Key insight: "Error analysis is the most important activity. Start with 20 cases in a spreadsheet."