AI PM Skills Library

Generate acceptance criteria for AI features with non-deterministic outputs. Define what 'done' looks like including quality thresholds, failure categories, and test cases.

Design an AI agent from user goal to autonomous execution. Define what an AI agent does, what tools it has access to, when it acts autonomously vs. pauses for human input, and how failures are handled.

Run A/B tests on AI features correctly. Design experiments for AI-powered features, understand how testing AI outputs differs from deterministic features, and interpret results.

Define what AI-assisted code review should and shouldn't catch. Set up standards for your engineering team on appropriate use of AI code review tools.

Build an AI governance framework for a product team. Define the processes, ownership, and standards that govern how AI features are built, reviewed, and monitored.

Spec AI memory systems for personalisation and agents. Design how an AI product remembers information across sessions, users, or agent runs.

Identify and build defensible AI moats for a product. Assess competitive defensibility, identify which moat types apply, and plan investments to widen the competitive gap.

Design AI-powered onboarding that activates users faster. Personalise, accelerate, or improve the new user experience from first sign-up to activation.

Write specs and briefs that work as effective AI pair-programming inputs for engineers. Structure requirements so engineers can use them directly with AI coding assistants.

Diagnose and strengthen AI product-market fit. Assess whether your AI feature has found PMF, identify blockers, and measure AI-specific fit signals.

Define AI quality metrics that your team can actually track in production. Go beyond vanity metrics and establish a measurement system for output quality, reliability, and user impact.

Run a structured red-team exercise on an AI feature to find failure modes before users do. Identify exploits, misuse vectors, and harmful outputs systematically.

Define AI safety requirements for a product feature. Identify and specify the safety controls that prevent an AI feature from producing dangerous, harmful, or out-of-control outputs.

Build an AI-driven experimentation system that runs, analyses, and learns from product experiments at scale. Automate the hypothesis-to-decision pipeline.

Specify an automated QA pipeline for AI features and AI-generated code. Define the test strategy for features with LLM calls or agent behaviour.

Define guardrails for autonomous AI task execution. Specify safety limits, circuit breakers, and human oversight mechanisms for AI agents that act without step-by-step confirmation.

Evaluate whether to build, buy, or partner for an AI capability. Make a structured recommendation with scorecard, TCO comparison, and exit strategy.

Set up and run an AI-assisted competitor monitoring pipeline. Track competitor activity, analyse strategic implications, and build a repeatable weekly monitoring system.

Set up confidence scoring so users know when to trust the AI and when to verify. Design the scoring system, UX, and calibration process for AI certainty signals.

Manage context windows effectively for complex PM tasks involving large documents, long conversations, or multi-session work. Diagnose context issues and apply practical fixes.

Build and run a continuous discovery system that surfaces user needs consistently. Set up an always-on feedback intake, automated tagging, weekly synthesis, and direct connection to the opportunity backlog.

Decide between copilot and agent interaction models for an AI feature. Choose the right level of AI autonomy using a structured decision framework with five autonomy tiers.

Design a data flywheel that makes your AI product harder to copy over time. Identify what data your product generates, convert it into model improvements, and measure whether the flywheel is spinning.

Choose and spec embedding models for semantic search, RAG, recommendations, or similarity tasks. Compare API-based and open-source models, and write the specification for engineering.

Build an evaluation framework for an LLM-powered feature before it ships. Define what to test, how to score it, and what the bar is for going live with four layers of evals.

Specify explainability requirements for AI decisions in a product. Define how an AI feature explains its outputs to users, satisfy regulatory requirements, and design the explanation UX.

Evaluate and reduce hallucination risk in AI-powered product features. Assess risk, design safeguards proportional to the stakes, and build monitoring for production.

Design human-in-the-loop evaluation workflows for AI features where automated checks aren't enough. Define who reviews, what they look for, how they score, and how findings feed back.

Spec hyper-personalisation features driven by AI. Design a system that adapts product content, recommendations, or experiences to individual users beyond basic segmentation.

Evaluate and specify MCP (Model Context Protocol) integrations for AI products. Decide whether to build or use existing MCP servers, and define security and trust requirements.

Review and write model cards as an AI PM. Evaluate third-party model cards for completeness and relevance, or document internally-built models with proper accountability documentation.

Detect and respond to model drift before users notice. Build a drift detection system, investigate suspected drift incidents, or define the monitoring strategy that catches quality degradation in a live AI feature.

Specify requirements for multi-agent systems. Design orchestration layers, agent contracts, handoff protocols, and failure handling for systems where multiple AI agents collaborate.

Decide between on-device and cloud AI inference for a product feature. Evaluate privacy, latency, cost, capability, and offline requirements using a structured decision framework.

Stress-test a PRD or AI feature spec to find assumptions, gaps, edge cases, and engineer pushback before it reaches the team. Four critic perspectives surface every gap.

Use AI to identify and act on churn signals before users leave. Define a churn prediction system, identify behavioural signals, design interventions, and measure whether anti-churn efforts work.

Design and build prompt chains for multi-step PM tasks where a single prompt isn't enough. Define data contracts between steps, handle failure modes, and build complete chain runner documents.

Write a production-quality system prompt for an AI feature and hand it off as a first-class spec artefact for engineering. Define AI behaviour through prompt design with versioning and test cases.

Choose the right prompting approach for a specific AI feature — between zero-shot, few-shot, chain-of-thought, and RAG. Understand tradeoffs and brief engineering on the right implementation strategy.

Synthesise qualitative research from multiple sources into structured insights without losing nuance. Find patterns, surface contradictions, and produce actionable findings from interviews, surveys, and feedback.

Build responsible AI requirements into a product from day one. Identify ethical risks in AI features, define safeguards before launch, and create a responsible AI checklist across five pillars.

Design semantic search for an AI product. Spec a search experience that finds content by meaning rather than keywords, including query pipeline, indexing strategy, ranking logic, and quality metrics.

Detect and report spec drift between an original AI feature spec and what was built. Compare spec requirements against build behaviours, classify drift types, and generate actionable drift reports.

Turn a completed feature spec into an engineering-ready handoff package. Prepare everything an engineering team needs: spec, test cases, data contracts, design assets, and kick-off brief.

Write AI feature specs using natural language requirements. Turn a rough feature idea into a structured spec written in terms of observable system behaviour with GIVEN/WHEN/THEN format.

Build synthetic persona sets to pressure-test product assumptions when real user access is limited. Explore how different user types would react to a feature before committing to full research.

Write effective system prompts that reliably shape AI behaviour. Diagnose failing prompts, apply a five-layer prompt architecture, and deliver tested, versioned prompts for PM workflows and product features.

Design and run AI-assisted user interviews at scale. Create interview guides, analyse transcripts, and synthesise findings across multiple sessions with three modes: pre-interview, post-interview, and cross-interview.

Evaluate and specify vector database requirements for an AI product. Decide whether you need one, choose between options (managed, self-hosted, pgvector, in-memory), and write implementation requirements.

Use AI coding tools effectively as a non-engineering PM to prototype, explore, and communicate technical ideas. Build quick prototypes, know when to hand off to engineering, and avoid common pitfalls.