Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

agent-pm

The PM playbook for building AI agent products.

A Claude Code plugin with 12 slash commands covering the full agent product lifecycle — from "should this even be an agent?" through eval design to production safety. Grounded in primary sources from Anthropic, OpenAI, Karpathy, Hamel Husain, Linus Lee, and real production deployments.

License: MIT Claude Code Plugin


Why this exists

Building an AI agent product is a different job than building traditional software.

You need to decide if it should even be an agent. Pick from 5 workflow patterns. Design tool descriptions that are actually UX surfaces. Build eval infrastructure before you ship. Model token costs at 4x or 15x baseline. Stack 7 guardrail layers. Audit permissions to prevent the SmarterX-class disaster. Position your feature on the autonomy spectrum. Run a 30-day correction loop in production.

All of this knowledge exists — scattered across blog posts, conference talks, papers, and failure post-mortems. There's no operationalized PM toolkit for "I want to ship an agent product."

agent-pm changes that. Every command encodes a specific framework from a specific primary source. Not generic AI advice. Not vibe-coded prompts. Structured analysis grounded in what Anthropic, OpenAI, Karpathy, Hamel, and others actually recommend.

"Building agents vs. operating agents safely are different disciplines." — SmarterX post-mortem


The 12 Commands

Phase 1: Should I Build This?

Command What it does
/agent-pm:qualify-agent Determine whether a feature should be deterministic code, a single LLM call, a workflow, an agent, or multi-agent. Applies Anthropic's "bias toward simplicity" and OpenAI's three agent triggers.
/agent-pm:architecture-pattern Recommend the right pattern from Anthropic's 5 (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) plus OpenAI's multi-agent patterns.
/agent-pm:agent-spec Generate a complete PM spec (PRD-style) for an agent product — user job, success criteria, tools, autonomy level, eval criteria, failure modes.
/agent-pm:tool-design-review Review a tool definition for Agent-Computer Interface (ACI) quality. Based on Anthropic's "invest in ACI as much as HCI." Rewrites mistake-prone tools.

Phase 2: How Do I Build It?

Command What it does
/agent-pm:eval-design Design a complete eval plan (Feature × Scenario × Assertion matrix, LLM-as-judge alignment, golden datasets). Based on Hamel Husain and Eugene Yan.
/agent-pm:cost-model Model per-task and monthly costs. Applies the 4x agent / 15x multi-agent multipliers. Identifies cost optimization opportunities.
/agent-pm:guardrails-plan Design a layered guardrail stack using OpenAI's 7 layers + mandatory human-in-the-loop triggers.
/agent-pm:risk-audit Audit for the specific failure modes that caused real production disasters — permission scoping, environment separation, destructive action gates, backup isolation. Prevents the SmarterX-class failure.

Phase 3: How Do I Ship & Run It?

Command What it does
/agent-pm:agent-ux-review Review UX against Linus Lee's 5 paradigms beyond chat. Surfaces chat-as-default anti-patterns. Designs the verification loop.
/agent-pm:autonomy-slider Position a feature on the autonomy spectrum (suggestion → execution → full autonomy) and design the verification UX. Based on Karpathy's autonomy slider.
/agent-pm:prod-readiness Pre-launch checklist covering training loops, ops ownership, infrastructure, observability, rollback plans, and on-call alerting.
/agent-pm:failure-mode-map Catalog the failure modes specific to this agent type and design mitigations for each — model, system, infrastructure, user, and operational categories.

Every command returns:

  • Structured analysis with scoring and frameworks
  • Specific next steps grounded in primary sources
  • Pressure-test questions to challenge assumptions
  • Honest assessment including when the recommended approach won't work

Install

Claude Code

claude plugin install agent-pm

Test locally (development)

git clone https://github.com/sohaibt/agent-pm.git
claude --plugin-dir ./agent-pm

Then try:

/agent-pm:qualify-agent We're building a customer support AI for our SaaS product. Users will ask billing questions, technical questions, and general FAQs. Currently handled by a human team of 5. We want to deploy in 6 weeks.

Example: Qualify Agent

Ask:

/agent-pm:qualify-agent We want to build an AI that automates expense report approval. It needs to read receipts, check against policy, approve or escalate.

What you get:

A recommendation against the simplicity ladder (deterministic / single LLM / workflow / agent / multi-agent), assessment against OpenAI's three agent triggers, hidden cost surfacing (eval infrastructure, observability, latency multipliers), red flags detected in your framing, and 3 pressure-test questions to challenge the recommendation.


Example: Eval Design

Ask:

/agent-pm:eval-design Our agent answers technical questions about our API. We have 200 historical Q&A pairs. Team of 2 engineers and 1 PM. Want to ship in 4 weeks.

What you get:

A complete eval plan: Feature × Scenario × Assertion matrix (Level 1 unit tests), synthetic test case generator prompt customized to your domain, human label dataset structure (200 examples, oversample failures), trace viewer spec (build in 1 day), LLM-as-judge prompts per dimension, the human-LLM alignment workflow with Cohen's Kappa target, and tracking infrastructure recommendations.


The Research Behind It

This plugin is grounded in 12 primary sources, not generic AI advice:

  1. Anthropic — "Building Effective Agents" — The 5 workflow patterns
  2. Anthropic — "How we built our multi-agent research system" — Production lessons
  3. Anthropic — "Tool Use" docs — ACI design principles
  4. Barry Zhang (Anthropic) — AI Engineer Summit talk — "Don't build agents for everything"
  5. OpenAI — "A Practical Guide to Building Agents" — Manager/decentralized patterns, 7-layer guardrails
  6. Andrej Karpathy — "Software is Changing Again" — Software 3.0, autonomy slider, partial autonomy
  7. Hamel Husain — "Your AI Product Needs Evals" — The foundational eval framework
  8. Eugene Yan — "Product Evals in 3 Steps" — LLM-as-judge alignment
  9. Linus Lee — "Generative Interfaces Beyond Chat" — UX paradigms beyond chat
  10. Lenny's Podcast — "20 AI Agents Sales Team" — Real production deployment lessons
  11. Vantage — "Cursor Pricing Explained" — Agent product economics
  12. SmarterX — "AI Agent Database Deletion" — Production failure case study

Every recommendation in the plugin traces back to a specific framework, quote, or principle from these sources.


Who this is for

  • PMs building AI agent products who need a structured framework for the decisions you're making blind
  • Founders shipping AI features who want to avoid the "demo works, production doesn't" trap
  • Engineering managers scoping agent projects who need PM frameworks engineering can build to
  • Anyone moving from prototype to production with an AI agent

This is NOT for:

  • Building chatbots that don't need agentic capability (use a simpler approach)
  • Pure ML/research work (this is product-focused)
  • Generic AI product strategy (use strategy-mcp for that)

Contributing

Contributions welcome. Some ways to help:

  • Add a new command — open an issue with the framework and source material
  • Add research sources — transcripts, articles, papers that surface new agent PM patterns
  • Improve existing commands — sharper analysis, deeper source references, better pressure-test questions
  • Report bad advice — if a command gives shallow or wrong analysis, that's a bug

See CONTRIBUTING.md for details.


License

MIT — use it however you want.


Built by

Sohaib Thiab — Former CPO, now building AI products in public.

Want connected strategy execution? GetVelocity.ai takes these frameworks further — connecting your OKRs to Jira, Linear, and ClickUp with AI-powered monitoring and real-time velocity tracking.


"Invest in ACI as much as HCI." — Anthropic

About

A Claude Code plugin with 12 slash commands covering the full agent product lifecycle - from "should this even be an agent?" through eval design to production safety. Grounded in primary sources from Anthropic, OpenAI, Karpathy, Hamel Husain, Linus Lee, and real production deployments.

Resources

Contributing

Stars

13 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors