pmstack — PM Toolkit for the AI Era

pmstack

This is how you ship evals as a PM.

[!TIP]

How do you measure the success of AI features?

Take the 2-minute Eval-Readiness check →   ·   8 questions · no email · instant score. Most PMs find out they're shipping on vibes — then it names the one gap to fix first.

An eval is how you prove an AI feature is actually good — and stays good after the next model update. In 2025 it became the line between a real AI PM and one who ships and hopes (OpenAI's CPO, Lenny, and a16z all said as much). But every eval tool is built for ML engineers: dashboards, SDKs, CI pipelines. pmstack is the one a PM runs. Here, /eval is a command — not a $2,000 course.

It's a set of commands you run inside Claude — terminal or phone, no code. The eval lifecycle is the core; the rest of your PM workflow (PRDs, competitive teardowns, launch-readiness gates, the Monday memo) rides along in the same place.

If gstack is the engineer's setup, pmstack is the PM's. Prove you know whether the model is any good. A feature flag won't tell you your model got quietly dumber overnight after a version bump. An eval will.

See it in 30 seconds. Point it at any AI feature — get a real test suite, not advice:

/eval "AI code-review bot that comments on pull requests"

_{A real run — the YAML shown is actual /eval output, not a mockup.}

→ a runnable eval YAML: tasks, graders (code / model / human), negative cases, and the pass^k bar you'd gate a launch on. See a full one › Then /run-eval scores it against the real system and flags the week it regresses.

That's the wedge. The PRDs, voice-of-customer synthesis, and stakeholder briefs are table stakes — every AI writes those. An eval a PM can actually run is the part nobody else ships.

Brand new? Run /onboarding once — it walks the whole stack on a real example, start to finish. If it earns a place in your workflow, star the repo so you can find your way back.

Two kinds of PM, one toolkit

You don't have to be "technical." You have to be able to run a command.

Terminal PM — you live in the shell. Install pmstack as Claude Code slash commands and get one muscle-memory across every artifact you ship.

Phone-and-browser PM — you work in claude.ai on web, desktop, or your phone. Upload the skills to a Claude Project and the same commands fire on plain English. No terminal, ever — same commands, same outputs.

Install

Path 1: claude.ai web, desktop, or mobile (non-technical PMs)

~30 seconds. No terminal.

1. claude.ai → Projects → new Project ("PM toolkit").
2. Settings → Skills → Upload skill → upload each folder from claude-skills/. All 18, or pick the 4–5 you'll use most.
3. Chat normally. "Write a PRD from this customer quote" auto-activates the skill; output appears inline — copy-paste anywhere.

Path 2: Claude Code plugin (technical PMs — recommended)

Two commands inside Claude Code. Versioned updates, clean uninstall, nothing piped to bash.

/plugin marketplace add RyanAlberts/pmstack
/plugin install pmstack@pmstack

Commands arrive namespaced — /pmstack:eval, /pmstack:prd, /pmstack:weekly — and all 18 skills auto-activate on plain English ("write a PRD from this customer quote"). Update later with /plugin, no re-install.

curl -fsSL https://raw.githubusercontent.com/RyanAlberts/pmstack/main/install.sh | bash -s -- --global

git clone https://github.com/RyanAlberts/pmstack.git && cd pmstack
./setup --global    # or: ./setup ~/work/my-pm-stuff

Start here

If you just installed pmstack, run this once. It's the only command you need to remember.

Prefer to browse first?