# PM Agent Stack — The Master Install Checklist

Companion to [The PM Agent Stack series](https://falkster.com/blog/pm-agent-stack-overview). This is the install order that actually works. Print it, mark off as you go, do not jump ahead.

---

## Week 1 — The base (90 minutes)

The agent does nothing useful until the base is in place. Three installs, no exceptions, no "I'll come back to it."

- [ ] **Claude Code** · `npm install -g @anthropic-ai/claude-code`
  - Verify: `claude --version` prints a version number
  - First task: run one real PM task through it tonight (a real interview transcript, a real PR review, a real digest). Not synthetic.
- [ ] **anthropics/skills** (the official skills repo)
  - Add via Claude Code's plugin marketplace; load the pdf, docx, xlsx, pptx skills at minimum
  - Verify: `/skill list` shows the official skills
- [ ] **github/github-mcp-server** (the official GitHub MCP)
  - Configure with a GitHub PAT that has repo:read scope
  - Verify: ask Claude "what issues are open in <repo>" and confirm a real answer

**Stop. Use the base for a week before installing anything else.** If you skip this, the rest of the stack will feel like noise instead of leverage.

---

## Week 2 — The first capability layer (2 hours)

- [ ] **obra/superpowers** · the brainstorm-spec-plan-TDD-review workflow
  - Install command: `/plugin marketplace add obra/superpowers-marketplace`
  - First use: try `/brainstorm` on a real assumption
- [ ] **One MCP server connecting to a system you actually use**
  - Choose one: postgres-mcp (queryable DB), playwright-mcp (web), or a vendor MCP
  - For postgres-mcp: point at a read-only role on a metrics replica, never production read-write
- [ ] **RonitSachdev/ccundo** · granular undo for Claude Code
  - Day-one install. The first time the agent does something destructive, you will be grateful.

---

## Week 3 — Memory and context (90 minutes)

- [ ] **thedotmack/claude-mem** · long-term memory via compression
  - Solves the single biggest pain point of long sessions
  - Install before you feel the pain, not after
- [ ] **yamadashy/repomix** · packs a codebase into one AI-readable file
  - Use when you need the agent to see the whole codebase
  - XML output works particularly well with Claude
- [ ] **wshobson/agents** (or equivalent) · pick 3 to 5 subagents
  - Recommended starter set: code reviewer, research synthesizer, devil's advocate
  - Do NOT install all 25 agents in the collection. Cognitive overhead is real.

---

## Week 4 — The PM-specific layer (one trailing how-to)

Pick ONE of the three trailing how-tos based on which stage of work most needs help right now. Walk through it end to end.

- [ ] **[Discovery agent stack](https://falkster.com/blog/build-discovery-agent-stack)** — gentlest entry point
- [ ] **[Build agent stack](https://falkster.com/blog/build-prototype-agent-stack)** — most transformative
- [ ] **[Measure agent stack](https://falkster.com/blog/build-measurement-agent-stack)** — highest calendar leverage

By the end of week 4 you have a working personal stack actually in use.

---

## Months 2 and beyond — Compounding

- [ ] Add hooks for repeating workflows (scheduled digests, pre-commit reviews)
- [ ] Add slash commands for muscle-memory tasks
- [ ] Read [the prompt engineering guide](https://github.com/ThamJiaHe/claude-prompt-engineering-guide)
- [ ] Subscribe to the awesome lists as feeds; check them monthly
- [ ] Build your own skills when the existing ones don't fit your work

---

## Critical limitations to keep in mind

- Single-PM scope. The agent only sees what you can give it access to.
- No team-wide memory. Each PM on the team rebuilds.
- Limited write access. Reads are easy; writes are fragile.
- Security boundary is your laptop. Treat data the way you treat it today.
- Open-source repos drift over time. Curation is ongoing work.
- Cognitive overhead of installed tools is real.
- Does NOT replace the data team, security team, or legal team.

---

## When the enterprise version arrives

The skills, subagents, prompts, and workflows you built migrate. The MCPs and memory layer get replaced by the enterprise platform's native equivalents. The institutional knowledge of which workflows save PM time lives in the people who built personal versions first.

**Sources:** [PM Agent Stack series on falkster.com](https://falkster.com/blog/pm-agent-stack-overview) · [The PM Agent Stack handbook chapter](https://falkster.com/os/pm-agent-stack)