How one PM runs a multi-product portfolio on Claude Code.
This repo publishes the operating model I use — the rules, contexts, commands, agents, hooks, and skills that shape Claude Code into a thinking partner across a real portfolio. Not a tutorial. Not a vendor product. Working artefacts a PM can copy into their own .claude/ and start using.
PMs (not engineers) who already know Claude Code exists and want a working operating model — not another article on prompting. If you're running multiple products and want Claude to behave like a senior partner across them instead of a one-shot autocomplete, the patterns here will compound.
A working PM operating model. Here's what's in the repo today, and what's landing next:
| Layer | What it does | Status |
|---|---|---|
CLAUDE.md.template |
The harness. Sets role, north star, product map, how-we-work for one PM running a portfolio. | Live |
rules/ |
Behavioural rules Claude follows every session — confidence tagging, strategic lens, operating discipline. | Live · 3 rules |
contexts/ |
Four operating modes — Strategy / Execution / Review / Analyse — switched per task. | Live · 4 modes |
commands/ |
Slash commands that codify the workflow — session start, mode switching, the /learn → /evolve memory loop, automation, and reporting. |
Live · 8 commands |
templates/ |
Spec template for non-trivial builds. | Live · 1 |
agents/ |
Specialised subagents for repeatable work (exec updates, risk scans, prompt tuning). | Live · 4 agents |
hooks/ |
Safety + automation layer — pre-bash guards, write guards, secret scans, pre-compact saves. | Coming next |
skills/ |
A small curated set of PM skills I've built and use. ~12 sanitized originals. | Planned |
Live = in the repo now · Coming next = the next weekly drop · Planned = on the roadmap.
- Copy
CLAUDE.md.templateto your project root asCLAUDE.md. Fill in your own product map, key dates, and goals. - Copy
rules/,contexts/,commands/,templates/into.claude/in your project (or your user dotfiles). - Run a session. The first command to try:
/context strategyfor a high-altitude session, or/context executionfor hands-on work. - After a non-trivial session, run
/learnto capture what worked into a memory file.
The artefacts are designed to be read in order: CLAUDE.md.template → rules/ → contexts/ → commands/. The whole tree is well under 1,000 lines.
- Not a tutorial. The README is short on purpose. The files are the documentation.
- Not opinionated about your stack — works whether you're using Claude Code, Codex, or anything that reads a CLAUDE.md harness and slash commands.
AI-agent-as-product — about building AI agent products. This repo is about working with one.
Weekly drops. The foundation landed first — harness template, rules, contexts, core commands, and the first round of agents. Each remaining drop layers in another subsystem (hooks next, then skills). See the commit history for what's shipped.
MIT. Take it, fork it, adapt it. If you build a better version, I'd like to read it — link your fork in an issue.
PRs welcome on operating-model patterns (rules, contexts, command protocols, harness shapes). Skills stay personal — fork and build your own.