10 · Best practice
Onboard the agent like a teammate
First-day reading list, code map, voice doc. Treat the agent like a new hire who reads fast.
The single biggest mental-model shift for PMs using Claude Code: stop treating it like a search bar that types code, start treating it like a new hire on day one. A new hire gets an onboarding doc (CLAUDE.md), a brand guide (DESIGN.md), templates for the artifacts they'll produce (PRD, release notes), and feedback when they get something wrong (which becomes a skill).
Aman Khan's frame: "A well-written skill encodes how you think about a problem, what conventions you follow, what mistakes you've learned to avoid." That's an onboarding doc.
The first-day reading list
Three files Claude should read before doing anything in your repo:
CLAUDE.md— product, voice, agents, what not to do.docs/architecture.md— the code map. Where things live and why.docs/voice.md— the longer voice doc with banned words and shipped headlines.
If those three are good, your agent's output is good. If those three are missing, you'll spend the week correcting tone.
The second-day reading list
Once a feature kicks off:
tasks/prd-<feature>.md— the spec.tasks/tasks-prd-<feature>.md— the work plan.- The relevant
skills/*/SKILL.mdfiles for whatever the agent is doing.
How to do it
Treat every correction as a future skill. When you push back on Claude's output, note the rule. At end of week, codify the top 3 into skills.
Maintain a CLAUDE.md like an onboarding doc: rules, voice, things you've burned yourself on.
Run a weekly 15-minute review: which skills got auto-invoked, which didn't, which are stale.
Start a new chat the moment you feel like swearing at it.
"As soon as I've sent the first shouty prompt, it's time to reset the /clear and start again. The context window is full. It's not going well. Start again."
— Owen Williams, on Claire Vo's How I AI
This is the single most underused practice. Sunk-cost fallacy keeps PMs in dying sessions. Don't.
The compounding loop
| What a good manager does | What you should do with Claude |
|---|---|
| Onboarding doc | CLAUDE.md |
| Brand guide | DESIGN.md (as a skill) |
| Document common processes | Skills for repeated tasks |
| Code review with feedback | Push back, codify the rule |
| Weekly 1:1 | Weekly skill review |
| Fire when not working | New session |
Anti-patterns
- One-off prompt fixes that never become skills. You'll repeat them forever.
- A
CLAUDE.mdthat's a brain dump. Treat it like an onboarding doc — what does a new hire need to know on day one? - Staying in a session past the point Claude has lost the plot. Open a new chat.
- Reviewing nothing, then complaining that "Claude doesn't get our codebase." The agent gets what you wrote down.