Docs / Claude Code for product managers

Guide

Claude Code for product managers: what it covers, and where a team record takes over

Claude Code works for product managers as a file-based assistant. It reads your notes, interview transcripts and specs from a folder, drafts PRDs, research syntheses and stakeholder updates, and reaches Slack or Linear over MCP. It covers solo work well. It stops where a second person needs your decisions, their reasons, and a later check on whether they worked.

Read against Anthropic's Claude Code docs, the product-management plugin README and Anthropic's post on how its own teams use Claude Code on 8 October 2026. Anthropic ships changes weekly, so check the docs for the version you run.

What can a product manager do with Claude Code?

The jobs below come from Anthropic's product-management plugin, which its README describes as built primarily for Cowork, Anthropic's desktop agent, and also working in Claude Code. The table maps each PM job to what the plugin does and what you set up first.

Two details from the README worth knowing before you install. The plugin's included connections cover Slack, Linear, Asana, monday.com, ClickUp, Atlassian, Notion, Figma, Amplitude, Pendo, Intercom and Fireflies, so most PM stacks have a path in. And the install is two commands in a terminal: claude plugin marketplace add anthropics/knowledge-work-plugins, then claude plugin install product-management@knowledge-work-plugins.

The honest limit is that a command is a starting template. The README's own brainstorm example shows the better use: Claude asks "what problem are your users hitting with search today?" before it generates anything. You still bring the problem, the customers and the judgment.

How to set Claude Code up as a non-engineer

Six steps, in the order that pays off fastest:

1. Install it on a surface you are comfortable with. Anthropic offers a terminal, a desktop app with a Code tab, a web version and IDE extensions. Most surfaces need a Claude subscription or an Anthropic Console account. If you have never opened a terminal, Anthropic's terminal guide covers it.
2. Make one folder per product. Put research, specs, decision notes and the current roadmap in it as plain text or markdown. Claude Code reads what is in the folder, so a messy folder gives messy answers.
3. Write a CLAUDE.md at the folder root. Who the users are, what the product does, your glossary, your PRD template, and "always do X" rules such as "cite the interview a claim came from". Anthropic recommends staying under 200 lines, because longer files reduce adherence, and putting multi-step procedures in skills instead.
4. Add skills for what you repeat. Install the plugin above, or write your own for a task you do twice a month. Anthropic describes skills as repeatable workflows a team can share.
5. Connect the tools your signal lives in. MCP, the open standard for connecting AI tools to outside data, is how Claude Code reads Google Drive, updates Jira tickets or pulls Slack threads.
6. Schedule what recurs. Anthropic's routines run in the cloud, so they keep running with your computer off. Desktop scheduled tasks run on your machine with access to local files. A Monday draft of the weekly update is a good first one.

A worked example, made up to show the shape and not a customer story. A PM at a nine-person B2B startup drops eight interview transcripts into a research folder on Monday and runs the synthesis. Claude returns four themes with the quotes behind each. She asks it to draft a one-page spec for the strongest theme using her template, and it flags that two of the eight interviews contradict the requirement she wrote. She rewrites the requirement herself. Total time: an afternoon, where it would have been two days. Everything above lives in one folder on one laptop.

That last sentence is the whole next section.

What Claude Code remembers, and what it does not

Anthropic is direct about this: each session begins with a fresh context window, and two mechanisms carry knowledge forward.

• CLAUDE.md files are instructions you write. A project CLAUDE.md is shared with teammates through version control, so sharing it means putting the folder in git. A CLAUDE.local.md holds personal preferences and stays out of it.
• Auto memory is notes Claude writes for itself from your corrections. It is stored under your home directory, and Anthropic states that it is machine-local: files are not shared across machines or cloud environments. Only the first 200 lines or 25KB of its index load at the start of a session.

For a PM, that means your corrections, your glossary and your style are remembered on your laptop. Your teammate's Claude Code does not know them unless you handed over the folder. The reasoning behind last Tuesday's call to cut a feature sits in a chat session, not in either file. Agent memory across sessions covers what an agent can and cannot carry from one run to the next.

When Claude Code alone is enough

Do not add anything if all of these are true:

• One person owns the product decisions. You are the only human who needs the reasoning.
• The work is a draft you will review and send. A spec, a synthesis or an update is consumed once, by you, and then lives wherever you already publish it.
• Nobody will ask "why" in a month. If the answer is in your head and the head is yours, a folder is a record.
• You can check outcomes yourself. You will look at the metric anyway.

Two of the guides ranking for this keyword, the free ccforpms.com course and ProductCompass's beginner guide, teach this solo setup well. As far as we could read on 8 October 2026, neither covers sharing with a team, which is the gap this page is about.

Where it stops: four boundaries for a PM

None of these are bugs. They are what you meet when the work stops being yours alone.

1. A second person needs the work. A designer or engineer asks what the research said. The honest answer is "it's in my folder". Even if you share the folder through git, you have shared files, not who owns what.

2. A decision needs its reason. A PRD records what was chosen. It rarely records the option you rejected, who overruled whom, or what you decided to leave alone. That context lives in the chat session that produced the draft.

3. A bet needs an outcome check. "The spec shipped" and "the pain went away" are different facts. Nothing in a folder of documents tells you which of your specs were tested against the number they promised. This boundary is our inference, not something Anthropic's docs address.

4. A signal needs to link to the decision it moved. The interview synthesis from March, the support-ticket pattern from April and the roadmap change in May are three files. Which one caused the other is a fact someone has to write down, or it disappears.

What a team record has to hold

You do not need a product for this. You need four things written down somewhere more durable than a session:

1. An owner per piece of work. One name, person or agent, on each item.
2. A decision with its reason. What was chosen, what was rejected, who approved.
3. A stated outcome condition. What "worked" means in terms you can check later.
4. Links between them. The signal that started it, the decision it led to, the tasks that carried it out.

A tracker you already use can hold this if you add fields by convention. Linear, Notion and GitHub Issues all work. The weak point is the convention: nobody enforces it, so the reason and the outcome condition are the first fields to go blank.

How Maskin keeps that record

Maskin is an open-source (Apache 2.0), MCP-native, self-hostable workspace. We build it, so weigh this section accordingly. Work moves as typed objects. A signal enters as an insight, from sources such as Slack, Intercom, PostHog or a scheduled scan. It is shaped into a bet, a testable hypothesis with a win condition (bet-based product planning explains the method). The bet breaks into tasks that people and agents share, each with a driver (an owner, person or agent) and a comment thread. When the work ships, the outcome is checked against the win condition.

For a PM already using Claude Code, two things matter. First, Maskin exposes its objects over MCP, and Claude Code connects to MCP servers, which is the route for pointing a session at the shared record instead of a private folder. The MCP tools reference lists what is exposed. Second, Maskin runs agent sessions on the Claude Code and Codex command-line tools, each in an isolated microVM, so the agents are the ones you already use. What changes is where the result is kept.

The product discovery loop is the closest match to the PM job described above. Interviews, tickets and reviews arrive as typed insights, agents cluster themes and draft candidate bets, and a person reviews and shapes each bet before it is approved. After it ships, the loop re-checks the signal to see whether the bet retired the pain. The human gate is the point: the agent gathers evidence and you own the call. See what an agentic workspace is for the full lifecycle and the closed-loop workspace for the longer argument.

Where Maskin is the wrong pick:

• You work alone. A folder and a good CLAUDE.md are lighter and cost nothing.
• You need a model other than Claude or Codex. Sessions run those two command-line tools, and local models are not a path we have validated.
• You need memory across conversations to be solved today. It is a known gap. Read Agent memory across sessions for what survives between scheduled runs.
• You do not want to run infrastructure. Self-hosting has a hard requirement on the agent-server host. The self-hosted workspace guide lists it, and there is a hosted option.
• It is early. Judge the architecture, not a long track record. AI product workspace vs AI coding agent covers where it sits next to the tool you already run.

A quick decision rule

• One PM, one product, work that ends in a draft you send: Claude Code with a folder and a CLAUDE.md. Stop there.
• You want it faster this week: install the plugin and connect your tracker and chat.
• Two or more people read your decisions: write down owner, reason and outcome condition in a tracker before you add more skills.
• You are asked "did that bet work?" and cannot answer from a document: you need the outcome condition on record, which is when a workspace like Maskin earns its place.
• A human must approve before anything side-effecting runs: see human in the loop AI.

FAQ

How do I use Claude Code as a product manager?

Install Claude Code, put your research, specs and roadmap in one folder, and write a CLAUDE.md describing your product and users. Then ask it to do PM jobs in plain language: synthesize interview notes, draft a PRD from your template, or write a stakeholder update. Add skills for anything you repeat, and connect tools over MCP when you want it to read your tracker.

Do product managers need to know how to code to use Claude Code?

No. Anthropic's write-up of its own teams says legal, growth marketing and product design staff use Claude Code without being engineers: lawyers prototyped a phone tree, and a growth team built a workflow that processes spreadsheets of ads. You do need to be comfortable giving instructions in text and, on some surfaces, opening a terminal. The desktop app's Code tab avoids that.

What are the best Claude Code skills for product managers?

Start with the seven in Anthropic's product-management plugin: feature specs, roadmap management, stakeholder communications, user research synthesis, competitive analysis, metrics tracking and product brainstorming. Then write your own. ProductCompass's beginner guide suggests building a skill from a task you already do twice, which is a good filter: skills for one-off work cost more to maintain than they save.

Can Claude Code connect to Jira, Slack or Linear?

Yes, through MCP, the open standard for connecting AI tools to outside data. Anthropic's overview lists reading Google Drive docs, updating Jira tickets and pulling Slack data. The product-management plugin's README also names Linear, Asana, monday.com, ClickUp, Notion, Figma, Amplitude, Pendo, Intercom and Fireflies.

Is the Anthropic product-management plugin for Claude Code or Cowork?

Its README says it was designed primarily for Cowork and also works in Claude Code, and it gives the Claude Code install commands. On 16 September 2026 Anthropic announced that Cowork and chat are merging into one Claude, rolling out to Pro and Max first. The announcement is about Cowork and chat, so check where the plugin installs on your plan.

Can my team share what Claude Code learns?

Only some of it. A project CLAUDE.md is shared through version control, so teammates who clone the folder get your instructions. Auto memory is machine-local and is not shared across machines. Decisions, reasons and outcomes are not stored by either mechanism, so a team needs a tracker or workspace that holds them outside any session.

How much does Claude Code cost?

Anthropic says most surfaces require a Claude subscription or an Anthropic Console account, and the pricing changes often, so this page carries no prices. Check Anthropic's pricing page for your plan.