A product manager's week with agents: briefs, research synthesis, specs and tickets
Publicado el 2026-10-12
In August 2026, Lenny's Newsletter profiled Daniel Blum, a product manager at Melio who hands 70% to 80% of his workday to Claude running in Cowork, Anthropic's agentic desktop app. The habits worth copying: a daily pass over Slack, email and notes, a spec-writing helper, and a weekly loop that compares Claude's original drafts with his final versions.
This guide turns that into a week you can set up yourself. It is for PMs and for founders who are still their own PM. No coding is needed: Cowork installs plugins from a web page, and Claude Code or Codex need a few commands pasted into a terminal.
The recurring jobs:
- Every morning: a brief of what changed and what is waiting on you.
- Weekly: interviews and support tickets turned into themes, with quotes you can defend.
- When a bet is approved: a spec built from a measurable goal, then tickets engineering can pick up in order.
- Before the product review: adoption and experiment numbers.
- Friday: the stakeholder update.
Before you start: pick your route
You need one agent, and each needs a paid plan. Cowork is included in Claude's Pro, Max, Team and Enterprise plans. Claude Code needs a paid Claude plan or Anthropic API billing. Codex comes with ChatGPT plans that include it, or with an OpenAI API key.
- Cowork, the route Blum uses: open a plugin's page on claude.com/plugins and choose Install in Cowork. No terminal.
- Claude Code: run the
claude pluginandclaude mcpcommands below in a terminal (Terminal on a Mac, PowerShell on Windows). Then type/mcpinside Claude Code to see each connection and sign in. - Codex: Anthropic's plugins have no Codex version, so install the standalone skills and the Codex plugins named in each section; the skill installs need Node.js. The prompts below use the slash form, such as
/write-spec. In Codex, start the prompt with$write-specinstead: OpenAI's docs say to type$to mention a skill, or run/skills.
| Step | Cowork or Claude Code | Codex | What you need first |
|---|---|---|---|
| Morning brief | Productivity; Slack plugin for channel digests | Notion (Codex plugin); Codex gets Slack's skills but not its server | Slack admin approval; a mail connection |
| Research synthesis | Product Management | synthesize-research | Node.js for the Codex install |
| Spec | Product Management | write-spec | Node.js for the Codex install |
| Tickets | Product Management, which connects Linear and Jira | Linear (Codex plugin) or Atlassian MCP | Linear or Atlassian Cloud account |
| Engineering task list | Task Master, set up by an engineer | Task Master | A model API key, or a Claude Code or Codex login |
| Review prep | PostHog MCP in Claude Code; Amplitude or Pendo through the plugin in Cowork | PostHog MCP | PostHog, Amplitude or Pendo account |
| Friday update | Product Management | stakeholder-update | Node.js for the Codex install |
If your company approves only Copilot, read the section on that near the end first.
Anthropic's Product Management plugin covers the research, spec, update and metrics steps in one install. It adds /write-spec, /synthesize-research, /stakeholder-update, /metrics-review, /roadmap-update, /competitive-brief and /brainstorm; the long form is /product-management:write-spec and so on. Its README says it is designed mainly for Cowork and also works in Claude Code. Its config already includes the hosted Slack, Linear, Notion, Asana, monday.com, ClickUp, Atlassian (Jira and Confluence), Figma, Amplitude, Pendo, Intercom and Fireflies servers, and you sign in to each one the first time you use it. The email and calendar entries are left empty.
claude plugin marketplace add anthropics/knowledge-work-plugins && claude plugin install product-management@knowledge-work-plugins
Avoid double connections. The Productivity plugin in the next section includes the same Slack, Notion, Linear and Jira servers, and the vendors' own Slack, Notion, Linear and Atlassian plugins connect to them once more. Each copy is a separate sign-in, so pick one route per service and add a vendor plugin only for its extra commands. With the Product Management plugin installed, skip the standalone skill installs further down: they are the same skills.
Morning: a brief from Slack, email and your notes
Hand over the reading. The agent goes through channels, threads, email and your notes since yesterday and returns decisions made without you, questions waiting on you, and anything that moves a date or scope.
Blum's version adds a useful step: the agent looks for terms, projects or goals missing from its context files and asks him targeted questions to fill those gaps. Anthropic's Productivity plugin is built around the same idea. /start creates a task list (TASKS.md), a memory (a CLAUDE.md file plus a memory/ folder) and a local dashboard, then asks about your role, team and priorities. /update --comprehensive scans email, calendar and chat for missed to-dos and suggests new memories, such as a person who keeps coming up in threads. Its README says it is designed mainly for Cowork and also works in Claude Code. There is no Codex version.
claude plugin marketplace add anthropics/knowledge-work-plugins && claude plugin install productivity@knowledge-work-plugins
Connect email first. The plugin's config leaves its Gmail and Google Calendar entries empty, so it cannot read your inbox until you connect one. In Cowork, add Gmail or Microsoft 365 from Claude's connector directory, if your plan and admin allow it. In Claude Code or Codex with Outlook, the community-maintained Microsoft 365 MCP Server reads mail and calendar through Microsoft Graph and has a read-only mode; work accounts need its organization mode, so ask IT before connecting a company mailbox. If your day is spread across more tools, Enterprise Search adds /enterprise-search:digest --daily, which its README says highlights action items, decisions and mentions of you across connected chat, email and docs.
For a Slack-only start, Slack's own plugin adds /slack:channel-digest payments-squad, product-leads for a digest across several channels, plus /slack:summarize-channel and /slack:standup.
claude plugin install slack@claude-plugins-official
The plugin README says your workspace admin must approve the MCP integration before you can sign in (Slack's docs say admins approve and manage MCP clients). In Codex the plugin ships only the skills, not the MCP server, so Codex cannot read your Slack through it yet.
claude plugin install notion@claude-plugins-official
It adds commands such as /Notion:search and /Notion:database-query, plus four skills including meeting prep, on top of the Notion connection the two Anthropic plugins already have. Claude Code only. In Codex, use Notion, Notion's plugin in OpenAI's Codex marketplace, which needs Notion authorization:
codex plugin marketplace add openai/plugins && codex plugin add notion@openai-curated
If you use the Productivity plugin, run /update --comprehensive first. Then:
Check #payments-squad, #product-leads, my DMs and my inbox since 6pm yesterday, plus my Notion page Weekly plan.
Give me: 1) decisions made without me, 2) questions waiting on me, with links, 3) anything that moves a date or scope.
Do not send or post anything. List any term or project you don't recognize and ask me about it before saving it to memory.
Without the Productivity plugin, in Codex for example, keep a short context.md (team names, code names, current goals, a glossary) in the folder where you run the agent. Codex reads a file named AGENTS.md there on its own; any other name is read only when you ask, so start the prompt with Read context.md first. and ask the agent to add your answers there.
Weekly: research synthesis with quotes you can defend
Hand over the first coding pass. synthesize-research takes pasted transcripts, uploaded files or connected sources, pulls out observations and verbatim quotes, groups them into themes, and ranks findings by frequency and impact, each with a confidence level. It aims for five to eight findings and ends with open questions. It comes with the Product Management plugin.
For Codex only (skip it if you installed the plugin; needs Node.js for npx):
npx skills add anthropics/knowledge-work-plugins --skill synthesize-research -a codex -y
The skill can only count what you give it: six interviews with power users say nothing about the people who churned. Its own guidance says a finding from two interviews is a hypothesis, not a conclusion, and that quotes should be attributed to participant type, not name.
/synthesize-research Why do new admins stall before inviting their team?
Sources: the 9 transcripts in ./interviews and support-tickets-sept.csv.
For every finding, give each quote verbatim with file name and line so I can check it. Attribute quotes by role and company size, never by name. Mark anything with fewer than 3 sources as low confidence.
Then open three quotes at random and find them in the source. If one has been paraphrased, ask for a redo.
When a bet is approved: a spec from a measurable goal
Hand over the first draft, and the questions it should ask you. write-spec asks about the user problem, target users, success metrics, constraints and prior attempts, then drafts a PRD (product requirements document) with goals, non-goals, user stories, P0 to P2 requirements with acceptance criteria, leading and lagging metrics, and open questions tagged with who should answer them. Its guidance says goals should be outcomes, not outputs. It comes with the Product Management plugin.
For Codex only (skip it if you installed the plugin):
npx skills add anthropics/knowledge-work-plugins --skill write-spec -a codex -y
/write-spec Goal: raise the share of new workspaces that invite a second member within 7 days from [current]% to [target]% by end of Q1, measured as the invite_sent event in PostHog.
Evidence: ./synthesis-oct.md. Non-goals must include SSO.
Ask me your questions one at a time before drafting.
Add the weekly learning loop. Blum's loop compares Claude's original drafts with his final versions once a week, finds the most common problems and recommends updates. Do the same with two folders: save each draft untouched in drafts/ and what you actually shipped in finals/, then on Friday:
Compare each file in ./drafts with its match in ./finals from this week. List the edits I made most often: tone, structure, what I cut, what I added.
Propose at most five changes to my instructions file (CLAUDE.md if you use the Productivity plugin, otherwise context.md or AGENTS.md) as a diff. Don't apply them until I say so.
Handoff: from PRD to tickets engineering can pick up in order
Hand over the splitting. If you installed the Product Management plugin, Linear is already connected, and so are Jira and Confluence through its Atlassian connection. You sign in as yourself, so the agent acts as you. Without the plugin, the Linear plugin connects Linear's hosted MCP so the agent can create and update issues, manage projects and states, and search your workspace. You need a Linear account and OAuth sign-in. Claude Code only.
claude plugin install linear@claude-plugins-official
In Codex, use Linear (Codex plugin), Linear's plugin in OpenAI's Codex marketplace, which needs a Linear account and authorization:
codex plugin marketplace add openai/plugins && codex plugin add linear@openai-curated
All of these connect Linear's full endpoint, which can write. While you only want the agent to look, Linear's MCP docs offer a read-only endpoint that exposes only read tools. Add it as a server of its own, then run /mcp in Claude Code to sign in. Creating the issues later still needs a full connection.
claude mcp add --transport http linear-readonly https://mcp.linear.app/mcp/readonly
If your team tracks work in Jira, Atlassian MCP does the same handoff: it searches and creates Jira issues, reads and writes Confluence pages, and manages sprints. You need an Atlassian Cloud account with OAuth or an API token.
# Claude Code
claude plugin install atlassian@claude-plugins-official
# Codex
codex plugin marketplace add openai/plugins && codex plugin add atlassian-rovo@openai-curated
If your engineers build with coding agents such as Claude Code, Cursor or Codex, Task Master parses a PRD into a dependency-aware task list and tells the agent which task comes next. An engineer adds it to the code project, where it keeps its own task list; your part is handing over the PRD. After the install command, its README says to add a model API key to the project's .env file, or to tell the agent Change the main model to claude-code/sonnet to reuse a Claude Code login (a Codex CLI login also works), then Initialize taskmaster-ai in my project. It needs Node.js. Agree with your tech lead whether its list or your tracker is the record.
# Claude Code
claude mcp add taskmaster-ai -- npx -y task-master-ai
# Codex
codex mcp add taskmaster-ai -- npx -y task-master-ai
The no-invented-dependencies rule below comes from the planning example in Linear's docs.
Read prd-invite-flow.md. Draft, but do not create, a project with milestones and issues. Each issue: problem, goal, approach if the PRD gives one, acceptance criteria copied from the PRD, open questions.
Mark an issue as blocked by another only where the PRD supports it; don't invent dependencies. Show me a table first. After I approve, create them in team PAY in Linear (or project PAY in Jira).
Before the review: adoption and experiments in PostHog
Hand over the queries. PostHog MCP lets the agent query product analytics with HogQL (PostHog's SQL), manage feature flags and A/B experiments, and look into errors and session replays. You sign in to PostHog to authorize. Connecting and calling tools is free, but standard PostHog usage billing applies to the data your queries consume, and a few LLM-powered tools count as PostHog AI spend. The PostHog MCP FAQ says the server does not store your analytics data and routes you to the US or EU region of the account you sign in with; the README also lists mcp-eu.posthog.com for EU users.
The same server can change things: the README's own examples include creating a feature flag rolled out to 20% of users. For review prep, use the read-only mode described in the FAQ. With ?readonly=true on the URL, every tool that creates, updates or deletes data is left out. Keep the quotes around the URL, because zsh, the default shell on a Mac, treats an unquoted ? as a wildcard.
# Claude Code
claude mcp add --transport http posthog "https://mcp.posthog.com/mcp?readonly=true"
# Codex
codex mcp add posthog --url "https://mcp.posthog.com/mcp?readonly=true"
The catalog's commands use the same URL without ?readonly=true; switch to those only when you want the agent to manage flags or experiments. The README's features and tools filters only narrow which tools appear, and do not make them read-only. In Cowork, or if your analytics live in Amplitude or Pendo, the Product Management plugin's /metrics-review can use those connections instead.
For the invite-flow experiment: exposure per variant, conversion to invite_sent within 7 days, and what PostHog reports about significance. Show every query you ran. Do not change any flag or experiment.
Friday: the stakeholder update
Hand over the formatting, keep the judgment. stakeholder-update asks for the type (weekly, monthly, launch or ad-hoc) and the audience (executives, engineering, partners, customers or the board), pulls from a connected tracker, chat and meeting notes when available, and writes for that reader. For executives that means a status color, a one-line summary, progress tied to goals, risks with mitigations, decisions needed and next milestones. It can rewrite the same progress for a second audience. It comes with the Product Management plugin.
For Codex only (skip it if you installed the plugin):
npx skills add anthropics/knowledge-work-plugins --skill stakeholder-update -a codex -y
/stakeholder-update Weekly, for the exec team.
Use the project PAY-invite in our tracker (closed and at-risk issues), the PostHog numbers in thursday-notes.md, and decisions in #payments-squad this week.
Leave the status color blank and give me your suggestion with reasons. Then write a five-line version for the engineering channel. Don't post anything.
The skill's own tip applies: the color should be your honest assessment, not what you think leadership wants to hear.
If your company approves only Copilot
First check which Copilot. Microsoft 365 Copilot is the assistant inside Outlook, Teams and the Microsoft 365 app. GitHub Copilot is a coding assistant (Copilot CLI and the VS Code extension), a separate product with its own license. None of the tools in this guide plug into Microsoft 365 Copilot Chat itself, so if that is the only assistant approved, the installs here don't apply. You can still paste the prompts from this guide; what it can see depends on your license and your IT settings.
If GitHub Copilot is approved too, Microsoft Work IQ connects it to your Microsoft 365 data: email, meetings, OneDrive and SharePoint files and Teams messages, through plain-language questions. Microsoft's admin guide requires a Microsoft 365 Copilot license for each user, and the README adds tenant admin consent and Node.js 18 or later. It is in public preview, so features may change.
The README's quick start runs inside GitHub Copilot CLI:
copilot
/plugin marketplace add microsoft/work-iq
/plugin install workiq@work-iq
For a morning brief, the README also lists workiq-productivity, a read-only plugin for email triage, meeting costs, org charts and channel audits, installed with /plugin install workiq-productivity@work-iq. If Claude Code or Codex is approved as well, the catalog commands are claude mcp add workiq -- npx -y @microsoft/workiq mcp and codex mcp add workiq -- npx -y @microsoft/workiq mcp.
The three skills in this guide are plain instruction files, and the skills CLI lists github-copilot as a target agent. Swapping -a codex for -a github-copilot in their install commands puts them into GitHub Copilot, not Microsoft 365 Copilot Chat. Ask IT first: an approved assistant does not always mean approved add-ons.
Limits and risks
- The agent has your access. Slack, Notion, Linear, Jira and PostHog connect through your own sign-in. If you can read a private leadership channel, so can the agent, and anything it puts in a summary can be forwarded. Use read-only modes where they exist (PostHog's
readonly=true, Linear's read-only endpoint) and keep do-not-send lines in your prompts until you trust the setup. - Customer data. Transcripts and tickets carry names, emails and sometimes account details. Remove them before synthesis, follow your company's list of approved AI services, and do not paste customer data into tools nobody has vetted.
- Text that looks like instructions. An agent reading your inbox reads whatever strangers send you. PostHog's FAQ gives the general rule: review MCP tool calls before they run. Anything the agent wants to send, create or delete because of something in an email or ticket should come to you first.
- Invented quotes and numbers. Spot-check quotes against sources and ask for the query behind every metric before it goes in a deck.
- Colleagues' AI slop. Critique the document, not the tool: ask which goal it serves and where each number comes from. A useful team norm: agent drafts are fine, but the author must have checked every quote and number and be able to defend them in review.
What stays with you
The agent can rank themes by frequency. It cannot decide that the segment you are betting on matters more than the loudest one. Prioritization, saying no to a stakeholder, choosing the non-goals, calling the status color and negotiating scope with engineering stay with you. Faster first drafts only give you more time for them.