Jonathon's AI Wiki

A Self-Improving Claude Cowork System for PM Work (Daniel Bloom, How I AI)

8 min read

Source: raw/I_built_a_Claude_Cowork_system_that_does_a_week_of_PM_work_in_a_day.md — How I AI (host Claire Vo), guest Daniel Bloom, a product manager (employer rendered “Mio” / “Amelio” in auto-captions, probably Melio, unconfirmed), youtube.com/watch?v=p2qmX6TM0kw (fetched 2026-09-29).

Daniel Bloom runs “about 70 to 80%” of his computer time through Claude Cowork and says “I’m really able to do in a day now what used to take me a week”, a self-assessment he stands behind on the show. The system is a Notion board that Cowork built and maintains, two scheduled tasks that anchor his week and his day, and a weekly self-improvement run with four parts. Two pieces are new to this wiki: a context-gap interview, where the system asks him about terms it doesn’t understand, and a draft-versus-sent diff, where it learns from the drafts he rewrote. He also turned a version of the setup into a Cowork plugin for his whole company.

Key Takeaways

  • Two properties matter more than the tool. The system must be able to rewrite its own core files, and it must connect to as much of your work as possible. Bloom says Codex or ChatGPT Work could meet both too. His company chose the stack (Gemini Enterprise first, then Claude Enterprise with Cowork “a few months ago”), and he works within normal token budgets.
  • Do the work inside the agent. Anything he does outside Cowork is context the system doesn’t capture, so he routes Jira, Slack and other work through it. Where there is no connector, the Chrome connector is “a critical … piece”.
  • One context file per topic. Areas of work, goals, colleagues and so on each get a file, seeded by feeding it links and decks and by long Whisper dictation. A recurring job updates the files every few weeks. The test of completeness: a “demo” skill replaced every name and project in his workspace with generic terms in one prompt.
  • The morning brief asks what it doesn’t know. It scans recent Slack, email and Granola notes for terms, files, milestones or goals missing from its context and asks him. Live example: “A term I didn’t know: settlement cap”, which he told it to save. Vo: “something we haven’t seen on the podcast yet.”
  • A weekly run improves the system from four signals: drafts he rewrote, recurring work worth turning into skills, friction logged by each skill, and outside tips checked by a critical “improve” skill (table below).
  • Teach it your habits. He keeps inbox zero on Slack and email, so the system treats a saved Slack message or email that disappears as probably done. Each task also gets a “closed scenario” describing what done looks like, to prepare for the system completing simple tasks on its own.
  • Package it for the team as a plugin. His company’s “workstation” plugin onboards anyone in chat in about 15 minutes (Vo’s estimate, which Bloom confirmed). They chose Cowork over Claude Code after PMs spent about three days installing dependencies in the terminal.
  • Expect the first weeks to feel worse. Until the system knows your “business logic”, it is “very frictionful” and you double-check everything; he says the value compounds if you keep “contextualizing ruthlessly” and centralising work.

The system

Notion board (read-only for him). Three sections: top of mind (large initiatives), this week (priorities), and inbox (items from Slack, email and calendar). Cowork built the board itself to replace a messy Google Doc, and it manages the board; he mostly reads it.

Weekly prep (Sunday morning, the start of his work week). A scheduled task made of several skills:

  • Pulls Notion, calendar, Slack and Granola meeting transcripts.
  • Recommends top-of-mind items to add or remove, and this-week items to move, remove or mark finished; does the same for the inbox.
  • Walks through each meeting of the week and asks how he wants to prepare: a task of its own, a quick reminder, or nothing.
  • After he confirms, it writes the result to Notion.

Morning brief (daily).

  • A one-liner for each of yesterday’s Granola meetings, with any action items; it offers to expand or draft summaries.
  • The context-gap interview: anything in recent Slack, email or notes it doesn’t understand becomes a question, and his answer is saved to the right context file.

Weekly self-improvement run (scheduled).

PartWhat it looks atWhat it produces
How my drafts landedDrafts Claude wrote that he did not send as written, compared with what he actually sentLessons about his writing; in the demoed week, nothing significant
Skills worth buildingWork he does repeatedlySuggested new skills; “many of my skills actually ended up being … from this flow” (the demo suggested a Cowork-to-Claude-Design handoff skill for prototyping)
Fixes to my setupLines in every skill and recurring-task file that log feedback and friction whenever he asks for a fixThe week’s top frictions, each with a proposed fix
Improve skillTips, posts and articles he sends from his phone to a Slack channelA critical review of whether the tip fits his system or is hype. Run on Vo’s loops episode, it said “loops at the end of the day are essentially like scheduled tasks” and that he lacked goal-driven loops

Vo’s reactions: the draft diff “takes the indirect signals of rewriting”, and she has not seen anyone log friction inside their own skills for weekly review (she cites Intercom doing something similar for shared skills).

Scaling it to a company: the workstation plugin

  • What it is. An “operating system” that started for PMs and expanded to all employees, built with colleagues and shared as a plugin in the company library.
  • Onboarding in chat. It connects your tools, confirms your role, maps colleagues and manager, reads your calendar and Slack, builds a writing voice “so that … from day zero, you’re writing in a way that sounds like you”, and captures goals.
  • Why guided UX matters. A year earlier Bloom built a spec-writing Gemini gem that fit how he worked; in live demos other PMs struggled with it. It spread only after he added a guided “here’s how you can work with me” introduction. His conclusion: “good UX and simplicity is critical even for … internal teams and PMs and technical people.”

What’s still missing

  • Running when his computer is off. He calls this the real gap and names Claude Tag (not available at his company, which he has heard is “expensive”), OpenClaw and Hermes as examples. His workaround: a Slack channel Claude polls, where he posts requests while away.
  • Acting on simple tasks alone. The “closed scenario” definitions are groundwork for letting it answer, share a file or schedule a meeting without him.
  • Where the freed time goes. Customer and user conversations, research, and improving the system itself week over week.

Try It

  1. Add a context-gap step to your morning brief: have it scan the last day of Slack, email and meeting notes for terms, files or goals it has no context on, ask you about each, and save the answers to your context files.
  2. Add a line to every skill telling it to log feedback and friction from each run to one file, and have a weekly scheduled task list the top frictions with a proposed fix for each.
  3. Schedule a weekly comparison of the drafts Claude wrote with what you actually sent.
  4. Create an “improve” skill and send it the tips you collect; ask whether each one fits your current setup before you build anything.
  5. If your team uses Cowork, build an onboarding plugin that connects tools, records role and colleagues, and builds a voice profile for each new person.
  6. Marketing teams: run the same draft-versus-sent comparison on emails and social posts; the gaps are the raw material for a brand-voice skill.

Open Questions

  • “A week in a day” and “70 to 80%” are Bloom’s own estimates.
  • There is no data on how often the draft diff produces a change; the demoed week produced nothing significant.
  • The exact wording of the friction-logging lines and the context-file structure was not shown.
  • His employer’s name is unconfirmed, and the site where he says the system can be downloaded is garbled in the captions.
  • Cowork scheduled tasks inside cloud projects now run with no device online (Claude Cowork), which may close part of his “computer off” gap; he does not say whether his setup could use them, and a project tied to a local folder cannot.

Graph View

Backlinks