Video summary

My Full Claude Cowork Setup (steal my workflows!)

Main summary

Key takeaways

Product Review

Product reviewed

Claude Co-work / “Claude co-work” setup — a local AI agent workflow for building software and automations — plus a Hostinger Claude Code VPS option for running it 24/7.

Key features mentioned

  • Daily “morning digest” delivered to Apple Notes, including:
    • investments summary
    • calendar events
    • emails and inbox items
    • action items for the day
  • Custom dashboards, including:
    • Investments dashboard: track/update investments and get news/advice aligned with an investing philosophy
    • Mission control dashboard: monitor projects/builds/schedules
  • Autonomous building (nightly builds):
    • PRD-first approach: Claude drafts a PRD; the user approves/signs off before building
    • A scheduled job processes PRDs from folders using a workflow like pending → in progress → done/failed
    • Build logs are kept
    • Example output: a resource hub built overnight
  • Core build methodology using a meta prompt with instruction sections:
    1. Building things: PRD first, open questions, explicit sign-off before building
    2. Push back & clarification: Claude challenges inconsistent/incorrect plans and flags trade-offs
    3. Note-taking: aggressive documentation + memory management
    4. Reversibility: confirm before hard-to-reverse actions; “interrupting is cheaper than destroying”
  • Project “execution plan” structured as 5 hours (hour-by-hour):
    • Hour 1: set up the data layer foundation (“data lake” analogy) + memory system
    • Hours 2–4: build dashboards and skills on top of the data lake
    • Hour 5: polish (notifications + end-to-end testing)
  • Skills suite (examples):
    • Today: fetch/update the morning brief
    • Research: deep dive on a selected stock/ticker
    • Prep: meeting prep with context and people details

Pros (as described)

  • High correctness / reduced surprises via PRD-first + sign-off (prevents building the wrong thing)
  • Better control: Claude pushes back, asks clarification questions, and highlights trade-offs
  • Useful daily automation: centralized daily digest, investment tracking, and meeting prep
  • Autonomous builds producing tangible artifacts overnight (e.g., resource hub)
  • Dashboards for monitoring/visualization to track what’s happening
  • Hostinger benefit (if used): avoids local-agent downtime—runs on a private VPS 24/7, so progress isn’t lost when a laptop sleeps

Cons / limitations mentioned

  • Not “completely smooth sailing”:
    • you must grant permissions and make tweaks/additions to the PRD as you go
  • Memory issues over time:
    • Co-work starts “forgetting things” as time passes
    • the memory system can bloat (the video mentions implementing a more complex memory system)
  • Local agent issue (context provided):
    • if running locally (e.g., Claude Code Remote) and the laptop sleeps, the agent “dies” and progress can be lost—motivating the VPS approach

Comparisons made

  • Compared to other local AI agents: mentions the same “agent dies when laptop sleeps” issue with OpenClaude, Hermes, and Cohort
  • Memory comparison: claims Co-work’s forgetting is not as bad as Openclaw (but still occurs)
  • VPS vs local: Hostinger VPS is positioned as the fix for persistence (“never sleeps”)

User experience / setup flow (how it feels)

  • Setup steps include:
    1. Download Claude Co-work
    2. Paste an operating instructions/meta prompt into Claude Co-work settings (prompt provided in the video/description)
    3. Generate a custom PRD via a linked/copy-paste prompt
    4. Create required folder/project scaffolding (claimed to take ~1 minute)
    5. Kick off the build using the PRD (follow the hour-by-hour plan)
    6. Optional: run the autonomous builder for ongoing improvements
  • Time to complete initial setup/build: ~5 hours (according to the creator)

Numerical / rating info

  • Time estimate to build the system: ~5 hours
  • Scaffolding setup time: ~1 minute
  • Autonomous builder polling frequency: every 30 minutes
  • The five-hour plan is explicitly described hour-by-hour
  • No star ratings or numeric product scores were provided

Overall verdict / recommendation

Recommended if you want an organized, PRD-controlled Claude automation setup that can deliver daily digests and continuously generate/extend workflows. Expect some permissions/tuning and eventual memory bloat, but the workflow design (PRD-first, pushback, reversibility checks) is positioned as a major advantage. If you’re running locally, strongly consider the 24/7 VPS approach to avoid interruptions.

Unique points mentioned (complete list)

  • Daily Apple Notes digest with investments, calendar, emails, action items
  • Custom investments dashboard with news/advice aligned to investing philosophy
  • Autonomous nightly software/workflow building
  • Mission control dashboard for monitoring builds/projects
  • Resource hub built by autonomous builder overnight (example)
  • Download Claude Co-work; use Claude hub/settings → paste operating instructions
  • Meta prompt includes 4 behavioral sections: Building things, Pushback & clarification, Note-taking, Reversibility
  • PRD requirement: problem, success criteria, scope, constraints, build plans
  • PRD includes open questions + explicit sign-off before building
  • Claude instructed to push back on inconsistent/incorrect plans + flag trade-offs
  • Aggressive note-taking + emphasis on memory management
  • Reversibility confirmations before hard-to-reverse actions; stop and ask if in doubt
  • Hour-by-hour 5-hour build plan (data lake foundation → dashboards → briefs/skills → polish)
  • “Data lake” analogy for data layer and transformations into dashboards/projects
  • Folder architecture: local agent runs under a Co-work folder; PRD maps folder structure; mentions a “three-layer system”
  • Folder scaffolding:
    • create project folder
    • create new project pointing to existing folder
    • drop in PRD file
    • start building command
  • Need to follow PRD and permissions; expect tweaks during build
  • Skills examples: Today, Research ticker, Prep for meetings
  • Autonomous builder mechanics:
    • Claude suggests → drafts PRD → user approval → pending folder → scheduled checks → in progress → done/failed + logs
  • Mission control dashboard used to see builds more clearly
  • Optional/“advanced bonus” general builder section can be skipped
  • Notifications setup + end-to-end testing in hour 5 (not shown)
  • Ongoing expansion: health tracking, planning skills, weekly plans, memory improvements
  • Memory system eventually forgets things; not as bad as Openclaw, but still bloat occurs
  • Mentions a possible Part 3 about the more complex memory system

Speaker views

  • Single main speaker/creator throughout; no separate speakers contribute distinct viewpoints.

Original video