TL;DR: Claude Cowork starts blank every session. 30 minutes of one-time setup (context files, global instructions, scheduled tasks) turns it from a generic assistant into a context-aware tool that already knows who you are and what you are working on. The 7 configurations below are what made the difference.
I have been using Claude Cowork since it launched in January 2026. It runs my morning briefs, synthesizes research across Slack and email, and produces weekly rollups before I ask for them. But it took me a few weeks to get there. Out of the box, Cowork does not know who you are. Every session starts blank. So it gives you generic, safe, middle-of-the-road outputs.
The fix was setting up the environment properly. About 30 minutes of one-time configuration, and then small tweaks over time.
Here is what I did.
1. Put a _Context.md in Every Working Folder
The underscore prefix pushes it to the top of the file list so Cowork picks it up first. Inside, I list which files in the folder are the source of truth and which ones to skip. This matters because Cowork will try to read everything in the folder and incorporate all of it. If you have 40 files but only 5 are relevant, you end up with bloated outputs that mix current strategy docs with old brainstorm notes. Keep this updated using Claude Cowork.
My _Context.md for a typical project folder looks like this:
# Context
## Read these
- product-roadmap-h1.md
- project-brief.md
## Skip these
- old-drafts/
- archive/
- meeting-notes-jan.md (outdated)
Short, explicit, and it keeps Cowork focused.
2. Add a Personal Context File
Separate from the project context, I have a file that tells Cowork who I am. My name, role, team, what I own, how I like outputs structured. Without this, you get the same output anyone else would get for the same prompt.
Mine includes things like: my role and what I own, that I prefer tables over paragraphs when comparing options, keep language simple, no filler phrases, and always cite source filenames when synthesizing research so I can trace where a claim came from.
You write this once. It takes maybe 15 minutes. Every session after that starts with Cowork already understanding your context.
3. Require a Plan Before Execution
This is a simple rule I added to my instructions: "Always create and share the plan before execution."
Cowork is autonomous. If you give it a task, it will interpret your request, make assumptions, and start working. Sometimes those assumptions are wrong, and you have burned 10-15 minutes on output you cannot use. A quick plan review takes 30 seconds and catches misalignment early.
The plan usually lists which files Cowork intends to read, what it thinks the deliverable should look like, and how it plans to structure the work. I approve, adjust, or redirect before it spends any real effort.
4. Set Global Instructions
These apply to every Cowork session regardless of folder. Settings > Cowork > Global Instructions. Set once, never think about it again.
Mine:
My name is Prateek. I am Head of Product at Grab.
Check \_Context.md first before starting any task.
Always share a plan before execution.
Be concise. Use simple English.
No em dashes. No corporate filler.
Use tables when comparing options.
Cite filenames when synthesizing across documents.
5. Layer Folder Instructions on Top
Global instructions handle who you are and how you work. Folder instructions handle what you are working on right now.
Lets say when you are working on a market expansion project, the folder instructions might say: target market is Cambodia, payment architecture is X, key stakeholders are Y, deadline is H1. When you switch to a different workstream, different folder, different instructions. The project context stays contained instead of leaking across workstreams.
This layering is what makes the setup work across multiple projects without constant re-prompting.
6. Use Parallel Sub-agents for Independent Work
Cowork can spin up multiple sub-agents to work on separate pieces at the same time. You have to ask for it explicitly. Something like: "Spin up parallel sub-agents to research each of these five competitors separately."
The key constraint is that the pieces need to be genuinely independent. Sub-agents run in isolated context windows. They do not share state. So "research five competitors in parallel" works well. "Write a document where section 3 builds on section 2" does not.
Do keep in mind sub-agent workflows use significantly more tokens than single-agent sessions. On a Pro plan, you can hit rate limits fast.
One practical habit: always ask for a summary file at the end. A what-changed.md or similar that lists what each sub-agent did. Makes review much faster than digging through individual outputs.
7. Automate Recurring Work with /schedule
Type /schedule in any Cowork task to set up jobs that run automatically on a daily or weekly cadence. This is where connectors make a real difference. I have Gmail, Google Calendar, Slack, and Google Drive connected, along with some internal work tools.
Here is what I have running right now:
- Morning Brief (daily, 8:30 AM) - Pulls unread emails, today's calendar, and overnight Slack activity. Produces a single-page summary with flagged action items.
- End of Day Brief (daily, 6 PM) - Summarizes what happened across email, Slack, and calendar. Highlights unresolved items. Drafts follow-ups where needed.
- End of Week Brief (Friday, 5 PM) - Synthesizes the full week. Decisions made, open items, suggested priorities for Monday.
- Daily AI News Brief (daily, 8:30 AM) - Searches the web for AI and product news, filters for relevance, delivers a short digest.
The caveat with /schedule: tasks only run while your computer is awake and Claude Desktop is open.
The 30 minutes you put into writing context files and setting instructions pays back in every session that follows.