Stop Prompting Claude. Architect It.

On their first day, you wouldn't hand a new hire a task and walk away. You'd brief them. Here's who we are. Here's how we work. Here's what good looks like.

Claude Cowork is no different. Standalone sessions start with zero memory. Every time you open one, you're talking to a stranger. The fix is a context architecture. A set of files that brief Claude automatically, so every session starts with a colleague who already knows your role, your voice, and your rules.

This takes about two hours to build. After that, prompts like this actually work:

Write a proposal for Acme Corp. Read all files first.

Ten words. Claude reads your context files, follows your voice, uses your template, and produces a 4-page proposal with correct pricing. Your tone. Your formatting. No 80-word setup prompt. No re-explaining yourself.

Here's how to build it. Five steps.

Step 1: Set up the office

You need two things before you start: Claude Desktop (the app, not claude.ai in a browser) and a paid plan. Pro is $20/month, Max is $100/month. Cowork is the third tab in Claude Desktop, next to Chat and Code.

First, create a folder on your computer. Your Desktop or Documents folder works fine. Call it whatever fits: "My Cowork", your company name, "HQ". Inside it, create three subfolders: About Me/ for your personal context files (Step 2), Outputs/ for work Claude produces, and Templates/ for reusable formats.

Now connect it. Open Claude Desktop, click the Cowork tab, then click the folder icon to choose which folder Cowork can access. Select the folder you just created. Claude will read everything in it at the start of each session.

This folder structure comes from Ruben Hassid, who's been stress-testing Cowork since launch day. It mirrors how you'd organize a shared drive for an actual employee. Cowork needs the same clarity.

If you work across multiple areas (say, client work and internal strategy), create a folder per area with the same three subfolders inside each one. You choose which folder to connect at the start of each session, so separate folders keep context clean.

Don't overthink this. You can reorganize later. The point is to start with something.

Step 2: Write the onboarding brief

This is the core of your setup. Three markdown files that tell Claude who you are, how you communicate, and what you expect. Drop all three into your About Me/ folder.

You can write these yourself, or let Claude do it. Open Cowork and say: "Interview me about my role, writing style, and work preferences. Then create about-me.md, my-voice.md, and my-rules.md in my About Me folder." Claude will ask you questions and write the files for you. Either way works. Here's what each file should contain.

about-me.md is your professional identity. Role, company, what you work on, who your audience is, what your priorities are. Keep it under 2,000 tokens (roughly 1,500 words). Shorter is better. Claude reads this every session, so don't pack in your life story.

Here's the shape of a good one:

# About Me
 
## Role
Head of Marketing at Acme Corp (B2B SaaS, 200 employees)
 
## What I Do
- Own content strategy, demand gen, and brand
- Report to the CMO
- Team of 6
 
## My Audience
Mid-market CTOs and VP Engineering at companies with 50-500 employees
 
## Priorities This Quarter
- Launch product-led growth motion
- Reduce CAC by 20%
- Ship 3 case studies

my-voice.md captures how you sound. This one matters more than people expect. Include your preferred tone (direct? warm? technical?) and paste in two or three samples of your real writing. Proposals you've sent. Emails you're proud of. Slack messages that sound like you. Claude picks up patterns fast when it has examples.

my-rules.md is your guardrails file. Every "never do X" and "always do Y" goes here. Hate bullet points in client emails? Say so. Want every proposal to include a pricing table? Put it here. Prefer British English? This is where it lives.

These three files are your onboarding brief. A new hire who read them would know how to write like you, what to prioritize, and what mistakes to avoid. Claude works the same way.

Step 3: File the cabinet

Once you have more than a handful of files, Claude needs a reading list. That's what _MANIFEST.md does.

You can create this manually, or tell Claude: "Look at all the files in my folder and create a _MANIFEST.md that tiers them into canonical, domain, and archival." Claude will scan your folder and organize the manifest for you.

If you want to do it yourself, create a file called _MANIFEST.md in your root folder. Inside, tier your files into three categories:

# _MANIFEST.md
 
## Canonical (read every session)
- About Me/about-me.md
- About Me/my-voice.md
- About Me/my-rules.md
 
## Domain (read when relevant)
- Templates/proposal-template.md
- Templates/weekly-report-format.md
- About Me/my-company.md
 
## Archival (read only when asked)
- Outputs/q1-strategy-doc.md
- Outputs/client-research-acme.md

Canonical files load every session, no matter what. These are your three onboarding files from Step 2. Domain files get read when the task matches. Archival files sit there until you specifically ask Claude to reference them.

Without a manifest, Claude reads everything in the folder every time. That's fine with 5 files. With 30, it wastes context window on old output drafts when you just want a quick email. The manifest tells Claude what matters most.

A note: this isn't a built-in Cowork feature. It's a community convention that works because Claude reads markdown files and follows instructions. You'll tell Claude to check for it in the next step.

Step 4: Post the standing orders

Your context files handle the "who" and "how." Global instructions handle the "what to do first."

In Claude Desktop, go to Settings, then Cowork, then click Edit next to "Global instructions." This text box applies to every Cowork session. Paste something like this:

At the start of every session:
1. Read _MANIFEST.md first
2. Load all canonical files
3. Check domain files for relevance to the current task
4. Before executing any task, ask me 2-3 clarifying questions
5. Follow my-voice.md for all written output
6. Follow my-rules.md as hard constraints

If you're on Max, you already have access to Claude's strongest model. It references your context files more carefully and produces better first drafts than the default.

One more thing. Add this line to your global instructions: "End every session by writing session-notes.md with key decisions and next steps." For standalone sessions, session notes become the bridge between conversations. Start your next session with "Read session-notes.md" and you pick up where you left off. (If you use Cowork's Projects feature, memory persists within a project automatically. Session notes are still useful for keeping a paper trail.)

Step 5: Teach it your first workflow

You've built the architecture. Now put it to work.

Skills are custom commands that turn repeated tasks into one-word triggers. Pick the task you do most often. Writing a weekly report? Drafting client emails? Summarizing meeting notes? That's your first skill.

The fastest way: just ask Claude. Open Cowork and say: "Help me build a skill for writing my weekly status report. Interview me about how I want it to work, then create the Skill.md file." Claude will ask what the report should include, what format you prefer, and how much autonomy to take. Then it generates the file for you.

Here's what a good skill looks like. Notice the mix of autonomous action and human checkpoints:

---
name: weekly-report
description: Generate my weekly status report
---
 
# Weekly Report Skill
 
When activated:
 
1. Read About Me/about-me.md for my role and priorities
2. Review my recent Cowork sessions and files in Outputs/ to identify what I shipped this week
3. Draft the report based on what you found
4. Ask me: "Anything I missed? Any blockers to add?"
5. Finalize following Templates/weekly-report-format.md
6. Keep it under 500 words, direct, no fluff
7. Save to Outputs/weekly-report-{date}.md

The key difference from a basic prompt: the skill tells Claude to look through your recent work FIRST, then fill in the gaps with questions. It does the heavy lifting. You just confirm and add what it missed.

To install: create a folder called weekly-report with this Skill.md file inside, ZIP it, then upload in Claude Desktop via Customize > Skills. Once installed, type / in any Cowork session to see your skills and pick the one you want.

Build one skill this week. Add another next week. Within a month, your most common tasks are all one-word triggers. Claude handles most of the work before you even type.

Do it now. Create the folder. Write your three context files. Add the manifest. Set your global instructions. Build one skill. The whole thing takes about two hours. After that, every session with Claude starts briefed, primed, and ready to produce real work. The people still typing paragraph-long prompts every morning are already behind.

Related alpha: The 10-Minute AI Setup Most People Skip covers Custom Instructions for ChatGPT, Claude, and Gemini. If you use more than one tool, start there first.