Brownstone Research Coworker Starter Kit
Getting Started

The Coworker Starter Kit

The point of this starter kit is to help you stop prompting and start delegating.

Step 1

Download and set up the files

1

Download the kit.

Download the Kit
2

Open your vault and look at the top level.

Obsidian hides file extensions, so the kit's instruction file just reads as CLAUDE — not CLAUDE.md.

Hard stop

Do you already have a file named CLAUDE at the top level?

Nothing to move. Go ahead to step three and unzip.

Drag that CLAUDE file out of the vault first, so the kit's version doesn't collide with it. Then continue to step three.
3

Unzip what you downloaded — outside the vault.

You should see CLAUDE, Jeff Brown, Larry Benedict, and folders like .claude, Copy, Prompts, and Handoffs sitting side by side — not all of that nested one level down inside a single outer folder. Those folders are supposed to be there; then drag everything you see into the vault's top level.
Step 2

What claude.md is

Claude.md is your permanent instruction file. This is where your orchestration system lives.

This file holds your preferences, your rules, your protocols, permissions, and anything you want Claude to always remember.

At first blush you think — hell, I'll put everything here — resist the urge. 

Why it stays lean

It's read before every session so you don't want to wind up with bloat racking up usage. Instead use this file to 'point' to other scenario-specific files. 

Your Coworker Starter Kit comes with a starter claude.md file, pre-loaded with skills, agents, and an orchestration layer.

ALL of it works in Claude Code. SOME of it works in Cowork.

Your claude.md file is your ultimate source of truth. In Claude Code it's read every session automatically and can be updated inside any session. 

If you want a rule add to your permanent claude.md file, ask for it inside Claude Code. 

To make it work in Cowork, you need to paste the contents of your claude.md file into the Cowork Global Instructions.

WARNING: If you update your claude.md file, always updated your Cowork Global Instructions and vice versa - otherwise they'll drift. 

Don't worry - it's simple, the instructions are on the next page, and I've included a triggered reminder in your new claude.md file that catches when you're trying to make something permanent, knows where you are, and tells you where to go and what to do (see rule 7 below).

BOTTOM LINE: COWORK MEMORY IS NOT AS RELIABLE AS CLAUDE.MD

Cowork has its own persistent memory, but it decides what to keep, decides what overrides what, and doesn't enforce hard rules — that's why anything permanent has to live in your claude.md file and copied to your Cowork Global Instructions directly.

What's in the kit

Your starter claude.md

Your starter claude.md has everything you need, for whichever environment you're in. It's strongly recommended to set this up in Claude Code the first time then copy it over to your Cowork Global Instructions. 

  1. 01Who you are and what you work on — This contains a prompt that triggers in your first fresh Claude Code session. Once you answer, it saves it, you never have to answer again.
  2. 02How you want to be talked to Yours to edit — I prefer plain language, direct answers, nothing padded. That's what yours comes with, but you're free to edit this.
  3. 03Formatting rules for anything you copy elsewhere Yours to edit — plain text, no stray formatting to clean up by hand.  Again, my preference. 
  4. 04Voice and compliance rules for Jeff and Larry — keeps each persona consistent and keeps the two from ever blending together. Recognizes trigger words and points to the respective persona files in your vault. 
  5. 05A rule for running commands safely — This makes it easier to work in Claude Code. Permissions are asked in plain-English heads-up before anything runs, and outputs results in plain-English after. 
  6. 06A rule for saving reusable prompts to the Prompts folder — This helps you build your prompt repository without bloating your claude.md file, drifting in your persistent memory, or constantly having to go back and hunt down where you saved a good prompt. When you get a great output, ask Claude to save the prompt. It will ask you for a trigger phrase, then save your prompt to the prompt folder included in your kit.
  7. 07A rule for memory versus file — when you tell Claude to remember something, it asks whether that belongs in claude.md or a one-time note, identifies where you are, then instructs you what to do next.
  8. 08A five-agent worker roster — five agents, each with one job, so you delegate instead of doing every step yourself in one thread. Each agent is scoped to the least expensive model so you're not asking Opus to do something Haiku can do. I prefer to default to Sonnet and then ask Sonnet if something I'm working on might need a higher model - it will honestly say whether or not you do.

The rest of the kit, one by one

Gopher

File conversion, formatting, filing, mechanical lookups, log parsing.

Claude Code

Invoked by your orchestration rules by identifying the scope of request then delegating it to the agent automatically.

Cowork

Not available in Cowork. Cowork has no mechanism to run any of these as agents.

Copy-drafter

First drafts and variants of promotional copy.

Claude Code

Routes on its own — no naming needed.

Cowork

Not available in Cowork. Cowork has no mechanism to run any of these as agents.

Data-puller

Pulls raw numbers from a connected source, no interpretation.

Claude Code

Routes on its own — no naming needed.

Cowork

Not available in Cowork. Cowork has no mechanism to run any of these as agents.

Copy-chief

Adversarial review of a copy draft before it ships as final.

Claude Code

Must be named — for fresh-session isolation, and because when it runs is a deliberate call.

Cowork

Not available in Cowork. Cowork has no mechanism to run any of these as agents.

Verifier

Cold-checks a "done" claim against evidence before you call something finished.

Claude Code

Must be named — for fresh-session isolation, and because when it runs is a deliberate call.

Cowork

Not available in Cowork. Cowork has no mechanism to run any of these as agents.

Saving reusable prompts

Say "save this prompt," give it a trigger phrase, it writes to the Prompts folder; saying that phrase again runs it. Works the same in both environments.

I like to say "Let's" in front of everything. Let's write an email. Let's update the split-test channel in slack. Let's call a test. That trigger points to the longer more specific prompt so you're not having to repeat yourself every time.

Claude Code

An explicit file-write instruction — writes to the Prompts folder, then runs on the trigger phrase.

Cowork

Same — it's an explicit file-write, not a vague "remember this," so it doesn't fall into Cowork's own memory the way a loose "remember this" would.

Subject-line-scorer

Scores or generates subject-line and preheader pairs.

Claude Code

Same trigger phrases, same behavior.

Cowork

Same trigger phrases, same behavior.

Lead-scorer

Scores or generates VSL and sales-letter opens.

Claude Code

Same trigger phrases, same behavior.

Cowork

Same trigger phrases, same behavior.
Step 3

Set up and use each interface

Same vault, same claude.md — what changes is setup, and what's actually reachable once you're in.

Pick the interface you're setting up. The two paths differ — choose the one you'll actually work in.

Claude Code

terminal, desktop app, IDE extension

Setup

Click the + sign to open a new session. Just above the entry field you should see your local hard drive icon followed by your user folder — that's your default root. Click on your user folder and select "open folder," then select your vault. It should spawn a little main worktree.

Then just start talking to Claude — say what you're working on, ask it anything. That first prompt doubles as onboarding: claude.md's Role section is still a placeholder at this point, so Claude asks who you are and what you work on, then rewrites that section itself with your answer. You never edit it by hand.

Once that's done, it opens a section in the sidebar with the name of that folder — from then on, open new sessions from that sidebar's + sign and claude.md auto-loads every time.

The web app version of Claude Code works the same way but needs a git repo instead of a local folder to mount — ask if you want help setting that up.

What's accessible

Everything

claude.md, both persona files, the prompt-saver skill, subject-line-scorer, lead-scorer, and the full 5-agent roster.

How to invoke it

  • Persona voice — ask for something "in Jeff's voice" or "as Larry." Automatic, no setup needed.
  • Subject-line-scorer / lead-scorer — just describe the task ("score these subject lines," "give me 5 leads for this promo"). Claude recognizes it and loads the right skill itself.
  • Saving a prompt — say "save this prompt." Claude asks for a trigger phrase and files it in your Prompts folder. Say that trigger phrase again later and it runs automatically.
  • The 5-agent roster — name the agent directly ("use the gopher agent to...") or lead with your saved trigger word ("let's use copy-drafter for this"). This does NOT happen on its own — a plain request with no name or "let's" attached runs in the main thread, even for work an agent's own description says to grab proactively. A bare mention of an agent's name in passing ("I don't think gopher should touch this") doesn't count either — it has to read as an actual ask.
  • Memory vs. file — when you tell Claude to remember or always do something, it asks whether to write it into claude.md or leave it as a one-time note. Answer directly.

Cowork

Setup

Go to Settings, click Cowork, click Global Instructions, and paste in the entire contents of your claude.md file — not a pointer to it. Cowork can't mount a live file the way Claude Code does, so a pointer like "read the claude.md file at the root of this project" has nothing to read; Global Instructions has to hold the actual text.

Then create a project and mount it to your vault.

Keeping it in sync

claude.md is still the one file you edit. Whenever you (or Claude, in Claude Code) add a permanent rule to claude.md, copy just the new text and paste it onto the end of Cowork's Global Instructions, then save. Global Instructions is a static copy you manually refresh — never treat it as a second place to make rules, or the two will drift apart and disagree with each other.

What's accessible

claude.md's content (pasted in), both persona files if you've also pasted or attached them, the prompt-saver skill, subject-line-scorer, lead-scorer.

Not available in Cowork

NOT the 5-agent roster — Cowork has no mechanism to spawn a separate agent, full stop, no workaround.

How to invoke it

  • Persona voice, subject-line-scorer, lead-scorer, saving a prompt — work the same way as Claude Code above (same trigger phrases, same claude.md-driven behavior), as long as the relevant files are pasted/attached into the project.
  • Memory vs. file — when you tell Claude to remember or always do something, it still asks whether that's permanent or one-time. If permanent, it can't write claude.md itself here — it tells you to make the edit in Claude Code, then paste the new text into Global Instructions yourself.
  • The 5-agent roster — since there's no agent to hand off to, you carry that same division of labor yourself in one continuous thread: draft, then deliberately switch hats and review your own draft the way copy-chief would, verify your own "done" claims cold before calling something finished, and so on. claude.md's rules still apply — you're just the one applying them instead of a separate agent.
{{ stepLabel }}