Set up the workspace once. Stop re-explaining your company.

What the agent sees

The agent starts each session knowing only what's in your files, plus a short Project memory if you use Cowork. It looks at those files and works out what to do from what it sees.

A folder with your ICP, your positioning, three posts you're proud of, and a list of words you never use produces drafts that sound like your company. A folder called Downloads produces drafts that sound like everyone else's.

So spend your setup time on the workspace. A better prompt helps one session. The folder gets read at the start of every session after it.

The five-layer stack

Each layer depends on the one below it. Layers 1 and 2 do most of the work on draft quality. Layers 3 to 5 are where you stop re-explaining and re-checking.

Start at layer 1 and build up.

Pick your track

Use the toggle at the top to filter the whole guide. You can switch any time, and the folder structure is the same on both.

Most students mix them: Cowork for daily drafting and research, a terminal tool for anything that touches data, automation, or a shared repo.

The starter folder structure

This tree is built for marketing work. context/ matters most. It holds everything the agent reads before it drafts for you.

my-marketing-workspace/
├── AGENTS.md          # the rules (every tool reads this)
├── CLAUDE.md          # one line: @AGENTS.md
├── README.md          # what this workspace is, in one paragraph
├── .gitignore         # what git never tracks: secrets, exports
│
├── context/           # read this before drafting anything
│   ├── company.md     # product, ICP, positioning, proof points
│   ├── personas/      # one file per buyer
│   ├── voice/         # voice guide, banned words, 5 best examples
│   └── brand/         # colors, fonts, logo rules
├── inbox/             # raw material lands here first
├── campaigns/         # one folder per campaign or project
│   └── 2026-q4-webinar/
│       ├── README.md  # goal, audience, owner, status
│       ├── inputs/
│       └── drafts/
├── output/            # shipped work: content, emails, decks
├── reference/         # research and source material you reuse
├── templates/         # starting points that aren't skills yet
├── archive/           # finished campaigns
└── .agents/skills/    # your saved playbooks (Part 6)

Why it's shaped this way

• context/ is the part that changes the output. Your ICP, voice, and positioning live in one place, so every draft starts from them.
• inbox/ keeps the top level clean. New files land there until they have a home.
• campaigns/ holds work in motion. Each one gets a short README with goal, owner, and status, so you and the agent can see where it stands.
• output/ keeps what the agent made apart from what it read. That one split ends most of the "where did it save that" hunting.

Naming that scales

• Lower-case and hyphenated: q4-pipeline-review, not Q4 Pipeline Review (final)(2).
• Dates first, as YYYY-MM-DD, so they sort: 2026-10-31-board-update.md.
• A prefix when it helps grouping: webinar-, abm-, acme-.

The rules file

This file tells the agent who you are, where files go, and what it must never do. Each tool looks for a different file name, so write the rules once in AGENTS.md and point everything else at it.

What goes in it

• What this workspace is. One paragraph: your company, your role, what you use it for.
• Read before drafting. Point at context/company.md, your personas, and your voice guide.
• Where things go. A table mapping kinds of work to folders. The agent checks it every time it saves a file.
• How you like things done. "Show me a plan before writing files." "Date every file." "Lead with the hook."
• What never happens. "Never send, post, or publish anything." "Never update the CRM without asking." "Never invent a statistic."

Keep it under about 200 lines and link out to detail. Past 200 lines, agents start skipping rules and you stop updating the file.

Two habits that make it better every week

• Keep the file short and specific. A study by Gloaguen and colleagues at ETH Zurich (ICLR 2026 workshop) tested context files on coding agents. Whether an AI model or a developer wrote them, the files didn't raise success rates on average and raised costs by over 20%. Agents did follow the instructions, but general overviews didn't help. So list the rules that change the work and skip the background. Let the agent draft a first version if you want, then cut everything that isn't true and specific.
• When the agent gets something wrong, add a rule so it doesn't happen again.

Skills: your saved playbooks

The rules file says how you work in general. A skill says how to do one job. It's a folder with a SKILL.md file: a name, a one-line description of when to use it, and the steps. Claude Code, Cowork, and Codex all support them.

The agent reads only the description until a request matches it, then loads the full playbook. So the description is the trigger. Write it in the words you type: "Use when I ask for a LinkedIn post or say 'turn this into a post.'"

Good first skills for marketers

When to make one: the third time you explain the same job to the agent, turn the explanation into a skill.

Connectors: let it read your stack

Connectors (built on MCP, the Model Context Protocol) give the agent access to tools outside the folder: your CRM, docs, call notes, analytics. They're the difference between "write an ABM email to a target account" and "write an ABM email to the five accounts that visited pricing this week."

Three rules

• Start every connector read-only. Add write access (sending email, updating records) only after weeks of watching it work, and keep "ask before any write" in your rules file.
• Don't trust the green dot. "Connected" only means the tool can reach the server. Run the first real test above and check the answer against the source.
• Connect only what you use weekly, since each one adds tools the agent has to choose between.

GitHub and teams

GitHub keeps every version of your workspace and lets a teammate work in it with you. You'll use a small slice of it: commit, push, and a private repo.

cd my-marketing-workspace
git init
git add .
git commit -m "Set up workspace structure"
gh repo create my-marketing-workspace --private \
  --source=. --push

Commit habits

• Small and often. One change per commit.
• Start with a verb: "add," "fix," "rewrite." Keep the first line under about 50 characters.
• Make repos private by default. Your context/ folder is your positioning.

Working with a team

Solo, you can skip branches. Once two people edit the same workspace, use them: each person works on a branch, opens a pull request, and a teammate approves before it merges. On GitHub's free plan a private repo can't require that approval, so it stays a team habit. GitHub Pro or Team can enforce it with a branch protection rule on main. Give each folder one owner, and treat a change to AGENTS.md like a policy change, since it changes how every teammate's agent behaves.

The quality check

Agents write fluent text fast. Generic phrasing and made-up numbers read as smoothly as the good parts, so you can't catch them by skimming. Before anything ships, run three checks.

1. Check the draft against context/voice/: banned words, sentence patterns you never use, how your best examples open.
2. Find a source you can open for every number, customer name, and claim. If the agent can't point to one, cut it.
3. Review in a fresh session. A session grading its own draft goes easy on it.

You are reviewing a draft you did not write. Read AGENTS.md and
everything in context/voice/ first.

Then review the draft below and return:
1. Every sentence that breaks a voice rule, with the rule it breaks
2. Every number, name, or claim, and whether a source exists in
   this workspace (cite the file) or not
3. The three weakest lines, and a rewrite for each
4. A verdict: SHIP, REVISE, or REWRITE

Do not rewrite the whole draft. Do not soften the verdict.

DRAFT:
<paste here>

Save that prompt as a skill once you've used it three times. If the draft is already a file and you have the draft-review skill, use Fresh-session draft review from the prompt library instead.

Safety

• Keep API keys and passwords out of the workspace. Put them in environment variables or a password manager. The .gitignore blocks the common ones.
• Don't drop full CRM exports with emails and phone numbers into the folder. Pull what a task needs through a connector, or strip personal details first.
• Keep your tool on the setting that asks before deleting files, sending, or writing to another system. Cowork asks before high-risk actions. In Claude Code, don't use --dangerously-skip-permissions or bypass permissions mode. In Codex, don't use --dangerously-bypass-approvals-and-sandbox. Keep the approvals on until you trust a workflow.
• A web page, PDF, or inbound email the agent reads can contain hidden text telling it what to do. Treat anything it reads as material to quote, never as instructions to follow, and never let an agent with send access process untrusted content unsupervised.

Per-tool quickstart

Keeping it healthy

Workspaces decay because files pile up faster than anyone sorts them. If it won't exist next session, write it down now. A decision, a preference, a mistake you corrected in chat. If it only lives in a chat window, the agent loses it.

Copy-paste starter kit

Copy each file into your workspace and fill in the blanks. Or use the workspace builder to get these filled in for your role.

# AGENTS.md
# The rules for this workspace. Codex reads this file directly.
# Claude Code reads it through CLAUDE.md. Paste it into your Cowork
# Project instructions. Keep it under ~200 lines. Edit by hand.

## What this workspace is
<One paragraph: company, your role, what you use this for.>

## Read before drafting anything
- context/company.md: product, ICP, positioning, proof points
- context/personas/: who we sell to
- context/voice/: how we sound and words we never use

## Where things go
| Type of thing               | Goes in          |
|-----------------------------|------------------|
| New, unsorted material      | inbox/           |
| Campaigns and projects      | campaigns/<name>/ |
| Shipped work                | output/          |
| Research and sources        | reference/       |
| Starting points             | templates/       |
| Finished campaigns          | archive/         |

## How I like things done
- Show me a short plan before creating or changing files.
- Date files as YYYY-MM-DD at the front of the name.
- <add your own>

## Never
- Never send, post, publish, or schedule anything. Draft only.
- Never write to the CRM or any connected tool without asking.
- Never invent a statistic, quote, or customer name. Cite a source file.
- Never store secrets or raw customer data in this folder.
- Never delete anything in archive/.

## Before anything ships
Check it against context/voice/, confirm every number has a source,
and have a fresh session review it.

# CLAUDE.md
# All rules live in AGENTS.md so Claude Code and Codex read one copy.
# Add Claude-only notes below the import if you ever need them.

@AGENTS.md

---
name: linkedin-post
description: Use when I ask for a LinkedIn post, or say "turn this
  into a post" or "draft a post about X". Drafts in my voice from
  context/voice/. Do not use for comments or ads.
---

# LinkedIn post

1. Read context/voice/ and context/personas/ before writing.
2. Ask me for the one point of the post if I haven't given it.
3. Write three hook options, each under 12 words. Let me pick.
4. Write the post: short paragraphs, one idea, one specific example.
5. Check the draft against the banned-words list. Fix every hit.
6. Save to output/content/YYYY-MM-DD-linkedin-<topic>.md
7. Never post it. Draft only.

# Secrets and credentials: never commit these
.env
.env.*
*.key
*.pem
secrets/
credentials.json
*-credentials*.json

# Raw data exports that may hold customer details
exports/
*.csv

# System and editor files
.DS_Store
Thumbs.db
.vscode/
.idea/

# Large files
*.zip
*.mov
*.mp4
node_modules/

Prompt library

Copy any of these into a session started from your workspace folder. They work the same in Cowork, Claude Code and Codex. Text in <ANGLE BRACKETS> is yours to fill in. Start with the three prompts in START-HERE.md. These come after.

Setup checklist

Check items off as you go. Progress saves in this browser and follows the track you picked at the top.