Give Claude persistent memory, your voice, and custom workflows — in one afternoon.
A quick reference for BRG business users across all teams
Every session begins from zero. Claude doesn't know your name, your role, your preferences, or what you were working on yesterday. You spend the first few minutes re-explaining context before any real work happens.
Generic tone. No memory. No shortcuts. Repeat yourself constantly.
Claude reads your constitution and memory files at the start of every session. It knows your role, your active projects, your preferences, and how you like to write. You jump straight into the work.
Your voice. Your context. Routes the right tasks to the right workstation automatically.
CoworkOS is just a folder on your computer. No external services, no lock-in.
You explain the job properly one time — the files, the rules, the format, the people involved. From then on you just ask for the thing, and it already knows. Like a colleague who has done your bank rec forty times and doesn't need the background again.
Most of your work is not a workstation, and that's the point. They're for the five jobs you've already done a dozen times and will do a hundred more.
Four things that sound similar and aren't. Nine out of ten mix up the first two.
Not hypothetical. An AR specialist in APAC built three reconciliation workstations that now run in live monthly use. A finance manager in Canada built five.
Each workstation is one specialist, in one folder: a local CLAUDE.md for its rules, a MEMORY.md for what it has learned, and a Resources subfolder for its reference files. The global files apply everywhere; the workstation files layer on top.
The global constitution. Tells Claude your preferences, rules, folder structure, and routing map. Every session starts by reading this.
The persistent brain. Claude reads it at session start and writes to it when you say "remember this." It builds up your context over time.
Your writing voice. Tone, sentence style, words to avoid, signature phrases. Claude reads this before writing any content on your behalf.
00_Resources/ folder. That's the whole system.
You're one afternoon away from a Claude that knows who you are, how you work, and what you're trying to get done.
Open Cowork with your CoworkOS folder selected, paste this prompt, and let it run. Claude will create all three core files, the folder structure, and then ask you a few questions to personalise everything.
I'm setting up CoworkOS for the first time. Please create the following in my CoworkOS root folder: 1. CLAUDE.md — with these sections: Memory System (include memory routing rules: root vs workstation MEMORY.md, and that behavioral rules go to CLAUDE.md not MEMORY.md), Preferences, Rules, Folder Structure, Routing Map 2. MEMORY.md — with these sections: About Me, Active Projects, Key Contacts, Key Decisions, Useful Context 3. 00_Resources/voice-principles.md — with these sections: Tone, Sentence Style, Words to Avoid, Words I Use, Formatting 4. 00_Skills/ (empty folder) 5. 00_Scheduled/ (empty folder) Use starter templates with [bracketed placeholders] where I need to customise. After creating the files, ask me 3 questions to personalise my CLAUDE.md for my role and working style.
That one prompt created your entire CoworkOS skeleton — three files, two empty folders, and a structure Claude will use on every future session.
The next three slides walk through each file with a copy-ready template. Customise the bits in [brackets] and you're good to go.
This file is Claude's operating instructions. It loads automatically and governs tone, behaviour, folder structure, and which workstation to use for which task.
# CLAUDE.md
## Memory System
Read MEMORY.md at the start of every session.
Write to MEMORY.md when I say "remember this."
+ new
Memory routing — which MEMORY.md?
- Cross-cutting (role, people, preferences, projects)
→ root MEMORY.md
- Workstation-specific (domain contacts, decisions,
output paths, tool configs) → [Workstation]/MEMORY.md
- Behavioral rules (always/never/before doing X)
→ CLAUDE.md (root or workstation), not MEMORY.md
Never duplicate the same fact across both files.
## Preferences
- Tone: [conversational / professional / casual]
- Language: [Australian / British / American English]
- Length: concise by default; expand only when I ask
- Format: prose first, bullets only when content is
genuinely list-like
## Rules
- Ask clarifying questions before complex tasks
- Never guess — if unsure, say so
- Read voice-principles.md before writing any content
on my behalf
## Folder Structure
| Folder | What lives here |
| 00_Resources/ | Reference files (voice, etc.) |
| 00_Skills/ | Installed skill archives |
| 00_Scheduled/ | Scheduled task outputs |
| [Workstation]/ | Workstation + Outputs/ folder |
## Routing Map
| Workstation | Route here when I... |
| [Workstation Name] | [Brief trigger description] |
Fill in the About Me section yourself — the rest Claude populates over time as you work together. Don't over-engineer this on day one; let it grow.
# MEMORY.md ## About Me - Name: [Your name] - Role: [Job title, team, company] - Key priorities: [Your top 2–3 focus areas] - Working style: [e.g. async-first, direct comms] ## Active Projects | Project | Status | Notes | | [Name] | Active | [One sentence] | ## Key Contacts | Name | Role | Why they matter | | [Name] | [Role] | [Context] | ## Key Decisions (Claude populates this over time as we work — don't edit manually) ## Useful Context (Claude adds facts here when you say "remember this")
Ask yourself: would this be useful in three different jobs, or only this one? Three jobs, it goes in the root notebook. Only this one, it goes in the workstation's own. That's the whole rule.
Things true no matter what you're working on — your role, the people you deal with, what you're currently on, how you like things done. Useful in three different jobs? It belongs here.
Things that only matter for one job — the contacts for it, the calls you've made, where the output goes, which entity or system it uses. If it's meaningless outside that job, it goes here.
Rules, not facts. Anything with an "always", "never" or "before you do X" in it. If you're telling Claude how to behave rather than what's true, it goes in CLAUDE.md instead.
When in doubt, write to root MEMORY.md. Never duplicate the same fact across both files. If a memory could plausibly live in either place, root wins — it's always in scope regardless of which workstation you're in.
Paste 5–8 of your own emails or Slack messages and ask Claude to extract patterns from them. It'll fill in the signature phrases and tone sections automatically.
# Voice Principles
## Tone
- Sound like: [smart peer / trusted advisor / warm colleague]
- More formal with: [senior leaders / new contacts]
- Avoid sounding: [corporate / over-hedged / AI-flavoured]
## Sentence Style
- Vary length — mix short punchy beats with longer ones
- Use contractions: [yes / no]
- Front-load context, then the ask
## Words to Avoid
- Never: "leverage," "synergise," "deep dive," "circle back"
- No em dashes (use commas or parentheses instead)
- No "I hope this email finds you well" openers
- No stacked adjectives ("incredibly valuable", "seamless")
## Words I Use
- [Your signature phrases — add them here]
- e.g. "heaps", "nerd out", "level up", "if helpful"
## Formatting
- Short paragraphs (1–2 sentences max)
- Bold sparingly — key terms and names only
- [Other formatting preferences]
Don't start with a workstation. Start with one scheduled task that saves you a job you resent — it takes two minutes and it proves the whole idea. Build the workstation once you already believe it.
A scheduled task is Cowork doing one small job on a timer, without being asked. No folders, no files, no setup. It is the fastest way to see the point.
Pick the check you do every morning out of habit — a balance, a queue, an inbox, a report that lands late. Tell Cowork to do it and message you the answer.
Set up a scheduled task for me. What I check manually right now: [describe it plainly — e.g. "every morning I open the bank file and compare actual cash to the forecast, then flag anything more than $5k out"] When it should run: [e.g. "every weekday at 7am Sydney time"] How I want the answer: [e.g. "a short Slack message to me — just the variance and anything unusual, not a full report"] Before you set it up: - Tell me what you'll need access to - Run it once now so I can see the output - Then confirm the schedule with me If anything is ambiguous, ask me rather than guessing.
This is the brief from slide 4, written down. Four short sections, in one file, so you never have to say any of it again.
One paragraph. Who Claude is in this workstation, what kinds of tasks route here, and — just as importantly — what doesn't belong here and should go elsewhere.
A table mapping reference files to "read when" conditions. Claude only loads these files when the trigger is met, keeping context efficient.
Numbered steps for the primary task in this workstation. Keep it simple — 5–7 steps. You'll refine it over time as you see what Claude gets right and wrong.
Always opens with "Follow voice-principles.md in 00_Resources." Then adds domain-specific rules on top — tone calibration, vocabulary, output formats for this specific domain.
Pick a job that passed all three questions on slide 4 — a reconciliation, a weekly report, a recurring email. Fill in the two brackets, paste, and Claude scaffolds the whole thing.
Create a new workstation called "[Workstation Name]".
Route here when I: [describe the task type — e.g.
"draft, edit, or review any Slack message"]
Create these three items inside the workstation folder:
1. [Name]/CLAUDE.md with sections:
- Identity (one paragraph: what routes here,
what doesn't)
- Resources (empty table: Resource | Read when...)
- Workflow (numbered steps for the main task)
- Editorial Rules (start with voice-principles.md
reference, then add domain-specific rules)
2. [Name]/MEMORY.md with:
- Contacts section (people relevant to this domain)
- Key Decisions section (domain-specific choices)
Note: cross-cutting facts (role, people, projects)
belong in root MEMORY.md, not here.
3. [Name]/[Name] Resources/ (empty folder)
After creating the files, add a new row to the
Routing Map in root CLAUDE.md.
If you remember one thing: a workstation is a specialist you brief once. The rest — folders, files, routing — is just how that brief gets stored so Claude can find it again next week.
The system only gets smarter if you save what you learn. One 30-second prompt at the end of each session — that's all it takes to turn CoworkOS from a config file into something that actually grows with you.
Preferences you mentioned, decisions you made, context Claude picked up on — none of it survives unless it's written to MEMORY.md. The AI resets completely.
One prompt at the end of every task fixes this permanently. Paste it into Cowork and Claude scans the whole conversation, then writes anything useful to the right place.
Scan this conversation for any unsaved preferences, decisions, or useful context. For each item found, route it to the right place: - Cross-cutting (role, people, projects, preferences) → root MEMORY.md - Workstation-specific (contacts, decisions, paths, tool configs for this domain) → [Workstation]/MEMORY.md - Behavioral rule (always/never/before doing X) → CLAUDE.md (root or workstation) Don't duplicate what's already there. Focus only on what's genuinely new from this session.
Pick the one job you resent repeating. That's your first workstation.
CoworkOS compounds. The more you use it, the more Claude knows you. Start small — one folder, three files — and add workstations as you go.
Questions? Find us in #bai-claude-ama on Slack.
Keep your master CLAUDE.md under 300 lines. Point to external resource files rather than stuffing everything in one place — this keeps baseline token costs low every single session.
Never repeat global rules inside local workstations. If the root already says "no corporate jargon," your Email HQ workstation shouldn't waste tokens restating it. Inheritance is free.
Sonnet 5 handles 80% of daily tasks — it's faster and cheaper. Reserve Opus 5 for complex workflows with three or more interdependent steps.
Haiku 4.5 for small, lightweight tasks; Fable 5 only for long-horizon autonomous agents. See the model slide for the full picture.
Four models, one dial: bigger = smarter and pricier. Match the model to the stakes, not the excitement.