Build your own Claude workstation system

What you're building

Every session, you re-explain the same context. Your email style. Your brand colors. A client's history. That repetition is the tax, and a workstation is how you stop paying it.

A workstation is a folder that holds everything Claude needs to do one kind of job well. You write the context down once, and Claude reads that folder whenever the work touches that domain.

Email work loads the email folder. Brand work loads the brand folder. Nothing else gets loaded, so Claude stays focused and you stop repeating yourself.

Think of it as hiring specialists instead of one generalist. Your email specialist knows your greeting conventions cold. She does not need your brand's hex codes, and loading them into her head would only slow her down.

Claude OS/                          ← your root folder, connected to Claude
├── CLAUDE.md                       ← standing instructions + routing map
├── MEMORY.md                       ← facts Claude remembers about you
├── 00_Resources/                   ← reference files any workstation can use
│   └── voice-principles.md
├── Email HQ/                       ← a workstation
│   ├── CLAUDE.md
│   ├── MEMORY.md
│   └── Email HQ Resources/
└── Brand HQ/                       ← another workstation
    ├── CLAUDE.md
    ├── MEMORY.md
    └── Brand HQ Resources/

The whole thing gets built in one guided conversation. You paste a prompt, answer questions once, and come out the other side with a working system.

Make a folder somewhere you'll actually see it, like your Desktop or the top of your Documents. Name it something plain: Claude OS, Work OS, or your company name.

Then connect it in the Claude desktop app using Add folder, so Claude can read and write inside it.

This is the whole build. Claude creates every file, then walks you through five phases, stopping after each one to ask you something. Answer as you go.

1. SkeletonClaude builds every file and folder. No questions.
2. How you like to workYour preferences and rules.
3. Facts about youYour role, company, the people you work with.
4. Your voiceYou paste writing samples, Claude extracts your style.
5. Your first workstationBuilt and wired into the routing map.

You're setting up my Claude workstation system from scratch. Build the
whole thing in one guided pass.

Work through the phases in order. Complete a phase, then STOP and wait
for my reply before starting the next one. Never ask questions for two
phases in the same message. When you do ask, put all of that phase's
questions in one numbered list so I can answer them together.

Don't guess my preferences, my facts, or my voice. If I skip something,
say so and leave it blank rather than inventing an answer.


PHASE 1 — BUILD THE SKELETON (no questions; just do it)

Create these at the root of my connected folder:

A) CLAUDE.md with these sections, in this order:
   ## Memory System
   ## Preferences
   ## Rules
   ## Routing Map
   ## References
   ## Creating New Workstations
   ## Updating Workstations

B) MEMORY.md containing:
   # [My Name / Company] Memory
   *Last updated: [today's date]*
   ## Contacts — a table with columns: Person | Role | Notes
   ## Key Facts — a table with columns: Fact | Value
   ## Key Decisions — a bulleted list; each entry states the decision,
     the date, and the reasoning behind it.

C) A folder called 00_Resources containing an empty file called
   voice-principles.md. Leave it empty for now.

Fill these sections in immediately:

MEMORY SYSTEM — instructions to yourself: read MEMORY.md at the start of
every session. Use two tests when deciding where new information goes.
Test 1: does it prescribe behavior ("always," "never," "before X do Y")?
→ CLAUDE.md. Test 2: does it describe a fact about the world that could
change? → MEMORY.md. When unsure, suggest a file and ask me to confirm.

ROUTING MAP — an empty table with columns "Workstation" and "Route here
when I...". Above it, a note: add a row here whenever a workstation is
created.

REFERENCES — a table with columns "Resource" and "Read when...".
Add one row: voice-principles.md | Writing any content on my behalf.

PREFERENCES and RULES — leave blank. Phase 2 fills them.

CREATING NEW WORKSTATIONS — write this section exactly as follows:

  When I ask you to create a new workstation, create a subfolder with my
  workstation name and add these three items inside:

  1. CLAUDE.md with these sections in this order:
     - Identity — One paragraph: who you are in this workstation, what
       routes here, what doesn't.
     - Resources — Table with "Resource" and "Read when..." columns.
       Start empty.
     - Workflow — Numbered steps for the primary task. Start simple;
       I'll refine over time.
     - Editorial Rules — Always opens with: "Follow my voice principles
       in 00_Resources (voice-principles.md)." Then add domain-specific
       writing rules that layer on top.

  2. MEMORY.md with this structure:
     - Header: "[Workstation Name] Memory"
     - Sections: Contacts (people relevant to this domain) and Key
       Decisions (reasoning behind choices).
     - You populate this over time as we work. I don't write it manually.

  3. [Workstation Name] Resources/ — Empty folder. This is where
     reference files for this domain go.

  After creating the workstation, add a new row to the Routing Map above
  so future sessions load it automatically.

UPDATING WORKSTATIONS — write this section exactly as follows:

  Two triggers keep workstations current. Both work in any workstation
  and always target the active workstation's files first.

  When I say "remember this" — capture what we just discussed and write
  it immediately, without waiting for further instruction:
  - A fact that could change (contact, rate, status, a decision and its
    reasoning) → the active workstation's MEMORY.md, in Contacts or Key
    Decisions.
  - A rule about behavior (contains "always," "never," "before X do Y")
    → the active workstation's CLAUDE.md, under Editorial Rules or
    Workflow, whichever fits.
  - A fact or rule that spans multiple workstations → the root MEMORY.md
    or root CLAUDE.md instead.
  Confirm in one line where you wrote it. If the file is ambiguous, make
  your best call, write it, and say which file you chose so I can
  redirect.

  When I say "update files" — review the full session and propose edits:
  1. Anywhere I corrected how you did something → a new rule for
     Editorial Rules.
  2. Anywhere you missed or misordered a step → an amendment to the
     Workflow.
  3. Any durable fact I stated → a new row in MEMORY.md.
  4. Any external document we referenced more than once → suggest adding
     it to the workstation's Resources folder and Resources table.
  Show me the proposed edits as a short list before writing. On my
  approval, write them all and stamp MEMORY.md with today's date under
  "Last updated."

  Always: append, don't overwrite. If a new rule contradicts an existing
  one, flag the conflict and ask which wins rather than silently
  replacing it.

When the skeleton is built, tell me in one line what you created, then
go straight into Phase 2.


PHASE 2 — HOW I LIKE YOU TO WORK

Ask me, in one numbered list:
- How long should a typical response be?
- Do I want one strong recommendation, or a menu of options?
- When should you use bullets, and when should you write in prose?
- What should you always do before starting a complex task?
- Do I default to meetings or to async (email, docs, recordings)?
- Anything else you'd need to know to write and work the way I prefer.

Wait for my answers. Then write them into Preferences (how you write and
respond) and Rules (what you must always or never do). Make each line
specific enough that I could check whether you followed it — "be
professional" is useless, "under 300 words unless I ask for more" is not.


PHASE 3 — FACTS ABOUT ME

Ask me 5-8 questions covering the stable facts you'd need to work well
with me: my name and role, my company and what it does, who I work with
most, what I'm currently focused on, and anything you'd otherwise have to
ask me repeatedly.

Wait for my answers. Then write them into MEMORY.md under Key Facts and
Contacts.


PHASE 4 — MY VOICE

Ask me to paste 20-30 samples of my real writing — sent emails, Slack
messages, posts. Tell me volume matters more than polish.

Wait for the samples. Then write 00_Resources/voice-principles.md with
these sections:
  ## Tone
  ## Sentence Style
  ## Words and Phrases to Avoid
  ## Words and Phrases I Use

Rules for writing it:
- Extract my actual voice, not the voice I'd claim to have.
- Every principle gets a real quoted example pulled from my samples.
- Describe what I actually do, including habits I might not notice.
- If something I told you in Phase 2 contradicts my actual writing, flag
  the conflict explicitly rather than picking one silently.

If I say I don't have samples handy, leave the file empty, tell me the
system works without it, and move on.


PHASE 5 — MY FIRST WORKSTATION

Ask me which single kind of work I do most often — the thing I'd most
like to stop re-explaining. Then interview me about it in one numbered
list:
- What kinds of task route here, and what explicitly does not?
- What conventions do I follow that you'd otherwise get wrong?
- What should never appear in this kind of work?
- What needs my review before it goes out?

Wait for my answers. Then build the workstation following the "Creating
New Workstations" instructions above. Show me the Identity paragraph
first and let me correct it before you write the rest. Make sure the
Identity paragraph ends with a decision rule for the edge case — a
sentence that settles what happens when a task is close to the boundary.

Finally, add the Routing Map row.


WHEN ALL PHASES ARE DONE

Show me the finished root CLAUDE.md. Then list what you created, and
remind me of the three phrases that run the system from here:
  "create a workstation for [X]"
  "remember this"
  "update files"

You're done building. From here the system runs on three phrases, with no more prompts to copy.

Build your second and third workstations whenever you next hit the work they cover, not in one sitting. A workstation written while you're actually doing the job comes out sharper than one written from imagination.

Good next workstations, in rough order of how often people build them: Email, Content (or Social), Brand, a specific Client or Account, Sales Pipeline, Finances, Recruiting.

Optional, and worth adding once you've used the system for a week or two. The base build covers what Claude knows about you. This step covers how Claude behaves when the work gets uncertain: when to keep trying, when to stop and ask, and which model to reach for by default.

A few things worth deciding before Claude has to guess:

• Retry limits. How many attempts should a failing task or automation get before Claude stops and asks you, instead of quietly retrying and burning tokens?
• Model defaults. Which model should Claude run on for routine work, and what earns a stronger one?
• Permission boundaries. What should always wait for your explicit go-ahead, no matter how confident Claude is: sending, publishing, deleting, spending money?
• Judgment calls. When your instructions conflict with what Claude thinks is the better move, should it flag the conflict and wait, or make the call and tell you after?
• Default verbosity. How long should a typical response be, and what earns more detail than that?

Interview me about how you should operate day to day, not what you
already know about me.

Ask me, in one numbered list:
- When a task or automation keeps failing, how many attempts should you
  make before you stop and ask me instead of continuing to retry?
- Should you default to asking before you're unsure, or default to
  acting and telling me after?
- Which model should you run on by default for routine work, and what
  earns a stronger one?
- What kinds of actions should always wait for my explicit go-ahead, no
  matter how confident you are: sending, publishing, deleting, spending
  money, anything else specific to how I work?
- When my instructions conflict with what you think is the better call,
  should you flag it and wait, or make the call and tell me after?
- How long should a typical response be, and what earns more detail
  than that?
- Anything else about how you operate, not what you know, that I'd
  otherwise have to correct you on more than once.

Wait for my answers before writing anything.

Then place each answer using the same test the rest of this system
uses: does it prescribe behavior? If yes, it's a Rule.
- If a rule applies to everything you do, add it to the Rules section
  of my root CLAUDE.md.
- If a rule only applies to one kind of work, add it to that
  workstation's Editorial Rules or Workflow section instead.

Confirm in one line where each rule landed.

You don't have to get this right the first time. Once it's in place, correct it the same way you correct everything else: catch Claude doing the wrong thing, then say "remember this."

The anatomy of a workstation

CLAUDE.md — the job description

Identity is one paragraph: who Claude is here, what routes here, and what doesn't. The exclusions matter as much as the inclusions. A real example:

Those last two sentences are a decision rule for the edge case, and they're worth more than three paragraphs of description. That's why the master prompt asks Claude to put one in every Identity it writes.

Resources is a two-column table, Resource and Read when.... It starts empty. Every file you drop into the workstation's Resources folder should eventually earn a row with a trigger condition.

Workflow is numbered steps for the primary task. Five or six is plenty for a first pass. You'll refine it every time you notice Claude skipping something.

Editorial Rules always opens with the same line, "Follow my voice principles in 00_Resources (voice-principles.md)," then layers domain-specific rules on top. This is where a workstation gets genuinely sharp. Greeting conventions, formatting conventions, what to never say.

MEMORY.md — the running record

Header is [Workstation Name] Memory. Sections for Contacts and Key Decisions. You don't write this by hand, Claude populates it as you work.

[Workstation Name] Resources/ — the filing cabinet

Empty at creation. Reference material for this domain goes here: a brand guide, a pricing sheet, past examples worth matching, a template.

The routing map

Creating a folder isn't enough, because Claude needs to know when to load it. That's the Routing Map in your root CLAUDE.md:

Write each trigger the way you'd actually say it out loud. "…need to draft a LinkedIn post" beats "content ideation and social distribution workflows." You're writing a lookup key for a machine that reads plain English, not a taxonomy for a filing system.

Where things go

The single most useful test in this whole system:

What deserves a workstation

Not every task needs one. A workstation earns its place when all three are true:

1. You do it repeatedly. One-off work doesn't need infrastructure.
2. It has its own rules. If the way you write emails is meaningfully different from how you write blog posts, that's two workstations. If it's the same rules with a different word count, it's one.
3. You'd have to re-explain context every time. That re-explaining is the workstation. Write it down once.

How it grows

Your first version of a workstation will be thin: an Identity paragraph, six workflow steps, two editorial rules. That's correct. Refinement comes from use, and the two triggers are the mechanism:

• Claude does something wrong → correct it, then say "remember this."
• Claude skips a step → correct it, then say "remember this."
• You explain a fact twice → say "remember this."
• End of a session with several corrections → say "update files."

"Remember this" is surgical and immediate. You use it mid-task, dozens of times a week. "Update files" is the end-of-session sweep, a broader review that catches what you didn't consciously notice.

Say "update files" at the end of any session where you corrected Claude more than once. That's the whole discipline. Everything else in this guide is setup you do once, and this is the part you do repeatedly. It's what separates a folder that gets sharper from a folder that gets abandoned.

After a month of that, a workstation stops feeling like a folder and starts feeling like someone who's worked with you for a year.

Quick reference

Setup checklist

• Folder created and connected in the Claude desktop app
• Master prompt pasted
• Phase 2 answered — Preferences and Rules written
• Phase 3 answered — Key Facts and Contacts written
• Phase 4 — writing samples pasted, voice file built
• Phase 5 — first workstation built and in the Routing Map
• Tested: fresh session, real task, confirmed it loaded the right folder
• Used "remember this" once and checked that it landed in the right file
• (Optional) Step 4 — operating rules added: retries, model defaults, permission boundaries

Your progress on this checklist is saved in your browser only. It isn't shared with anyone else viewing this page.

Appendix: build it piece by piece

The master prompt does all of this in one pass. These are the same instructions split apart, useful if you're teaching the system, rebuilding one file, or adding a piece you skipped.

Create a file called CLAUDE.md at the root of my connected folder.

Use this exact structure, with these five sections in this order:

## Memory System
## Preferences
## Rules
## Routing Map
## References

Leave the Routing Map and References tables empty for now — just the
headers and column names.

For Memory System, write instructions telling yourself to read MEMORY.md
at the start of every session, and to use two tests when deciding where
new information goes:
- Test 1: does it prescribe behavior ("always," "never," "before X do Y")?
  → CLAUDE.md
- Test 2: does it describe a fact about the world that could change?
  → MEMORY.md
When unsure, suggest a file and ask me to confirm.

For Preferences and Rules, ask me questions first. Don't guess my
preferences — interview me, then write what I tell you.

Create MEMORY.md at the root of my folder. Structure it as:

# [My Name / Company] Memory
*Last updated: [today's date]*

## Contacts
A table with columns: Person | Role | Notes

## Key Facts
A table with columns: Fact | Value

## Key Decisions
A bulleted list — each entry states the decision, the date, and the
reasoning behind it.

Populate it by asking me 5-8 questions about the stable facts you'd need
to work well with me: my role, my company, what we do, who I work with
most, and anything you'd otherwise have to ask me repeatedly.

Create a folder called 00_Resources at my root, and inside it a file
called voice-principles.md.

I'm going to give you 20-30 samples of my real writing — sent emails,
Slack messages, posts. Read them and extract my actual voice, not the
voice I'd claim to have. Structure the file as:

## Tone
## Sentence Style
## Words and Phrases to Avoid
## Words and Phrases I Use

Rules for how you write it:
- Every principle gets a real quoted example pulled from my samples.
- Describe what I actually do, including habits I might not notice.
- If I tell you a rule that contradicts my actual writing, flag the
  conflict explicitly rather than picking one silently.

Wait for my samples before writing anything.

Add this section to my root CLAUDE.md, exactly as written below.

## Creating New Workstations

When I ask you to create a new workstation, create a subfolder with my
workstation name and add these three items inside:

1. **CLAUDE.md** with these sections in this order:

   - **Identity** — One paragraph: who you are in this workstation, what
     routes here, what doesn't.
   - **Resources** — Table with "Resource" and "Read when..." columns.
     Start empty.
   - **Workflow** — Numbered steps for the primary task. Start simple;
     I'll refine over time.
   - **Editorial Rules** — Always opens with: "Follow my voice principles
     in 00_Resources (voice-principles.md)." Then add domain-specific
     writing rules that layer on top.

2. **MEMORY.md** with this structure:

   - Header: "[Workstation Name] Memory"
   - Sections: Contacts (people relevant to this domain) and Key Decisions
     (reasoning behind choices).
   - You populate this over time as we work. I don't write it manually.

3. **[Workstation Name] Resources/** — Empty folder. This is where
   reference files for this domain go.

After creating the workstation, add a new row to the Routing Map above so
future sessions load it automatically.

Add this section to my root CLAUDE.md, exactly as written below.

## Updating Workstations

Two triggers keep workstations current. Both work in any workstation and
always target the active workstation's files first.

**When I say "remember this"** — capture what we just discussed and write
it immediately, without waiting for further instruction:

- A fact that could change (contact, rate, status, a decision and its
  reasoning) → the active workstation's MEMORY.md, in Contacts or Key
  Decisions.
- A rule about behavior (contains "always," "never," "before X do Y") →
  the active workstation's CLAUDE.md, under Editorial Rules or Workflow,
  whichever fits.
- A fact or rule that spans multiple workstations → the root MEMORY.md or
  root CLAUDE.md instead.

Confirm in one line where you wrote it. If the file is ambiguous, make
your best call, write it, and say which file you chose so I can redirect.

**When I say "update files"** — review the full session and propose edits:

1. Anywhere I corrected how you did something → a new rule for Editorial
   Rules.
2. Anywhere you missed or misordered a step → an amendment to the Workflow.
3. Any durable fact I stated → a new row in MEMORY.md.
4. Any external document we referenced more than once → suggest adding it
   to the workstation's Resources folder and Resources table.

Show me the proposed edits as a short list before writing. On my approval,
write them all and stamp MEMORY.md with today's date under
"Last updated."

**Always:** append, don't overwrite. If a new rule contradicts an existing
one, flag the conflict and ask which wins rather than silently replacing it.

Create a workstation called Email HQ.

Before you write the CLAUDE.md, interview me:
- What kinds of email route here, and what explicitly doesn't?
- How do I open an email? How do I sign off?
- Does my formality change by recipient type? Walk me through each type.
- What should never appear in an email I send?
- What needs my review before it goes out?

Then write the four sections, add the Routing Map row, and show me the
Identity paragraph first so I can correct it before you write the rest.