Claude plugin · Claude Code and Cowork · how it works
A Claude that actually knows you.
Out of the box, Claude is smart and knows nothing about you. Not your team, not your accounts, not what "urgent" means to you, not the words your group uses that appear in no manual. So you re-explain yourself in every chat, and drafts sound like a polite stranger wrote them.
This plugin fixes that. It gives Claude a memory that lives in plain files you own, six routines that run your day on a schedule, and a nightly pass where Claude reads what happened, learns from it, and proposes changes. It measures whether it's getting better instead of assuming it is.
The part that makes it a system is the handoff. The briefing, the recap and the nightly dream each read what the one before them wrote, so the day gets gathered once and nothing falls on the floor.
Install
Two lines, then say "set me up."
In Claude Code or Cowork, add the marketplace and install the plugin. You need Python 3 on the machine that runs setup; macOS and most Linux installs already have it.
/plugin marketplace add jwm-axoni/claude-workbench/plugin install workbench@claude-workbench/workbench:setupor just say "set me up"
What's in the box
Ten skills. Six of them never wait on you.
A plugin is just the box it ships in. Once installed, these show up individually. Six run themselves on a schedule, one fires on its own while you chat, and three you run when you need them. The skills tab explains each one and what to say to it.
Set me up
The guided onboarding. About ten minutes of questions, and it builds your whole folder, wires up every routine, and writes your project instructions.
Daily briefing
What you owe, what landed overnight sorted by your definition of urgent, today's meetings with context, and numbered questions. Sends a short digest to you by Slack, email or Claude. Tops itself up at midday.
End-of-day recap
The one record of the day, every line with its source. Grades the morning's claims and checks every draft you were given against what you actually sent.
Nightly dream
The one that learns. Files the day into people, projects, orgs and tools, builds the relationships between them, and turns your corrections into rules.
Weekly
The week-ahead plan, a "is it getting better?" scorecard, and the librarian's checks with the safe fixes applied.
Calendar sync
A dated copy of your calendar that records moves and cancellations, so the briefing reads a file instead of guessing what changed.
Capture
When a chat settles a decision, a fact, or a correction, one line lands in today's capture file and the dream files it that night. Drafts get saved too.
Librarian, Schedules, Migrate
A health audit of the folder, a reset of your scheduled routines, and safe upgrades that never overwrite what you edited.
Getting started
Say "set me up." Here's exactly what happens.
Run this even if you've been using Claude for months. Onboarding is not about learning Claude, it's about Claude learning you, and it builds files and routines that regular use never creates. It takes about ten minutes, and you never touch a file by hand.
Look at the folder before touching it
An empty folder gets the full build. An existing workbench gets a short "what changed?" instead of the whole interview. An existing notes vault gets adopted: nothing moved, nothing overwritten, and each folder move suggested one at a time for your yes.
Who you are
Name, role, team, who you report to, your company and what it does, and where you live, which sets the timezone every routine runs on. It fills in what it can from your account and asks you to confirm. Your role picks a role pack: sales, customer success, marketing, engineering, product, operations, people, IT and security, finance, leadership, or general.
"Looks like you're a Senior AE on the Mid-Market East team. I'll set you up with the sales pack. Right?"How you work
Your hours and workdays, the tools you live in, the two or three things you need to land this quarter, what "urgent" actually looks like for you, how you like answers written, and any calendar blocks that should never show up in a briefing.
Check what Claude can actually reach
One real, read-only call per tool. Each one is marked connected, needs sign-in, missing, or manual. A broken connection never blocks setup; you get the exact place to fix it, and every routine says "couldn't check" instead of implying nothing happened.
Decide how the day reaches you
A Slack DM to yourself, an email to yourself, a message in the Claude app, or just the file. It sends a test before moving on. It also offers Obsidian, free, for browsing the folder and seeing the graph of who and what connects.
Build the folder
The operating contract at the top, a schema in every folder, and the first pages already linked: you, your manager, your company, one page per tool. A hash of every file it writes is recorded, so later upgrades never overwrite what you change.
Turn on the routines
Six scheduled tasks, all named [Workbench], timed off your
hours. If you already have older tasks doing the same job, it lists them and
asks before replacing each one, so you never get two briefings.
Make one project your home
It writes your project instructions and has you paste them into one Claude project with the folder attached. From then on every chat starts from your memory, and everything a chat settles flows back into it. It ends with three things to try today.
Where your day comes from
Your role decides what gets watched first.
The recap sweeps every tool Claude can reach, once, at the end of the day. The role pack sets the order and what each routine pays extra attention to, so a seller's briefing leads with deals and a manager's with their people.
Sales
Calendar, email, CRM, call recorder, chat. Watches deals closing this month with no activity, and follow-ups promised on calls that haven't gone out.
Pack: salesCustomer success
CRM, CS platform, ticketing. Watches renewals in the next 90 days, rising ticket volume, and promises to customers that are past due.
Pack: customer-successEngineering
Code host and issue tracker first. Watches PRs waiting on your review and tickets that changed since yesterday.
Pack: engineeringLeadership
Calendar, mail, chat, docs. Watches 1:1s, commitments made to your team, and direct reports you haven't met with in three weeks.
Pack: leadershipPeople and HR
Interviews, debriefs, open roles. Keeps candidate and employee personal data out of the folder entirely, by rule.
Pack: peopleEveryone else
Marketing, product, operations, IT and security, finance, and a general pack. If nothing fits, it says so and you pick.
Packs: 11 in allNothing is hardcoded. The tools above are examples because most people have them. Name whatever you use and setup will probe it. A tool that can't be reached is written down too, so a routine can say it couldn't check rather than implying nothing happened.
Routines read, they never act. No routine replies, posts, updates a record or changes a ticket. The one exception is your own digest, and it only ever goes to you.
Your folder
One folder tree, about you, in plain text you own.
Setup creates one folder. The suggestion is Workbench, but it's about you, so name it anything you'll recognize. Everything Claude knows about you lives in it, as markdown you can open, edit, back up, or take with you.
- Workbench/
- CLAUDE.md — the operating contract: folder map, who writes what, conventions
- Context/ — me.md (how you work), voice.md (how you write), surfacing.md (what your briefing always or never raises)
- Daily/ — briefings, recaps, calendar copies, weekly plans
- Projects/ — one page per project, plus INDEX.md
- People/ — one page per person, plus the name-variants table
- Organizations/ — your employer, customers, vendors, partners
- Resources/ — durable reference, Tools/ inside, the glossary
- Tracking/ — goals, commitments both ways, friction, the dream log, improvements
- Captures/ — what your chats settled, one file a day
- Drafts/ — every draft you were handed to send
- Signals/ — what each routine claimed, how it scored, and the run ledger
- Meetings/ — one file per meeting ← only if your recorder exports notes instead of connecting
- Archive/ — closed chapters, read-only
Every folder carries its own schema. Each holds a small CLAUDE.md that says what a page in that folder looks like, with a shape you can copy straight out of it. There is deliberately no templates folder. A second copy of a schema drifts from the first, nothing checks the copy, and you find out when half your pages have one shape and half another.
One writer per file. Each routine owns specific files and writes nothing else. Your chats write only to Captures and Drafts. Nothing races anything, which is what keeps a synced folder from silently stalling. The one deliberate exception is the run ledger, where every routine appends a dated line and none edits one.
Promises get a register. COMMITMENTS.md holds every commitment in flight, yours and other people's. The dream adds rows and flags the ones that look done or stale, the briefing asks about overdue ones as questions, and only the weekly closes a row, on evidence.
Conflicts become questions. When two sources disagree, the page
gets a Needs reconciling section instead of a guess. The weekly asks you about the
oldest three, and your answer supersedes the losing claim without erasing it. New
claims carry their evidence and confidence, so /workbench:why can show
you where any fact came from.
Nothing is deleted. A fact that stops being true is marked superseded, with a link to what replaced it, then archived once nothing points at it.
Short-term and long-term memory are different folders. Captures hold what your chats settled until the dream files them that night. Daily files are a diary, frozen once the day is over. People, project and org pages are the long-term memory, kept current by state rather than age, and a page nobody has mentioned in 90 days gets flagged for you to confirm, never moved. Meeting notes come through your recorder's connector when there is one, so transcripts don't pile up here.
Read the page, not the source
One page per project, person and account, always current.
A recorded call is an hour of talking. Nobody reads that again, and neither should Claude. So every call, thread, ticket and capture gets folded into the page it's about, kept current overnight. When you ask what's going on, the answer comes off that page in one read, with its sources.
Open a recording only when the page names something unresolved and points at it.
Pages link through their properties, not just their prose. A
project lists its owner, people, orgs
and tools; a person lists their org,
reports_to and projects. Each is a wikilink, so
Obsidian's graph and backlinks use it, and every link is kept on both pages.
Put Sam on a project and the project shows up on Sam's page the same night.
A reversed decision doesn't disappear. It moves down the page marked superseded, with a link to what replaced it. That's the difference between "we never agreed to that" and "we agreed it on the 12th and changed it on the 19th."
The whole idea
Each routine reads what the last one wrote.
This is what makes it a system instead of six jobs that happen to run on the same calendar. Every handoff is consumed by the next thing along, and the last one feeds tomorrow's first. Nothing gathers the same day twice.
The loop closes overnight. Tomorrow's briefing opens with the dream, asks what it couldn't settle, and the chain starts again.
The dream never goes back out to re-read your day. The recap already did that and wrote it down, so the dream learns from the record rather than from the raw feeds. That's why two routines never tell two different stories about the same day, and why a missed evening doesn't break the morning.
Calendar sync runs alongside, keeping a dated copy of today and tomorrow and logging every move and cancellation. The weekly sits beside the chain: on your first workday it reads the week's recaps and writes the plan, the scorecard, and the maintenance pass.
| Routine | Default time | What it writes |
|---|---|---|
| Calendar sync | Before work, 11:30, end of day | A calendar copy for today and the next workday |
| Daily briefing | 30 minutes before work | The morning briefing file, and your digest |
| Midday update | 12:00 | An update appended below the briefing |
| End-of-day recap | 30 minutes after work | The daily recap, and the day's scores |
| Nightly dream | 22:00 on workdays | Its own blocks on your pages, the links between them, the dream log |
| Weekly | Before work, first workday | The week-ahead plan, the scorecard, structural fixes, archive moves |
| Librarian | When you ask, plus inside the weekly | A report; nothing until you approve a fix |
The part that compounds
The nightly dream
Every other routine reports. This one learns. While you're asleep it reads what the day's recap and your chats left behind, works out what actually changed, and updates the pages that hold your long-term knowledge. Four things come out of it, all landing in the same place. A system that can't update itself is just six reminders on a timer.
Your day is the input. Nobody fills in a form.
Every change carries a receipt. The dream writes one log entry per run and ends each finding with a "Traces to" line pointing at the evidence. If you disagree with something it changed, you can find out what caused it in one hop.
What it can't settle, it asks. Once. Two people who might be the same person, a term it's never seen, a commitment that looks done: it never guesses on those. Tomorrow's digest asks, numbered, and "1 yes, 2 it's Jo Tan" is enough. The recap reads your answer and the dream acts on it that night.
It starts careful. For its first three nights it only describes what it would change, so you can see how it reads your work. Then it goes live, writing only the blocks marked as its own. Your words, your goals and your notes it never rewrites.
"Nothing notable" is a valid result. A quiet day gets a short log saying so. It does not invent a lesson to justify having run.
On autopilot
What your week looks like once it's running
plan, scorecard and maintenance, then the daily cycle
briefing, recap, dream
briefing, recap, dream
briefing, recap, dream
briefing, recap, dream
The repetition is the point. The same steps run every workday, so the chain never has a gap to climb out of. Your part is one digest a day: read it, answer what it asked, and that's the whole of what it needs from you. Your days off stay quiet, and your workweek is whatever you told setup.
One honest catch: where the routines run. That follows from two setup answers, which Claude you have and where the folder lives, and setup works it out rather than asking you to pick. A folder on your laptop means routines that need the laptop. Only a private GitHub repo with Claude Code lets them run with the laptop closed. For a team rollout or more than one computer, setup recommends the repo.
On Team and Enterprise plans, Cowork runs sessions in the cloud only if your organization turns on "Run Cowork in the cloud" (Organization settings > Cowork), and a task that needs local files still runs locally.
| You have | The vault lives in | Routines run on | What's guaranteed |
|---|---|---|---|
| Cowork only | a local folder | Cowork local | Runs only while the computer is on and the desktop app is open. No documented catch-up: the run ledger and gap detection recover missed runs. |
| Cowork only | a private GitHub repo, worked from a local clone | Cowork local, on the clone | Runs only while the computer is on and the desktop app is open. No documented catch-up: the run ledger and gap detection recover missed runs. The repo adds backup, history and safe use from more than one computer. |
| Claude Code | a local folder | Code Desktop local | Runs while the desktop app is open. Exactly one catch-up run for the most recent missed run within 7 days. |
| Claude Code | a private GitHub repo, worked from a local clone | Code cloud | Runs with the laptop closed. Research preview: fresh clone per run, pushes to claude/ branches, minimum interval one hour. No documented catch-up: the run ledger and gap detection cover misses. |
Missed runs show up instead of vanishing. Every routine appends one line to a run ledger when it finishes. The next briefing checks it before anything else and says in plain words what didn't run and what that cost: "the dream didn't run last night, so 4 captures are unfiled." The dream then files every unfiled capture whatever its date, a day with no recap gets a short file marked reconstructed that lists what it couldn't sweep, and a late run says so in its own file. None of it depends on a scheduler promising to catch up.
Getting the day
You don't have to open the folder. The day comes to you.
Most people won't browse a folder every morning, and they shouldn't have to. Setup asks where the digest should land and sends a test. You answer its questions by replying, and the answers flow back into the folder that night.
A DM to yourself
Focus, risks, meetings that need prep, and numbered questions. Reply in the thread.
One email to you
The same digest in your inbox. Reply to answer. Needs an email connector that can send.
A message in the app
The scheduled task's own message. Answer in your workbench project and it's captured.
Just the file
The full briefing waits in Daily/ for whenever you want it.
The digest is the only thing a routine ever sends, and only to you. Never to anyone else, never into a shared channel. It's written into your folder's contract as a dated exception, so a later reader can see the decision was made on purpose.
Measured, not assumed
Three ways it learns, and how you'll know it's working.
"It learns" is easy to claim. Each loop here leaves evidence in the folder, and once a week a scorecard compares this week with the last four.
| The scorecard tracks | Good direction |
|---|---|
| Drafts sent as written or lightly edited | Up |
| Drafts rewritten before sending | Down |
| Briefing claims that held up | Up, or steady above 85% |
| Repeat corrections on a topic that already has a rule | Down |
| Digest questions answered within two days | Up |
| Proposals built and verified | Up |
Silence is not agreement. If you aren't answering, the confidence bar holds rather than drifting, and the scorecard says there's too little data instead of printing a flattering percentage.
You stay in charge of the rules. Routines can propose a change to the operating contract but never write it. You say yes, and the change is made and logged.
The one project
Work from one project, and everything it settles comes home.
Separate projects split your memory. One keeps it whole. Setup writes
your instructions to .workbench/PROJECT_INSTRUCTIONS.md; paste them into
one Claude project with the folder attached and start every piece of work there.
They have four parts.
The folder first, every time
The contract, then how you work, the name variants, your goals and the glossary, and today's captures. Look things up in the folder and cite the page; if it isn't there, say so instead of guessing.
One line per thing settled
Decision, process, fact, correction, or an answer to a digest question, logged on the turn it's settled, in its own words, with people and projects linked. It tells you in one line what it logged.
Saved as handed over
Every email, message or post it drafts for you to send is kept exactly as given, so the recap can see what you changed.
Captures and drafts only
Everything else has an owner. Ask it to change a specific page and it makes that one edit and logs it.
How you use it
Ten skills. You rarely type a command.
A skill runs one of two ways: on a schedule, without you, or from
something you say in plain words. Claude picks the skill from your words. Every one
also has a command, /workbench: plus its name, if you'd rather be exact.
Setup
What it does. The guided onboarding: checks your tools, learns your role, your rhythm and how you want the day delivered, builds the folder, turns on the routines, writes your project instructions.
How you use it. Say "set me up". Later, "my role changed" or "I switched to Outlook" runs it again for just that.
What it writes. The folder, once. On a re-run, only files you haven't edited.
Daily briefing
What it does. What you owe, what landed overnight sorted by your idea of urgent, today's meetings with context, what the dream learned, and its questions.
How you use it. Nothing, it arrives. Ask "what's my briefing" any time.
What it writes. One file a day in Daily/, and your digest.
End-of-day recap
What it does. Sweeps every tool once and writes the day down, every line sourced. Checks the morning's claims and your drafts against what happened.
How you use it. Nothing. Or "recap my day" to run it early.
What it writes. The recap in Daily/ and the day's scores.
Nightly dream
What it does. Files the recap and your captures into the right pages, links people, projects, orgs and tools both ways, and learns from corrections.
How you use it. Nothing. Its questions arrive in tomorrow's digest.
What it writes. Only the blocks and links marked as its own, and its log.
Weekly
What it does. Priorities, carryover, meetings needing prep, the scorecard, and the maintenance pass: frontmatter, links, one-sided relationships, archive moves.
How you use it. Nothing. Or "plan my week".
What it writes. The plan, its scores, the improvements list, and structural fixes.
Calendar sync
What it does. Keeps a copy of today and tomorrow and logs every move and cancellation. Passcodes stripped, outside guests shown as a count.
How you use it. Nothing.
What it writes. One calendar file per day, frozen once the day's over.
Capture
What it does. Logs what a chat settled, one line each, from any chat in any surface, and saves drafts you were handed to send.
How you use it. It happens on its own. Say "remember this" or "log that decision" to make sure.
What it writes. Today's capture file and the drafts folder. Nothing else.
Librarian
What it does. Audits the folder across 34 checks: stale pages, dead links, duplicates, one-sided relationships, routines missing from the run ledger, a malformed commitments register, conflicts left unreconciled, pages gone quiet.
How you use it. "Run the librarian" or "audit my folder".
What it writes. A report. Nothing until you approve a fix.
Schedules
What it does. Makes your scheduled tasks match the plan. Finds older look-alikes and asks before replacing each, so you never get two briefings.
How you use it. "My briefing didn't run" or "move the dream to 11pm".
What it writes. Your scheduled tasks, after you see the plan.
Migrate
What it does. Upgrades after a plugin update, merges anything you edited with the new version, or brings an older notes vault into this layout.
How you use it. "Upgrade my workbench" or "migrate my vault".
What it writes. Only what you approve, after a snapshot.
Cheat sheet
Things to say to it
"set me up"runs onboarding, or updates it when something about you changed"what do you know about Sam?"everything on Sam's page and what's open between you, with sources"where are we on the Acme rollout?"status, decisions, and what's owed, off the project page"remember this..."captures a fact, decision or correction for tonight's dream"run the librarian"audits the folder and hands you a fix per finding"why do you think the pilot starts in January?"runs/workbench:why: the evidence, confidence, and what that fact replaced"1 yes, 2 it's Jo Tan"answers this morning's digest questions
And one honest expectation. The first week is the worst week. Claude knows the least about you on day one, and the value compounds from there. Every answer and every correction makes the next week better, and the scorecard shows you whether it is. Give it two weeks before judging it.
Worth understanding once
Claude has three kinds of memory. Trust them differently.
| Built-in memory | Project instructions | Your folder | |
|---|---|---|---|
| What it is | Claude remembers things from chats on its own | Standing rules in the project's settings | The pages in your folder tree |
| Can you read it? | Not fully | Yes, one text box | Yes, plain files |
| Who updates it | Claude, silently | You paste what setup writes, once | The routines, with a receipt for every change |
| Treat it as | A bonus, never the plan | How to use the folder, nothing more | The source of truth |
The guidance in one breath: don't trust the black box with anything important, keep the project instructions about how to use the folder rather than what's in it, and put everything that matters in the folder, because it's the only memory you can read, correct, back up, and take with you. When you wonder "does Claude know this," the answer is whatever's on the page, and you can go look.
If you set this up before
Already running an older version, or your own notes? Start here.
If you have earlier scheduled tasks or a notes vault of your own, don't just install this on top. The old tasks keep firing, you get two briefings every morning, and within a week you're ignoring both.
Say you already have it
Run setup or "migrate my vault". It looks at the folder first, recognizes an existing workbench or notes vault, and adopts it instead of starting over.
Pause, never delete
Older tasks that do the same job are listed one by one. Each gets paused on your yes, not deleted, so the wording inside survives and you can switch it back on.
Your edits survive upgrades
The plugin remembers every file it wrote. Files you changed stay as they are, and the new version lands beside them for you to merge, with a snapshot taken first.
Your old notes come with you
Existing folders get mapped to the new ones, moved one at a time on your yes, with every move logged so it can be undone. Nothing is left stranded.
Switch over in one go rather than running both for a week. Two systems writing two briefings teaches you to skip both, and the habit is hard to undo once it forms.
Open source under the MIT license: github.com/jwm-axoni/claude-workbench