How to Build a Trading Desk in Claude Projects (2026 Guide)

You already have an AI trading assistant. It is one chat window, it is holding forty unrelated jobs, and by Wednesday it has forgotten your risk rules and started telling you that gold looks bullish.

The fix is not a better prompt. It is a different shape. Stop running one assistant and start running a desk: four narrow seats, each with one job, one set of files, and a written list of things it is forbidden to do. A Chief of Staff who tells you whether you should be at the screen today. A Research analyst with no opinion on direction. A Journal that grades process rather than profit. A records clerk who keeps the tax year straight. None of them place a trade. You remain the only person on the desk who can size, enter, or override.

Claude Projects are a good place to build that desk, and one architectural fact decides how you build it. Get that fact wrong and you will spend three months having the same conversation every Sunday with something that keeps losing the last one. There is a sibling guide that builds the identical roster out of Grok bots if you are on that stack instead. The desk model is the same. The plumbing is not.

The one fact that changes the design

A Grok Bot channel is a room. The bots are in it together, they can see each other work, and they can hand tasks along. A Claude Project is not a room. It is a filing cabinet with a note taped to the front.

By default, every chat inside a Project starts fresh. It does not inherit what happened in the chat you had yesterday, or an hour ago, in the same Project. Three things can reach across: the custom instructions on the Project, the files in its knowledge panel, and, if you have the setting switched on, Claude’s ability to search your past chats and the memory built from them.

Those three are not equivalent, and the gap between them is the whole design problem. Instructions and files are exact. Same text, every time, read verbatim. Past-chat search is retrieval, and what comes back is frequently a summary of a session rather than the session itself. That is genuinely useful for continuity. It is not a source you would size a position from.

One thing does not cross at all: attachments. Upload a CSV of fills into a chat, work through it, close the tab, and a later chat can read what was said about that file but cannot reopen the file. The data is gone. For a desk built on statements and journal exports this is the single most practical line on the page. If a document matters more than once, it belongs in project knowledge, not in a message.

Traders meet this the expensive way. You spend a Sunday walking Claude through your last thirty trades, reach a genuinely useful conclusion about your Tuesday afternoons, feel understood, and never write it down. Months later the outline of that session can probably be dug back out of your history. What will not come back is the exact figure, the precise rule you decided to tighten, or the spreadsheet you pasted in. The conclusion was said. It was never recorded, and those are different things.

Once you accept that, the Claude desk becomes straightforward, and in one respect it is better than the Grok version: because nothing passes between seats automatically, no bot on this desk can quietly influence another. You are the switchboard. On a desk where the failure mode is a research seat talking a journal seat into forgiving a rule break, that is worth having.

The surfaces you build on

Four layers, from broadest to narrowest. Use all four and the desk holds together. Use only the first and you have a chat window with opinions.

• Profile instructions. Account-wide, applied everywhere. Put nothing trading-specific here. One line is enough: British spelling, direct answers, tell me when you cannot verify something.
• Projects. Custom instructions plus a knowledge panel plus a scoped chat list. This is where a seat lives. Files up to 30MB each, and a capacity meter you will not trouble for a while – one tester filled a project with roughly 845,000 words across nine files before hitting 98 per cent, at which point the next upload was refused rather than silently truncated.
• Skills. The distinction Anthropic draws is clean: a Project is what Claude needs to know, a Skill is how it should be done. A Skill is a reusable procedure that runs in any chat, inside a Project or outside one. Your journal-grading routine is a Skill. Your playbook is a Project file.
• Connectors and scheduled tasks. The automation layer, covered further down. This is where the morning brief stops being something you remember to ask for.

Anthropic’s help centre is clear on availability. Projects are open to all users including free accounts, and a free account can create a maximum of five. That is the four-seat desk with one spare. What free does not get is the retrieval layer: enhanced project knowledge, where Claude searches a body of documents too large to hold at once, is a paid feature on Pro, Max, Team and Enterprise. In practice a free desk runs fine on a lean playbook and a short journal file. Sharing a Project with another person is Team and Enterprise only, which for a solo trader is irrelevant and for a trading group is the entire question.

Two layouts

Pick one based on how many Projects you can actually create.

Layout A: one Project per seat

Four Projects: Chief of Staff, Research, Journal, Tax and Ops. Each gets its own custom instructions and only the files that seat needs.

This is the cleaner build and the one to use if you can. The Journal never sees your event research. Tax and Ops never sees your playbook. Each seat is genuinely narrow, which is the whole reason the desk works, and there is no risk of a chat drifting from one role into another because you switched topic halfway down.

The cost is that you carry findings between Projects by hand. That is the switchboard problem, and there is a protocol for it below.

Layout B: one Project, four Skills

If Projects are capped on your plan, or you simply do not want four of them, build a single Project called Trading Desk, load every file into it, and put the four roles in as Skills with a routing line in the custom instructions.

ROUTING
This project holds four separate roles. I will name one at the start of
every message. Adopt only that role and follow only its rules.
BRIEF    - Chief of Staff. Diary, calendar, open risk, trade or no-trade.
RESEARCH - Analyst. Event map and sources. No direction, ever.
JOURNAL  - Debrief. Process grade against my playbook. Blunt.
TAX      - Records clerk. Statements only. No advice.
If I have not named a role, ask which one. Do not choose for me.
Never carry a conclusion from one role into another inside the same chat.
If I switch roles mid-chat, start a new chat instead and tell me to.

The last two lines are the ones that matter. Role-switching inside a single chat is how the Journal ends up softening a grade because the Research role spent four paragraphs explaining why the market was difficult that day. It was difficult. It is always difficult. The trade still broke the rule.

What you upload once

The same five documents the Grok desk runs on, because they are properties of your trading rather than of any product: your playbook with named setups and invalidation criteria, your risk rules on one page, your session template with fixed fields, an account list with no credentials in it, and your prop firm’s own rules PDF if you are funded. Add your last twenty journal rows so the Journal seat has patterns to see rather than starting cold.

If any of those do not exist yet, write them before you create a single Project. A desk cannot enforce rules you have never written down, and asking Claude to draft your risk rules for you defeats the point in a way that will not become obvious for several months.

Two Claude-specific notes on files. First, keep each document short and single-purpose rather than merging everything into one master file, because when a Project grows past what fits in the context window Claude switches to searching your files rather than reading all of them, and search finds a focused document more reliably than a section buried in a long one. Second, when instructions and a file disagree, the outcome is not reliably predictable. Do not put your risk numbers in both places. Put them in the file and have the instructions point at it.

The four seats, as Project instructions

These are shorter than their Grok equivalents and each carries an extra clause the Grok versions do not need: a standing order to take every number from the knowledge files rather than from recall or retrieved chat history. Paste them into the Project custom instructions field, or into a Skill if you are running Layout B.

Chief of Staff

ROLE
You are my Chief of Staff. You run the desk diary. You do not run the trades.
Every number you give me comes from the files in this project, not from
recall and not from anything you retrieve out of an older chat.
ALWAYS
- Open with one line: TRADE or NO-TRADE, and the reason.
- Read my open risk and my rules from the files. Never from recall.
- List scheduled high-impact events for my instruments, in my local time,
with the source.
- Flag the rule in my session rules that today's calendar puts under pressure.
- Write "I could not verify this" rather than guessing a time or a figure.
- Under 250 words.
NEVER
- Give me a direction, a bias, or a level.
- Invent an event, a time, a price or a quotation.
- Quote a figure from an older chat. Chat history is a summary, not a record.
OUTPUT
1. TRADE / NO-TRADE, one sentence
2. Calendar: time, event, instrument, source
3. Open risk: position, R at risk, distance to daily loss limit
4. The rule under pressure today
5. Overnight tape, three lines maximum

Research

ROLE
You are my Research analyst. You build the event map. You have no view
on direction. Web search on.
ALWAYS
- Cite a primary source with a date for every factual claim: exchange notice,
filing, central bank release, official statistic.
- Label commentary, social media and aggregators as SECONDARY.
- Answer "what changed since my last session" first.
- Split FACT from INTERPRETATION under clear headers.
- Write "I could not verify this" and leave the gap visible. Do not fill it.
NEVER
- Recommend a trade, a bias or a level.
- Present a forecast as a fact.
- Give me a price I should treat as tradeable. Prices come from my platform.
- Describe a chart you have not been shown.
OUTPUT
1. What changed, five bullets maximum
2. Confirmed events ahead: date, time, instrument, source
3. Facts
4. Interpretation, labelled
5. Unverified and gaps

Journal

ROLE
You are my Journal. You grade process, not outcome. You are allowed to
be blunt. Work from the journal file only. If a trade is not in the file,
it did not happen, whatever an older chat appears to say.
ALWAYS
- Use my session template exactly. Same fields, same order.
- Grade each trade A to D on process against my playbook file, regardless
of profit or loss.
- Name the setup from my playbook, or write "not in playbook".
- Separate what I planned from what I did.
- Flag any breach of the 1% risk rule or the 1:2 minimum in plain words.
- Ask one question about the trade I least want to discuss.
- End by giving me the entry as a block I can paste back into my journal file.
NEVER
- Congratulate me on a profitable trade that broke a rule.
- Soften a process failure because the outcome was good.
- Invent a fill, a price or a time I did not give you.
- Suggest what I should trade next.
OUTPUT, per trade
Setup | Planned | Actual | Process grade | Rule breaks | One question
Then three sentences of session summary. No encouragement.

That second-to-last ALWAYS line is the whole Claude adaptation. Because the chat evaporates, the Journal has to hand you something to save. Paste it back into the journal file, re-upload weekly, and the seat compounds. Skip that step and you are having the same conversation every Friday with something that has forgotten the last four. If you want the manual version of this loop before automating it, our walkthrough on reviewing your trading journal with AI each week covers the analysis pattern in a single chat.

Tax and Ops

ROLE
You are my Tax and Operations clerk. You keep records straight.
You are not my accountant and you never pretend to be.
ALWAYS
- Work only from the statements and CSV files in this project.
- State the jurisdiction assumption at the top of every output.
- Keep a running fee and commission log per account and per venue.
- Track lots, dates and holding periods exactly as the source shows them.
- Report realised P&L year to date separately per account, and separately
again for prop firm payouts.
- End every tax-adjacent answer with: "Confirm with a qualified practitioner
in your jurisdiction."
NEVER
- Name a relief, allowance, deduction or election unless it appears in a
document in this project.
- Estimate a missing figure. Write MISSING and name the file to fetch.
- File, submit or draft anything addressed to a revenue authority.
- Merge accounts or currencies without saying so.
- Ask me for a login, an API key or a password.
OUTPUT
1. Jurisdiction assumption
2. Realised P&L year to date, by account
3. Fees and commissions, by account
4. Prop payouts received, by firm
5. MISSING items
6. Practitioner note

The credential prohibition matters less here than it does on a Grok desk, where bots run on a shared cloud computer and sign into tools as you. Claude Projects have no browser and no filesystem of their own; they read what you upload. That is a meaningful safety advantage for a trader and one of the better reasons to choose this stack. It does not make an uploaded statement private, though. Strip account numbers, and check your provider’s retention policy rather than assuming it.

The bench, as Skills

The Risk Auditor and the Playbook Librarian work better as Skills than as Projects, because both are procedures rather than bodies of knowledge. The Auditor reads your journal file against your risk rules once a week and reports breaches as a count, never as a story, and is never told whether you made money. The Librarian answers one question – does this match a setup I already trade – with MATCH, PARTIAL or NO MATCH, quoting your own criteria back at you. Full instruction blocks for both are in the Grok desk guide; they transfer unchanged apart from the file-first clause.

The switchboard protocol

Layout A costs you the handoffs Grok gives for free. Six lines cover it, and the important thing is what you carry rather than how you carry it.

1. Carry findings between Projects, never conversations. One or two sentences of conclusion, not a pasted transcript, and never a “as we discussed last week” that relies on retrieval to fill in the number.
2. A finding worth repeating goes into a file, not into the next chat. If you paste it twice, it belongs in the knowledge panel.
3. Research output reaches the Chief of Staff only as a dated line in a catalyst file. Never as narrative.
4. The Journal receives fills and your own notes. It never receives Research output, because today’s explanation is tomorrow’s excuse.
5. The Risk Auditor reads the journal file and the risk rules. Nothing else, and nobody briefs it first.
6. Nobody changes size, stops or targets. That is one person, and it is you.

Rule four is the one traders break. Rule two is the one that makes the desk compound.

Making the morning brief actually fire

A brief you have to remember to ask for is a brief you will skip in the week you most need it. Claude has a genuine answer to this, and it is worth knowing where it lives, because it is not in the Projects interface.

Scheduled tasks live in Claude Cowork. You describe a task once, pick a cadence, and it runs on its own, delivering a finished output rather than a reminder to go and do the thing. Per Anthropic’s help centre the route is Scheduled in the left sidebar, then New task in the upper right, then one of two options. Create with Claude interviews you and asks you to confirm before it saves anything. Set up manually gives you a form: task name, the prompt, an approval mode, the cadence, and optionally a model and a working folder. The cadences on offer are hourly, daily, weekly, weekdays, or manual.

One detail is worth the whole section. Scheduled tasks run remotely, so they fire on their cadence even when your computer is asleep or the desktop app is closed. For a trader whose instrument opens while they are asleep, that is the difference between a brief waiting for you at breakfast and a brief that never ran.

There is an exception worth designing around. Anthropic notes that a task requiring local files or local apps will only run locally. A desk whose knowledge lives in files saved to your Claude account keeps working overnight. A desk pointed at a folder on your laptop does not. Build for the former.

Scheduled tasks are available in Cowork on the paid plans, Pro, Max, Team and Enterprise, and each run starts as its own fresh session, reading your files and connectors at execution time rather than remembering the last run. That is the same rule as everywhere else on this desk: the files are the memory.

Three worth scheduling, and only these three at first.

• Weekday pre-session brief. The Chief of Staff prompt, an hour before your session, with your calendar and risk files attached.
• Friday journal pattern report. The Journal reads the week’s entries and returns the recurring three sentences you have been avoiding.
• Sunday risk audit. Breach counts against your written rules. Read it before you plan the week, not after.

Run each new task in review mode for two or three cycles before you trust it. Read the output properly, check one factual claim in it against a primary source, and only then let it become background. A scheduled task that has been wrong quietly for a month is worse than no task, because you have been making decisions around it.

Connectors, and where the chart comes in

The desk so far runs on files you upload. Connectors let Claude reach live sources through the Model Context Protocol, which is where a research seat stops being a well-organised guess. Availability differs by connector and by plan, and the list moves often enough that printing it here would date the page within a month. Anthropic maintains a connectors section in its help centre, and your own settings screen is the authority on what you have.

Two sensible uses on a trading desk, and a hard boundary. Sensible: a read-only connection to wherever your journal or trade log already lives, so the Journal seat is not waiting on a manual export; and a chart connection so Research is looking at your actual instrument rather than describing one from memory. We covered the second in detail in our guide to connecting TradingView to Claude, including where it works and where the honest answer is that it does not.

The boundary is unchanged from the Grok desk and it is not negotiable: nothing on this desk connects to an account you can lose money from. Not the broker. Not the exchange. Not the prop dashboard. The desk handles prep, records and review. Execution is a human, and the human is you.

Six ways this goes wrong on Claude specifically

1. You mistake a chat for a record. The single most common failure, and past-chat search makes it worse rather than better, because you can now retrieve the gist of that brilliant Sunday session and mistake it for the session. Gist is not a figure. Fix: nothing counts until it is in a file. Make the last step of every session a paste-back into your journal or catalyst file.
2. You upload the CSV to the chat instead of the Project. It works beautifully for an hour and is unreachable by tomorrow, and nothing warns you. Fix: a document you will need twice goes in the knowledge panel. A document you will need once can be an attachment. When unsure, knowledge panel.
3. Instruction and file disagree. Your custom instructions say 1 per cent, your risk rules file says 0.75 per cent because you tightened it in March, and the output is now unpredictable. Fix: numbers live in exactly one place. Instructions point at the file.
4. Retrieval mode arrives unannounced. As a Project fills, Claude stops reading everything and starts searching. Nothing tells you this has happened, and answers can quietly become less complete. Fix: keep each Project lean and single-purpose, split long documents into focused ones, and archive last year’s statements out of the Tax Project rather than letting them accumulate.
5. Project sprawl. Four seats becomes eleven Projects, three of which you have not opened since June. Fix: the same rule as the Grok desk. If you have skipped a seat’s output twice, retire it. An unread Project is a lie you tell yourself about how prepared you are.
6. Role drift inside one chat. Particularly on Layout B. You open with JOURNAL, drift into asking why the market moved, and the grading softens. Fix: the routing instruction above, plus a personal habit of opening a new chat whenever the role changes. It costs nothing. Chats are free.

The thirty-day install

Same order as the Grok desk, because the order is about your habits rather than the product.

Week one: Chief of Staff and Journal Projects. Write your risk rules, session rules and session template first. Paste every Journal entry back into the journal file before you close the chat. That habit is the build.

Week two: Research Project, web search on. Test the source hygiene deliberately by asking about something you already know cold. If it hands you a confident authority that does not exist, tighten the block before you rely on it for anything you cannot check.

Week three: Tax and Ops. Upload a real CSV and count the MISSING items. That number measures your record-keeping, not Claude’s competence. Fix the records.

Week four: the Risk Auditor Skill, plus your first scheduled task. Start with the Friday pattern report rather than the daily brief, because a weekly cadence gives you three chances to catch a problem before it becomes habit.

Day thirty-one: cull, then re-upload. Journal file refreshed with a month of entries, catalyst file trimmed, anything unread retired.

For readers who already live in the terminal

If you use Claude Code daily, there is a more elaborate version of this desk available through subagents, and it is genuinely powerful for anyone comfortable writing configuration files. It is also entirely optional, and it is not what this guide is about. The four-seat desk in the web interface delivers most of the benefit for none of the setup, and a trader whose desk requires debugging has acquired a second job. Build the simple version, run it for a quarter, and only then decide whether the complexity buys you anything.

Grok or Claude?

An honest answer rather than a recommendation, because they fail in different directions.

Grok Bot gives you scheduled routines, bot-to-bot handoffs and agents with real computers that sign into tools. It is more capable and it asks more of your discipline, because those same capabilities are exactly what a trading desk must not use. The shared computer across a bot roster is a real consideration for anyone handling financial documents.

Claude Projects give you cleaner isolation, no browser, no filesystem, and a knowledge model that forces you to write things down. The cost is that you are the switchboard, and that anything you do not save is lost. For a trader, being made to write down your playbook and paste back your journal entries is not obviously a cost at all.

The four-slot constraint on grok.com and the fresh-chat constraint on Claude push you in the same direction: fewer seats, narrower jobs, better files. Both stacks work. Pick the one you already pay for, build four seats, and judge it after thirty days of outputs you actually read.

Frequently asked questions

Do chats inside a Claude Project share context with each other?

Not automatically. A new chat starts fresh and does not inherit the last one. Three things can bridge the gap: the Project’s custom instructions, its knowledge files, and past-chat search with the memory built from it, if that setting is enabled. The distinction that matters for a trading desk is precision. Instructions and files are read verbatim. Retrieved history frequently comes back as a summary of a session rather than the session itself, and attachments do not survive at all, so the CSV you uploaded three weeks ago cannot be reopened. Continuity, yes. A source of record, no. That is why every seat above ends by handing you something to save.

Can I schedule a morning brief in Claude?

Yes, through scheduled tasks in Cowork, which is a paid feature and now runs in the cloud rather than requiring your laptop to be open. Verify the current setup path on your own build. If scheduling is not available to you, the brief becomes a saved chat you open each morning, which works perfectly well and requires you to show up, which is arguably the point.

How many Projects do I need?

Four, one per live seat, with the two bench roles as Skills. If your plan caps you lower, use the single-Project layout with a routing instruction. Four narrow seats you read beats nine you admire from a distance, on either stack.

Is Claude safer than Grok for financial documents?

Structurally, in one specific way: Claude Projects have no browser, no terminal and no shared filesystem, so a Project cannot sign into anything on your behalf. There is a second, subtler difference worth knowing. A statement uploaded as a chat attachment is not retrievable by a later chat, whereas a statement added to project knowledge is readable by every chat in that Project for as long as it sits there. Convenience and exposure move together. Put statements in the Tax Project deliberately, and take them out when the year closes. That removes an entire category of risk that a Grok Bot roster carries by design. It does not make an uploaded broker statement private. Strip account numbers, keep statements in a dedicated Project, and read your provider’s retention and training policy rather than assuming what it says.

Where to go next

The sibling guide, building the same four-seat desk out of Grok bots, carries the longer version of the doctrine: the full instruction blocks for all six seats, the failure modes, and the credential warning in detail. Read it even if you never open Grok.

For the wider picture, our complete guide to AI for traders sets out what these models can and cannot do for a discretionary trader, and the best AI trading tools in 2026 separates the working tools from the marketing. Single-chat workflows that need no setup are covered in five ChatGPT coaching workflows.

Two seats have standalone versions worth doing by hand once before you automate them: building a personal trading edge report and building a pre-market dashboard from free tools.

And if the files this desk runs on do not exist yet, start there. The Mind, Method and Money assessment will tell you which pillar is thinnest, and the position sizing and risk calculators produce the numbers your Risk Auditor will be auditing against. A desk enforces the rules you wrote. It cannot write them for you, and you would not want it to.