How to build a Claude Project for your role in one sitting

One sitting of about an hour gives you a Claude Project that holds your function's standing documents and runs one recurring task. You pick the task first, load four or five files, write the instructions once, then run three jobs through it and check the output against something you already know. Here is the order.

How do I start with AI?

Start with one task you already do every month, not with a plan. Most answers to "how to start with AI" hand over a week-by-week curriculum. A curriculum has no owner and no deadline, so it stops in week two. A monthly task has both.

Say you run marketing for India at a personal care business. The recurring task is the monthly business review: the deck you build from sales data, what moved, and what you need the board to decide. Pick that.

Then write down, in one line, what a good answer looks like. Movement by brand. Three reasons behind the movement. Two decisions you need from the board. That line is the only specification you will need all sitting.

I pick the task by asking which business problem on my desk today AI could help me solve faster. The answer is the project you build. Everything else waits for the second sitting.

What goes into project knowledge, and what stays out?

Load the documents you would hand a new analyst on their first day, and nothing beyond that. A Claude Project has two parts you fill in: a knowledge base you upload files to, and project instructions that apply to every chat inside the project.

For the monthly review, four to six files is enough. This year's brand plan. The last three review decks. The category and channel definitions your team argues about. The template the board expects.

Leave out anything carrying customer or employee personal data. IT owns the AI policy and nobody has told you where the line sits, so put the line where you can defend it in a meeting. No names. No field that identifies one person. No salary data. Strip those columns before you upload.

Leave out volume as well. On paid plans a large knowledge base gets searched rather than read whole, which keeps a big project working. But model accuracy on very long inputs is uneven, and researchers now have a name for that problem, context rot. So curate. Six right documents beat sixty. Done looks like four to six files in the knowledge base, none of them holding a column that identifies a person. Check by opening each uploaded file and reading the first ten rows.

Write the project instructions as a standing brief

Project instructions are where your judgment goes, so write them once and write them plainly. The shape that holds up is role, context, task, format, constraints.

Role: you are working for the person who runs marketing for India at a personal care business. Context: the brand plan and the last three reviews are in the knowledge base, and the definitions file governs every category name. Task: read the sales file I paste, report movement by brand against plan, and name the reasons you can support from the knowledge base.

Format: one page, a table of brand-level movement, then three short paragraphs, then the two decisions needed. Constraints: flag any number you cannot trace to an uploaded file. Never estimate a figure. Ask before assuming a definition.

That last constraint is the one that earns its place. It turns a confident draft into a draft with its weak points marked. Done looks like five short labeled lines a colleague could read in under a minute. Check them by asking the project, in a new chat, what these instructions would refuse to answer and what it would flag.

Run three tasks through the project before you trust it

Run the task you already know the answer to through the new Claude Project first. Take last quarter's review, paste the sales file, and compare what comes back with the deck you actually presented. Done looks like a one-page answer in your format. Check whether the brand movement matches and whether the reasons are ones you recognize.

Second, run a task that needs two documents joined. Ask for the three categories where the plan assumed growth and the last three reviews show none. Done looks like a short list with the plan page and the review month named for each. Check the page references by opening the files.

Third, run this month cold. Paste the current sales file and ask for the review. Done looks like the page you would have spent an afternoon building. Check the flagged numbers first, because those are the ones the instructions could not trace.

How do I get better with AI after the first project?

You get better by checking output you could not have produced yourself, and by finding the place it breaks. Pick three claims from the cold run and trace each one back to a file. Note which traced cleanly, which traced loosely, and which did not trace at all.

When the output is wrong, and part of it will be, do not rewrite the sentence. That teaches you nothing and the same error returns next month. Find the cause instead. A definition missing from the knowledge base. A review deck you forgot to upload. An instruction that allowed an estimate. Fix that, then run the same task again and see if the error is gone.

Two or three rounds of this and the project stops being an experiment. You will also start reading every AI output the same way, including the ones your juniors bring you. That is the skill the board is asking you for when it asks for your view on AI.

Module 2 of A.D.A.P.T. AI Coaching, Inside Claude, includes this build, a Claude Project for your role.

What else comes up when you build your first Claude Project?

Do I need a paid Claude plan to build this?

Projects are available on every Claude account, including free ones, where you can keep up to five. A paid plan is not the gate. One function's standing documents, four to six files, fit inside a single project either way, so build it on the account you already have.

How is this different from uploading a file into a chat?

A chat's uploaded file lasts one conversation, and your instructions have to be retyped every time. Project knowledge and project instructions apply to every chat inside the project, so the standing context is typed once and the work starts at the task instead of the setup.

When does this start saving me time?

Not on the first run. The first run is where you find the missing file and the loose definition. By the third run of the same task the setup is stable and the draft arrives in minutes instead of an afternoon. Your review time stays, and should.