Build an AI Assistant in Claude or Cowork (One Evening)

You can have a working AI assistant tonight. One project, one instructions file, no connections, no code, nothing to maintain.

This works in a Claude Project or in Cowork, whichever you already open. (ChatGPT has its own version of the same idea.) Everything below is identical in either.

This is the version most people should build, and most people skip it because it sounds too simple to work. It works. The built version connects to your tools and runs on a schedule, and it's covered in the full build guide. Start here regardless.

What you're actually making

A project is a container that holds standing instructions and files. Every conversation you start inside it already knows everything in there, so you stop re-explaining yourself.

The instructions are the job description. That's the whole product. Everything else on this page is about getting a good one.

Step 1: Don't write the instructions

The blank instructions box is where most attempts die. You're being asked to write a job description for a coworker you have never met.

So don't. Make it interview you first.

I want to build you into a dedicated assistant for [one specific part of my life]. Before you write anything, interview me. Ask me one question at a time about what this assistant should handle, what it should never touch, what a good day looks like, what I keep forgetting, how I want you to talk to me, and what would make me stop using you. Wait for each answer before asking the next. When you've got enough, tell me.

Answer out loud if you can dictate. Talking gets you further than typing, and the useful answers arrive once you stop performing.

Step 2: Have it write the job description

Now write the system prompt for this assistant based on everything I just told you. Write it in my voice, not corporate. Include how it should talk to me, what it should do without asking, what it should never do, and what to do when it isn't sure. Then tell me what you left out and why.

The "what you left out" instruction surfaces the gaps: the situations you never brought up, where it has no guidance and would just guess. You want to see those now, while you can still answer them.

Then edit one thing yourself: the tone. Tone is what the model gets slightly wrong every time and it's what determines whether you keep using this. Default AI voice is cautious, over-explaining, and faintly apologetic.

A tone line that works:

Talk to me like a competent chief of staff. Don't ask permission for obvious things. Don't add disclaimers. If you're not sure, propose a path and tell me what you'll do if I don't reply.

Paste the finished instructions into the project.

What a finished instructions file looks like

So you know what you're aiming for, here's the shape of a complete one. This example is for a client-work assistant; yours will differ, but every section here should exist in some form:

You're my assistant for client work.

TONE: Talk to me like a competent chief of staff. Don't ask permission
for obvious things. Don't add disclaimers. If you're not sure, propose
a path and tell me what you'll do if I don't reply.

WHAT YOU HANDLE: Drafting client emails and follow-ups. Prepping me
before calls using the notes in this project. Turning my voice memos
into organized next steps. Tracking open loops per client and surfacing
anything that's gone quiet for more than a week.

WHAT YOU NEVER TOUCH: Pricing decisions. Anything legal. Never send or
commit to anything on my behalf; everything you write is a draft for me.

HOW TO WORK: Short over long. Bullets over paragraphs. When I give you
a messy brain dump, organize it and ask me at most two questions. If a
task doesn't fit anything above, say so instead of improvising.

WHAT GOOD LOOKS LIKE: I should be able to read your output once and act
on it. If I have to re-read it, it was too clever.

Notice the last section. Telling it what "good" means is the instruction people never think to write, and it does more than any other paragraph.

Step 3: Load the files it needs

This is the step that separates an assistant that feels like it works here from one that feels generic.

Drop in whatever it needs standing access to. For a business assistant: how your week runs, who does what, your current projects, your recurring commitments, the things you always forget. For a health one: your training history, what you can't eat, what you actually cook. For anything: the vocabulary of your world, so it stops asking what things mean.

Plain text or markdown files work best. If you'd have to explain something twice, it belongs in the project.

Let it tell you what is missing:

Look at everything in this project. What do you still not know about me or my work that would make you better at this job? Ask me for it one category at a time.

Step 4: Give it a standing routine

The built version runs on a schedule. This version doesn't, and you can get most of the benefit with a habit instead.

Pick one moment a day and one prompt you paste every time. (In Cowork you can go a step further and set it up as a scheduled task, which is the first taste of the built version.) A morning briefing is the classic:

Here's today: [paste your calendar and anything on your plate]. Flag anything complex or back-to-back, tell me where my deep work block should go, what's urgent, what's left, and what's blocked. And remind me if there's no space to eat lunch.

Pasting takes twenty seconds. It's most of what the automated version produces, and you can decide later whether the twenty seconds is worth automating away.

Step 5: Correct it out loud for a week

Have it draft things you don't send. Let it propose plans you don't follow.

When it gets something wrong, tell it in the moment and in plain language: "too formal," "you keep suggesting meetings, I hate meetings," "stop asking me, just pick one." Then add the correction to the instructions file, because a correction in a chat evaporates and a correction in the instructions is permanent.

That loop is the entire difference between an assistant and a novelty you stop opening.

When to graduate

Move to the built version when one of these is true:

You're pasting the same context every day and it has become a chore
You want it to run without you opening anything
It needs to act on your behalf rather than tell you what to do
You're managing three of these and the pasting has multiplied

Until then, this version is the correct amount of machinery for the job.

Three assistants worth building this way

If you want a worked example instead of a blank slate, I've written up the full build for each of the three I run:

A personal chief of staff, the one that sends a 7am briefing and pre-triages the day
A nutrition assistant, the one that plans meals around your actual calendar and keeps the running log
An astrology companion, the one with no job, which is exactly why it works

Each of those pages includes the context to load and the standing instructions, so the interview in step 1 gets you to a working version faster.

Common mistakes

Writing the instructions yourself first. The interview produces something better in ten minutes.
Leaving the default tone. It's the single highest-impact edit and it takes one paragraph.
Building three at once. Get one reliable first. Three half-working assistants is the standard outcome of enthusiasm.
Never updating the instructions. If you have corrected it three times about the same thing, that belongs in the file.
Making it general. "My assistant" fails. "My assistant for client work" succeeds. Narrow scope is why this works at all.

FAQ

Do I need a paid plan? For projects with standing instructions and files, generally yes. The free tier won't hold context between conversations, which is the whole mechanism.

Claude or ChatGPT? Both do this. Use whichever you already have. The instructions transfer between them with almost no editing.

How long should the instructions be? Mine run several hundred words. Long enough to cover tone, scope, and the things it must never do. If it reads like a policy document, it's too long.

Can I have more than one? Yes, and keep them separate. One assistant per job. A single assistant holding your work, your health, and your personal life starts bleeding between registers and gets worse at all three.