How to Use Claude Projects (2026): Stop Re-Explaining Your Business Every Morning

TL;DR: The Feature Is Boring. The Habit It Kills Is Expensive.

Every tutorial about Claude Projects walks you through the buttons. The buttons take four minutes to figure out. What nobody writes about is the thing Projects actually fixes: the tax you pay at the top of every AI conversation. You know the one — "we're a 12-person HVAC company in Austin, we don't do commercial, our tone is friendly but not cutesy, here's our pricing, here's what we did last time." Five minutes, every time, forever. I watch owners pay that tax forty times a month and then conclude AI isn't worth $20. The tool wasn't the problem. Starting from zero was.

What a Project Actually Is

Strip the marketing away and a Claude Project is three things bolted to a chat window:

• Project knowledge — documents Claude can see in every conversation inside that Project. Your offers, your voice, your policies, examples of work you consider good.
• Custom instructions — standing orders. Who it's working for, what it should assume, how it should answer, what it should never do.
• Contained chats — every conversation you start inside the Project inherits both, and stays grouped with the rest.

That's the whole feature. As of 2026 Anthropic put Projects on the free tier with a five-Project cap; paid plans (Pro at $20/month and up) remove the limit and give each Project a far larger knowledge base that Claude searches rather than holds all at once. Which matters less than people think — see the sizing section below.

The Twenty-Minute Setup

1. Pick one recurring job, not one client. "Writing customer follow-ups" is a good Project. "Acme Corp" is usually a bad one, because you'll open it twice.
2. Write the custom instructions first, before uploading anything. Three or four sentences: who you are, who the reader is, what good output looks like, and one thing to never do. That last one does more work than the rest combined — "never use the word leverage" is worth ten paragraphs of style guidance.
3. Upload only stable documents. Offers and prices, positioning, two or three examples of writing you'd be happy to sign. Not the whole Drive.
4. Test it cold. Open a new chat in the Project and ask it something you'd normally have to set up first. If you still have to explain yourself, your instructions are too vague — fix them now, while you remember what was missing.
5. Prune it monthly. A Project rots. Prices change, positioning shifts, and stale knowledge produces confidently wrong output, which is worse than no output.

What Belongs in Project Knowledge (and What Ruins It)

The single most common mistake is treating project knowledge like cloud storage. It isn't a backup drive; it's a briefing. Every document you add competes for attention with every other document.

Belongs: your offers and prices, your positioning in plain language, three examples of finished work you consider good, standard policies you repeat constantly, and the objections you answer every week.

Ruins it: anything that changes weekly, meeting-note dumps, forty PDFs added "for context," and anything confidential you wouldn't put in an email. A tight ten-page knowledge base beats a sprawling two-hundred-page one nearly every time, because you can actually predict what it will do.

Do You Need the Bigger Knowledge Base?

Probably not, and this is where the upgrade advice usually gets sold to you backwards. Paid Projects handle much more source material by switching to retrieval — searching your documents for relevant passages instead of keeping everything in view. That is genuinely useful if you're working against a real corpus: a contract library, a codebase, years of case files.

For most owners, the free five-Project cap and a lean knowledge base is enough to prove the habit works. Upgrade when you hit a wall during actual work twice in a week — not before. If you're weighing the paid seat generally, I've written the honest version: is Claude worth it, and which Claude model to use for business work.

The First Three Projects Worth Building

1. The job you personally do most. Whatever eats your Tuesdays. Proposals, quotes, scopes, client updates. Load two examples you were proud of and one you weren't, and say which is which — the contrast teaches more than either alone.

2. Customer-facing writing. Follow-ups, objection replies, the awkward email about a delay. Knowledge: your voice examples and your policies. Instructions: your actual tone, described honestly. If you're blunt, say blunt.

3. The recurring analysis you dread. Monthly numbers, pipeline review, the report nobody reads but everyone asks for. This is the one owners skip and the one that compounds, because it converts a two-hour chore into a fifteen-minute review.

Where Projects Stop Being Enough

Here's the honest ceiling, and it arrives faster than people expect. A Project makes you faster at a task. It does not make the task happen without you. You still open the tab, still paste the inputs, still send the output.

That's the real jump — from a context store you drive to systems that run whether or not you show up: intake that answers itself, follow-up that fires on its own, reporting that compiles overnight. Projects are the cheapest possible rehearsal for that, which is exactly why they're worth twenty minutes. Once three Projects are running and you can feel where the manual part still hurts, you've diagnosed your own automation roadmap for free.

If you want the done-with-you version of that jump, the CFSB Playbook ($397) walks the install. If the question is which workflows to automate in what order, that's the AI Readiness Assessment ($2,500 flat — two weeks, a written roadmap, fee credited against a build within 90 days). And if you just want to know whether this is worth your time at all, that's what a strategy call is for.

How to Use ChatGPT Projects (and How They Differ)

ChatGPT Projects work the same way at the level that matters: a folder that carries files and custom instructions into every chat you start inside it. If you're already paying for ChatGPT rather than Claude, the setup advice on this page transfers almost unchanged — pick one recurring job, write the instructions before uploading anything, keep the file set small, prune it monthly.

The differences are practical rather than philosophical. Claude Projects lean on a larger, retrieval-backed knowledge base on paid tiers, which suits a real document corpus. ChatGPT Projects sit closer to the rest of the ChatGPT surface, so they inherit whatever else you're already using there. Both are free to try with limits, and both cost $20/month at the entry paid tier — so the honest answer to "which should I set up" is usually "the one you already pay for." If you're genuinely choosing, that's a different question, and I've answered it directly: is Claude better than ChatGPT.

The failure mode is identical on both, and it's worth saying plainly: people build eleven Projects, use two, and conclude the feature doesn't work.

Team Projects, Briefly

On Claude Team ($20/seat/month billed annually or $25 monthly for teams of 2 to 150 as of September 2026, per claude.com/pricing — the old 5-seat minimum and $125/month floor are gone, so two people pay $40 a month on annual billing) and Enterprise, Projects can be shared, so everyone works from the same instructions and the same knowledge, with permissions separating who can edit from who can just use. The practical win isn't collaboration; it's consistency. Five people improvising five different prompts produce five different voices. One shared Project produces one.

The Micro-Action

Do this today, before you close the tab: open Claude, create one Project named after the job that ate the most of your last week, and write four sentences of custom instructions. Don't upload anything yet. Start one chat inside it and notice what you didn't have to explain. That gap — the five minutes you just stopped paying — multiplied by every conversation you'll have this month, is the entire argument.