Claude Projects for Founders: Setups That Ship Real Work
How to set up a Claude Project so you stop re-explaining your business every session — with seven concrete project setups solo founders actually use.
Quick Answer
A Claude Project is a workspace that remembers. You set custom instructions once and upload a small knowledge base of files, and every chat you start inside that project inherits both — so you stop opening each session with three paragraphs re-explaining who you are and what you're building.
For a solo founder, that's the whole value. The bottleneck in working with AI isn't the model. It's that you re-brief it forty times a week.
The setup takes about fifteen minutes:
- Create the project from the Projects page in the Claude sidebar.
- Write custom instructions describing your business, your customer, and how you want Claude to respond.
- Upload three to five files — not thirty.
- Start scoped chats inside the project instead of in the general chat window.
The rest of this article is what to actually put in those fields, plus seven project setups that earn their keep.
Projects vs Claude Code — They Are Not the Same Thing
This trips people up constantly, so let's separate them.
Claude Projects live in claude.ai. They're conversational workspaces with persistent instructions and a file-based knowledge base. You use them for thinking, writing, planning, analysis, and research.
Claude Code is a terminal tool that reads and writes files in your actual codebase. You use it for building.
They solve different problems, and most founders need both. Projects hold the context about your business. Claude Code holds the context about your code. If you've been trying to run your customer research through a coding tool or asking claude.ai to edit your repo, that's the mismatch.
Worth saying early: a project is only as good as the business you point it at. If you're still deciding what to build, the startup ideas library is the better first stop — come back and set up the workspace once you've picked one.
If building is what you're after, what to build with Claude Code covers that side. This article is about the other one.
Step 1: Write Instructions That Constrain, Not Just Describe
The custom instructions field holds roughly 8,000 characters — about 2,000 words. Most people use 200 and wonder why the output feels generic.
Weak instructions describe. Strong instructions constrain.
Weak:
You are a helpful marketing assistant for my startup.
Strong:
I run a scheduling tool for independent physiotherapists in the UK. My customers are solo practitioners, not clinics. They are not technical and they are time-poor.
When I ask for copy, write at a grade-8 reading level, never use the words "seamless", "empower", or "solution", and lead with the specific hour of admin the feature removes.
When I ask for advice, give me one recommendation, not a list of options. If you think my premise is wrong, say so in the first sentence.
Never invent statistics. If you don't have a number, say you don't have a number.
The second one changes every response in the project. The first one changes almost nothing.
The highest-leverage lines are the negative ones — what not to do. Banned words, banned formats, banned hedging. Claude follows those reliably, and they're the difference between output you can use and output you have to rewrite.
Step 2: Upload Few Files, Not Many
A project can hold a lot of uploaded material — up to roughly 200K tokens of documents. That is not an instruction to fill it.
The failure mode is uploading your entire Google Drive and getting worse answers, because now Claude is weighing your abandoned 2024 pitch deck equally against this quarter's positioning.
Upload the small set that is actually true right now:
- One positioning document. Who it's for, what it does, what it costs, why it's different.
- Raw customer conversations. Transcripts or notes, unedited. This is the single most valuable file in most projects.
- Your best existing work. Two or three pieces of copy you'd be happy to see imitated.
- Constraints. Pricing, technical limits, things you've already tried that failed.
That's it. Four files beats forty, and it stays maintainable — which matters, because a knowledge base nobody updates quietly becomes a knowledge base full of lies.
If you don't have customer conversations to upload yet, that's the actual gap, and the Mom Test approach to customer interviews is the faster fix than any project setup.
Seven Project Setups Worth Building
1. The Customer Voice Project
Instructions: Answer only from the uploaded transcripts. Quote directly. If the transcripts don't cover something, say so rather than generalising.
Knowledge: Every customer call transcript, support email, and cancellation reason you have.
What you ask it: "What words do customers use for this problem?" "What did the three people who churned have in common?" "Which objection comes up most?"
This is the highest-value project on the list and almost nobody builds it. It turns scattered conversations into something you can query.
2. The Positioning Project
Instructions: You are a positioning critic. When I show you copy, identify the claim a competitor could make identically. Be blunt.
Knowledge: Your positioning doc, three competitor homepages saved as text, your own landing page.
What you ask it: "What here is undifferentiated?" "Rewrite this headline five ways, each committing to a narrower audience."
3. The Weekly Writing Project
Instructions: Your voice rules, banned words, structure preferences, typical length.
Knowledge: Your five best previous posts.
What you ask it: Drafts. Not final copy — drafts you edit. The uploaded samples are what make the drafts sound like you rather than like everyone.
4. The Idea Triage Project
Instructions: Score ideas against my constraints. I build alone, in weekends, with no ad budget. Reject anything requiring a sales team, a two-sided marketplace, or regulatory approval.
Knowledge: Your skills, your time budget, your audience if you have one.
What you ask it: Paste in an idea, get a verdict. It kills bad ideas fast, which is most of the value. Pair it with the startup ideas library when you want candidates to run through it.
5. The Pricing Project
Instructions: Argue for higher prices. Assume I am underpricing, because I am.
Knowledge: Competitor pricing pages, your costs, what customers said about budget.
What you ask it: "What would justify doubling this?" "Which of these features should be the paid tier?"
6. The Launch Project
Instructions: You are running the checklist. Ask me one question at a time and don't move on until I answer.
Knowledge: Your launch plan, previous launch retros, target platform norms.
What you ask it: Walk me through what's missing. The one-question-at-a-time constraint is what makes this work — it turns a vague plan into an interrogation.
7. The Support Reply Project
Instructions: Draft replies in my voice. Never promise a delivery date. Never apologise more than once per message.
Knowledge: Your FAQ, your refund policy, ten previous replies you were happy with.
What you ask it: Paste the incoming message, get a draft. Saves ten minutes each time, several times a day.
What Goes Wrong
You built one giant project. Instructions that try to cover writing, pricing, support, and code contradict each other. One project per job. They're free to create.
You uploaded and forgot. The knowledge base silently goes stale. Anything that changes — pricing, positioning, feature set — needs a re-upload when it changes, or the project confidently tells you last quarter's story.
Your instructions are all positive. Add the constraints. "Don't hedge." "Don't give me options." "Don't use the word 'leverage'." These do more work than any amount of role-playing preamble.
You're still starting chats in the general window. The project only applies to chats started inside it. This is the most common reason people conclude Projects "don't do anything".
Does This Actually Save Time?
Honestly assess it after two weeks. The setups that pay off are the ones where you were already repeating yourself — support replies, weekly writing, customer research. The ones that don't are the ones you built because they seemed clever.
Delete the ones you haven't opened in a fortnight. A projects sidebar with three live workspaces beats one with fifteen dead ones.
And if the reason you have nothing to set up a project about is that you haven't picked what you're building yet, the fix isn't a better workspace. Start with an idea that fits a weekend — the startup ideas library is filtered for exactly that shape.
FAQ
What is a Claude Project?
A workspace in claude.ai that bundles custom instructions and a set of uploaded files. Every chat started inside the project inherits both automatically, so you don't re-explain context each session.
How is a Claude Project different from Claude Code?
Projects are conversational workspaces in claude.ai for thinking, writing, and analysis. Claude Code is a terminal tool that reads and writes files in your codebase. Projects hold context about your business; Claude Code holds context about your code.
How many files should I upload to a Claude Project?
Three to five current, true documents beats dozens of stale ones. Prioritise raw customer conversation notes — they're the most useful and the most commonly missing.
Do I need a paid plan for Claude Projects?
Projects are part of Claude's paid tiers rather than the free tier. Check current plan details on claude.ai, since tier contents change.
What should go in custom instructions?
Your business, your specific customer, and — most importantly — constraints. Banned words, required formats, and "don't do X" lines change output far more than descriptive role-play.
Can I use one project for everything?
You can, but the instructions start contradicting each other. One project per job works better, and creating projects costs nothing.
TL;DR
A Claude Project is persistent context: custom instructions plus a small knowledge base, inherited by every chat inside it. Write instructions that constrain rather than describe — the negative rules do the heavy lifting. Upload four true files, not forty stale ones. Build one project per job: customer voice, positioning, writing, idea triage, pricing, launch, support. Delete anything you haven't opened in two weeks. And remember Projects and Claude Code solve different problems — one holds your business context, the other holds your code.
Sources: Claude Projects guide, Claude custom instructions guide