You don't need a developer account. You don't need to understand embeddings or RAG pipelines. You need three things: a system prompt, a knowledge base, and ten minutes.
This is how to build your first Claude project from scratch.
What Is a Claude Project?
A Claude Project is a persistent configuration inside Claude.ai that gives an AI a consistent identity, set of instructions, and access to documents. Every conversation inside that project inherits those settings automatically.
Think of it as: system prompt + files + conversation thread, bundled together.
You're not building software. You're configuring a collaborator.
Step 1: Open Projects
From Claude.ai, click Projects in the left sidebar. Then click New Project.
Give it a name. Something specific works better than something vague. "Customer Support Bot" is better than "My Bot". "Content Review Assistant" is better than "AI Helper".
The name is just for you — it sets your mental model for what this thing does.
Step 2: Write Your System Prompt
Click into Project Instructions. This is your system prompt — the foundational document that defines how Claude behaves inside this project.
A minimal but effective system prompt has four parts:
Identity — Who is this agent?
You are a customer support assistant for Meridian Software.
Your name is Max.
Purpose — What does it do?
Your job is to help customers troubleshoot issues, understand billing,
and navigate product features.
Tone and style — How does it communicate?
Be concise and friendly. Avoid jargon. If you don't know the answer,
say so and offer to escalate.
Boundaries — What should it not do?
Do not make promises about refunds or SLA commitments.
Do not discuss competitor products.
That's it. Four sections, ten sentences. You can always expand later.
Step 3: Add Your Knowledge Base
Click Add Content (or the document icon, depending on your interface).
Upload files that contain information your assistant needs. Good candidates:
- Product documentation (PDF or plain text)
- FAQ documents
- Internal guides or runbooks
- Pricing sheets
- Tone and brand guidelines
Keep files focused. A 500-word FAQ outperforms a 50-page PDF where 90% of the content is irrelevant. Claude reads what you upload — garbage in, garbage out.
Name your files clearly. refund-policy-2026.pdf is more useful than doc1.pdf. Claude can reference file names when explaining where information came from.
Start with two or three files. You can always add more.
Step 4: Test the System Prompt
Open a new conversation inside your project. Don't start with edge cases — start with the normal use case.
If you built a customer support bot, ask a normal support question:
"I can't find where to update my billing address."
Read the response. Ask yourself:
- Did it stay in character?
- Was the answer accurate based on your KB?
- Did it match the tone you asked for?
Then probe the edges:
"Can you promise me a refund?"
Check that it respects your boundaries. If it doesn't, your system prompt needs tightening.
Step 5: Iterate Once
Resist the urge to over-engineer on the first pass. Make one improvement based on what you observed.
Common first-round fixes:
- Too formal → Add "Use a conversational, direct tone" to your style section
- Too long-winded → Add "Keep responses under 150 words unless the question requires detail"
- Ignoring KB → Rewrite docs to be more explicit and structured; bullet points beat paragraphs
Make the change, run the test again. Ship it.
What You've Built
By now you have:
- A named agent with a defined identity
- A system prompt that sets behavior
- A knowledge base grounded in real content
- A tested, working configuration
That's a Claude Project. It's not magic — it's structure.
The best projects aren't built in one session. They're iterated on every time you notice a gap. Every bad response is a system prompt bug waiting to be fixed.
Start small. Build something useful. Improve it when it fails.
That's the whole loop.
Comments 0