Back to guides

How Do I Build My First Claude Project That Actually Knows My Business?

Jake McCluskeyBeginner25 min read
How Do I Build My First Claude Project That Actually Knows My Business?

I spent about six months using Claude as a smart notepad. Every new chat, I'd paste my company description, my voice guide, my current offer, my latest case studies. Every. Single. Chat. Then I discovered Claude Projects and felt like an idiot. One afternoon of setup replaced roughly a hundred copy-paste rituals and — more importantly — gave me a Claude that sounds like my business, not like a well-meaning intern who just walked in.

This is the walkthrough I wish I'd had. Twenty-five minutes, start to finish, and you come out with a Claude that actually knows what you do, who you talk to, and how you want to sound.

Why this matters

Claude Projects are persistent workspaces inside Claude.ai. Each Project gets its own custom instructions (what Claude should always know about you), its own knowledge base (files it can reference in every chat), and its own chat history (separate from your other work).

Without Projects, Claude is a brilliant contractor you hire fresh every morning. With Projects, Claude is a staffer who already knows the playbook.

For consultants, solopreneurs, and small teams, the ROI shows up inside a week. You stop re-briefing. You stop drifting off-voice. You stop explaining the offer. Good Claude work becomes the default instead of the exception.

Before you start

You need:

  • A Claude.ai account. Free works; Pro is better because Projects on the free plan cap knowledge-base size tight. Pro is $20/month.
  • About 25 minutes. Ten of that is gathering files you already have.
  • Three things to have in front of you: your brand voice guide (or a sample of writing in your voice), your current offer or services list, and two or three of your best case studies or testimonials.

If you don't have a brand voice guide, that's fine. I'll show you how to generate one from sample writing in Step 2.

Step 1: Create the Project

Sign into claude.ai. In the left sidebar, click Projects. Top right, click Create Project.

Name it something you'll use daily — Elite AI Advantage or Acme Marketing — not Test or Claude Project 1. The name shows up in your sidebar and the URL.

Add a short description. One line. "Brand content, client work, and SEO for Acme Marketing." This isn't just UI cosmetics — Claude reads this description when it picks up context in a new chat.

Step 2: Write custom instructions that survive model updates

Click Edit system prompt (sometimes labeled Custom instructions depending on the Claude UI version). This is the single most important field in the whole Project.

Don't paste a marketing bio. Write instructions for Claude. Here's the template I use, with my own values swapped in:

text
You are supporting Jake McCluskey, owner of Elite AI Advantage.

About the business:
- We build custom AI-powered web tools for small and medium businesses.
- Our audience is owner-operators, marketers, and technical founders.
- We are NOT an enterprise consultancy. Keep examples grounded in
  teams of 1-20, not 10,000.

Voice rules (non-negotiable):
- First person singular when Jake is talking.
- Contractions are fine. No jargon. No hype words ("revolutionary,"
  "game-changing," "synergy," "leverage"). If a phrase would sound
  like a LinkedIn post, rewrite it.
- Be specific. "Cut my Anthropic bill 75%" beats "significant
  savings."
- Admit what you don't know.

When I ask you to write content, default to:
- Short paragraphs, 1-3 sentences each.
- Concrete examples over abstract principles.
- Code blocks for any command or config I'll need to run.

When I ask you to research, default to:
- Primary sources over summaries.
- Cite links inline.
- Flag when a claim is opinion vs. documented fact.

When I ask for opinions, push back if I'm wrong. Don't agree by default.

Three things to notice. First, this is written TO Claude, not ABOUT my business. Second, it's rules and examples, not adjectives. Third, that last line — "push back if I'm wrong" — is the single most valuable sentence in there. Claude's default mode is agreement. You have to explicitly permit disagreement.

Hit save.

Step 3: Upload the knowledge base

Scroll to the Knowledge or Project files section. This is where you park reference material Claude can cite in any chat.

Upload three things to start:

  1. Your voice guide — a one-page doc with your tone rules, five example paragraphs of your best writing, and five phrases you'd never use.
  2. Your current offer / services sheet — a plain-language summary of what you sell, who it's for, what it costs, and what's in vs. out of scope.
  3. Two or three case studies — the actual write-ups, not the one-line bullet versions.

That's it for v1. Resist the urge to upload every file you've ever written. A tight, well-curated knowledge base beats a dump. Claude's retrieval degrades on bloat — 10 surgical files outperform 100 dumped files, every time.

Don't have a voice guide? Upload a folder of your best emails, blog posts, or sales pages and ask Claude (inside the Project, once it's saved):

text
Read my uploaded writing samples. Write a one-page voice guide
covering: tone, contractions, sentence length, jargon rules,
three example paragraphs in-voice, and five phrases I'd never
use. Call it VOICE.md and format it as Markdown.

Save what it produces as VOICE.md and upload it back to the knowledge base. Now Claude has a formal voice reference instead of having to re-learn from raw samples every time.

Step 4: Test the Project with a real task

New chat inside the Project. Ask something that would have gotten a generic answer before:

text
Write the opening paragraph for a blog post titled
"Why most small businesses waste their first AI budget."

Read the output. Does it sound like you? Does it cite the audience size you defined? Does it avoid hype words?

If anything's off, fix it in the custom instructions — don't fix it in the chat and move on. That drift is exactly what you set up the Project to prevent.

Good first edit signals:

  • Voice too formal? Add "Write like you're talking to a single person at a bar, not a room of 100" to instructions.
  • Too agreeable? Strengthen the "push back" rule with "If my premise is wrong, say so in the first sentence."
  • Ignoring your offer? Add "Every content draft should tie back to one of our three services (AI tools, SEO, automation) when it naturally fits."

Iterate on the instructions. Re-test. Two or three rounds and it's locked in.

Step 5: Share it with your team (if you have one)

If you're on Claude Team or Enterprise, the Share button lets you invite teammates into the Project. Their chats are separate, but the custom instructions and knowledge base are shared.

This is the single best onboarding artifact you can hand a new hire. "Here's the Project. Run your work through it. Ask Claude any business question before you ask me."

Verify it worked

Three quick tests:

  1. Voice test. Ask Claude to rewrite a generic paragraph "in our voice." Does it match?
  2. Offer test. Ask "what do we sell?" and check the answer matches your offer sheet exactly.
  3. Knowledge test. Ask "which of our case studies is most relevant for a SaaS client doing SEO?" — does it reference the actual case studies you uploaded, or does it hallucinate?

If all three pass, the Project is working.

Where this breaks

  • Uploading a 50MB PDF. Projects have a knowledge-base size cap. A big PDF can eat the whole budget and push out the files that actually matter. Convert PDFs to Markdown first (or upload the Markdown source if you have it), and split long files into focused chunks.
  • Custom instructions that are biography, not rules. "We're a leading agency founded in 2018 specializing in…" is useless. Claude already has your website. Rules that are actionable — "always cite sources, never use the word 'leverage'" — are what move the needle.
  • Treating the Project as read-only. Revisit it monthly. Offers change, voice evolves, case studies get added. If your Project matches the business you ran a year ago, it's drifting you backwards.
  • Using one giant Project for everything. One per distinct client or domain. Mixing "my SaaS product" and "my freelance copywriting" into one Project produces schizophrenic output. Split them.
  • Relying on Projects alone for security. Project chats aren't private from Anthropic under the default data-retention settings. Flip on the "do not train on this data" toggle in Settings → Privacy if you handle anything sensitive.

What to try next

  • Read How Do I Set Up Claude Code on macOS Without Breaking My Shell Config? — Projects live in the browser; Claude Code lives in your terminal, and the two together are where the real productivity lives.
  • Browse the blog for specific ways to use Projects (content ops, sales enablement, research workflows).
  • Want this set up for you, tuned to your voice, with knowledge-base files generated from your existing content? Let's talk.
Want this built for you instead?

Let's talk about your AI + SEO stack

If you'd rather skip the how-to and have it shipped for you, that's what I do. Start a conversation and we'll figure out the fastest path to results.

Let's Talk
Questions from readers

Frequently asked

Do I need a Claude Pro subscription to use Projects?

Free accounts get Projects, but the knowledge-base size cap is tight. Pro ($20/month) gives you enough room for a real voice guide, offer sheet, and a handful of case studies — which is all you need to start.

What's the single most important field in a Project?

The custom instructions. A good set of rules (written to Claude, not about your business) beats a huge knowledge base almost every time. If you only tune one thing, tune those.

Should I upload every document I have?

No. Ten surgical files outperform a hundred dumped files, every time. Claude's retrieval degrades when the knowledge base is bloated. Start with your voice guide, offer sheet, and two or three case studies.

How do I stop Claude from just agreeing with everything I say?

Explicitly permit disagreement in the custom instructions. A line like 'When I ask for opinions, push back if I'm wrong. Don't agree by default.' moves the needle immediately. Claude's default is conversational harmony; you have to override that.

Is it safe to upload proprietary info?

With the 'do not train on this data' toggle flipped in Settings → Privacy, content stays out of Anthropic's training. Still, treat Projects the way you'd treat any SaaS — don't upload secrets or anything you wouldn't want on a third-party server.