Claude Projects: Build Your AI Assistant in 5 Steps

You open Claude. You paste the same brief you pasted yesterday. And the day before. This is week three.

Every session resets. You re-explain your business to your assistant like it's Monday morning at a temp agency. You've been using Claude Pro for months, and it still greets you like a stranger.

I run four marketing assistants that manage my content on about an hour a week. Their memory doesn't live in the chat window. It lives in Projects.

Here's the 5-step method.

Why building an AI assistant feels harder than it should

The magic prompt doesn't exist. What's missing is persistent memory. And clear guardrails.

You copy-pasted your brand voice guide twice this week. Once for a caption. Once for an email. You'll do it again tomorrow. In six months you'll still be pasting the same 400-word brand voice intro into a fresh chat, still opening twelve tabs, still telling yourself Claude "just doesn't get it yet."

The problem is upstream. You never gave the assistant a place to remember.

A Claude Project is a folder with three things the chat window can't hold: instructions that persist, files it always sees, and a purpose narrow enough to be useful.

Claude Projects: 5 steps to build the assistant you'll actually use

1. Name one job. Not five.

The first assistant I built tried to write my emails, plan my week, manage my content, review my launches, and coach me on pricing. It did all of it badly.

Now every Project has one job. "Writes captions in Flavie's voice." "Drafts client onboarding emails." "Turns Fathom transcripts into show notes." One job. One reader. One output.

Sign it worked: you can describe the assistant in a single sentence to a friend without stumbling.

2. Feed it your real voice.

Custom instructions are where most Projects go wrong. People paste a personality prompt from Twitter. The assistant becomes a caricature.

Feed it real material instead. Ten captions you actually wrote and would post again. Three emails from your inbox that landed. One transcript of you speaking about your work. Not five hundred pages. Ten to fifteen pieces of your real voice, saved as a text file inside the Project.

Then in the custom instructions, name what the reader should never hear from you: banned words, banned structures, forbidden clichés. Boundaries make the voice.

3. Load 5 to 7 files. No more.

Every file you add is a slot in Claude's working memory. Load ten and it starts averaging. Load twenty and it starts fabricating.

Five to seven is the sweet spot. My content assistant sees six files: brand voice notes, ideal client profile, offer stack, content pillars, a swipe file of my best posts, one file of banned words. That's it. Everything else lives outside the Project and gets referenced when needed.

Sign it worked: the assistant refers to your actual client by name and your actual offer by price without you copy-pasting them into every prompt.

4. Set the guardrails.

Guardrails are the never-dos. Most custom instructions overweight the shoulds and underweight the never-dos, and that's why the assistant keeps drifting.

Mine include: never invent client quotes. Never use em-dashes. Never say "solution" or "unlock." Never write in the second person plural. Never confirm a claim without a source in the loaded files. If a fact isn't in the Project, flag it as a guess.

The assistant becomes trustworthy the moment it starts saying "I'm guessing here" instead of confabulating.

5. Test it on one real task from yesterday.

The instinct is to build the assistant on Monday and start using it next Monday. That gap kills it.

Give it a task from your yesterday. A caption you wrote at 11pm. An email reply you drafted three times. A show notes doc you started and abandoned. Let it do the task. Compare its output to yours.

If it's 70% of your voice on the first try, you have an assistant. If it's 30%, you have a prompt. Go back to Step 2.

✦ The point isn't more context. It's removing the context that confuses it. A messy Project is a bad assistant with better memory.

The prompt: copy and paste

Here's the custom instructions block I use for a voice-assistant Project. Fill the brackets with your actual business. Drop it into "Custom Instructions" in the Project settings.

ROLE
You're the voice assistant for [your business], a [description] serving [ICA in one sentence].

CONTEXT
- Business: [what you sell, one line]
- ICA: [who you sell to, one line, not "everyone"]
- Non-negotiables in tone: [three to five words: warm, direct, specific, no-fluff]
- Banned words: [list. Mine includes "solution," "unlock," "aligned," "level up," "authentic," "empowered"]
- Banned structures: em-dashes, negation-pivot binaries ("It's not X, it's Y"), fake studies, made-up client quotes

OUTPUT
- Match the voice files loaded in this Project. Never write in a voice not present in those files.
- Use contractions. Real numbers only. One idea per line.
- If you're guessing on a fact, say "I'm guessing here" before the sentence.

FORMAT
- Return the piece requested. No preamble. No sign-off. No "Here's your caption:" wrapper.
- If the request is ambiguous, ask one clarifying question before writing.

Lazy version, for comparison: "Write like me." That produces a caricature every time.

The point of the assistant

You still open Claude every morning. You just don't re-explain your business. Your Project already knows your ICA, your voice, your banned words, and the last five pieces you published that worked.

You save the twenty minutes of re-context you were spending. You use them to actually run your business.

Frequently Asked Questions

How is a Claude Project different from a regular Claude conversation?

A conversation resets when you close the tab. A Project keeps custom instructions and loaded files across every session. The assistant reads them at the start of every message, so you don't repeat yourself.

Do I need Claude Pro to use Projects?

Yes. Projects are a Claude Pro and Team feature. The free tier gives you the chat window, without persistent Projects.

How many Projects should I run at once?

One per job. If you juggle content, client work, and admin, that's three Projects. If you try to run everything through one Project, you'll get averaged output every time.

Can I share a Project with a team member?

On Claude Team plans, yes. On Claude Pro, Projects are personal. For a solopreneur, that's a plus. You can build the Project around exactly your voice without diluting it for collaborators.

★ Key takeaway

You build an AI assistant so you stop being the only person who knows what your business needs. Saving time is the side effect. Claude Projects is the shortest path there. A folder, one job, a handful of real files, and boundaries the assistant respects.

Next step

Wednesday I'm running a Behind the Scenes. I show you how I use Claude and the four Marketing Assistants that run my content. Live. Free. Save your seat here HERE.