Most people use Claude like Google. New chat, explain who you are, get a mediocre output, repeat tomorrow.
The people getting real value from it did one thing differently: they spent ten minutes on setup. After that, every output starts with full context already baked in. No re-explaining. No re-training. Claude writes like you, prices like you, talks like you.
Here are the five steps. Ten minutes upfront.
The five steps
Create a project instead of starting a new chat every time.
Run an interview prompt so Claude knows exactly who you are and how you work.
Save the output as custom instructions inside your project.
Drop in five examples of your writing so Claude can mirror your voice.
Add behavior rules that tell Claude what to cut.
The details are below.
Step 1: Create a project
A project is a persistent workspace inside Claude Cowork. It carries instructions and files across every chat you open inside it, so you're not starting from scratch every session.
Open the Cowork desktop app, hit New Project, and give it a name that makes sense for what you do most: "Quote Drafts," "Sales Follow-Ups," "Weekly Reports." Whatever you're going to use it for the most is where you start.
Note where the project folder lives on your computer. You'll be dropping files into it in Step 4.
Step 2: Run the interview prompt
This is where Claude learns who you are. Open a fresh chat in any AI tool (Claude, ChatGPT, Gemini, all work) and paste the prompt below. Use the voice feature if you'd rather talk than type; it's faster and the answers come out more naturally.
You're going to interview me and use my answers to write a Claude project instruction set. The goal is that every chat inside my project already knows who I am, how I write, how I make decisions, and what my work looks like — without me re-explaining any of it.
Ask me one question at a time. Cover my role and industry, what I'll use the project for, how I like things written, and what Claude does that I hate. Wait for my full answer before moving to the next question. Stop when you have enough to write the instruction set.
Once you have enough, produce a clean instruction set I can paste directly into my project. Use my words where I gave specific language. Do not paraphrase into generic AI phrasing. Output it inside a single markdown code block.
After you produce the instruction block, stop. No commentary, no follow-up suggestions.Copy the full instruction block it outputs.
Step 3: Paste the instruction set into your project
Open your project and find the instructions field. In Claude Cowork, this is at the top of the project view. Paste the full block from Step 2 and save.
You can also paste a version into your global Claude settings if you want the same base context to carry across all your chats, not just this project. Both are worth doing. The project-level instructions are more specific; the global settings are a useful fallback.
Step 4: Drop in five examples of your writing
The instruction set tells Claude how to write. Examples show it what you actually look like on paper.
Go to the project folder you noted in Step 1 and drop in five documents. These could be follow-up emails you've sent, call recaps, job estimates, internal updates, client proposals; anything you'd be happy to use again. PDFs and Word docs both work.
Then open a new chat inside the project and run this prompt:
Review the writing samples I've added to this project folder. Identify the patterns in how I write: sentence length, structure, the way I open and close things, vocabulary I use consistently, what I leave out. Do not summarize the content of the documents. Analyze the style only.
Then add a new section to the project instructions called "Writing patterns from examples" that captures what you found. Write it as direct style rules, not observations. Save it to the project instructions.
Do not invent patterns that aren't there. If something is inconsistent across the samples, flag it as inconsistent rather than picking one version.After this runs, every draft Claude produces in this project is pulling from both the instruction set and your actual writing history.
Step 5: Add behavior rules
This is the step most people skip. It's also the one that does the most to cut bad outputs.
Behavior rules tell Claude what to stop doing. Add a new section to your project instructions called "Behavior rules" and fill it in with whatever Claude does that you routinely delete.
Common ones:
No bullet point summaries at the end of responses.
Do not add disclaimers or caveats unless I ask for them.
No phrases like "Great question" or "Certainly."
Do not pad the opening. Get to the point in the first sentence.
No em dashes.
If you're unsure about something, say so rather than filling in with a guess.
The more specific you are, the more useful the rule. "Don't ramble" is vague. "No sentence should exist just to transition to the next sentence" is something Claude can actually act on.
To add these, open a chat in your project and type:
Add the following to my project instructions under a new section called "Behavior rules":
[paste your list here]Do it this week
Pick the project that eats the most of your time right now. Run the interview prompt. Build the instruction set. Drop in five documents. Add your behavior rules.
The ten minutes compounds. Every output after is faster, closer, and sounds more like you.
Boots & Bots is a weekly newsletter for home service operators and sales managers. Every issue is one practical AI workflow you can run inside your business. No theory, no generic tips, just the actual setup. If you're not on the list, signup is at the top of the page.
— Liam