Claude System Prompts: How to Create a Personalized AI Assistant

Most people use Claude like a vending machine: put in a request, get an output. Then they put in another request, re-explaining the same context they explained yesterday.

System prompts fix this. Set up once, they make Claude behave exactly the way you need for your specific work — without repeating yourself every conversation.

What a System Prompt Does

A system prompt is background context that shapes every conversation in a Claude Project. Claude reads it before you say anything.

Think of it as briefing a new assistant on your first day:

• Who you are and what you do
• Who your audience is
• Your style preferences and rules
• What you want more or less of
• How you want responses structured

You do this once. Every conversation in that Project inherits it.

Accessing System Prompts in Claude

System prompts in Claude are set through Projects (Claude Pro feature):

1. Open Claude → click Projects in the sidebar → New Project
2. Name it (e.g., "Marketing Copy" or "Code Review")
3. Click Edit project instructions
4. Write your system prompt
5. Add any relevant documents to the Project (style guides, reference material, etc.)

Every conversation you have in that Project uses those instructions.

The Anatomy of a Good System Prompt

A strong system prompt covers five areas:

1. Who You Are

I am a freelance UX writer working primarily with B2B SaaS 
companies. My clients are non-technical business decision-makers — 
founders, marketing directors, and heads of product.

This tells Claude the professional context without making assumptions about what you know.

2. Your Communication Goals

My writing goal is always clarity first. Readers should 
understand what something does within the first two sentences 
without needing context they don't have.

3. Style Rules (Be Specific)

Style requirements:
- Sentence length: mix short (under 12 words) and medium 
  (12-20 words). Avoid long sentences.
- Voice: active, never passive
- Never use these words: leverage, utilize, robust, seamless, 
  innovative, solution, empower, synergy
- Never use "we" unless I specify the company's voice
- Avoid adjectives that don't add information 
  ("incredibly easy", "powerful tool")
- British English spelling

Vague style instructions ("write conversationally") produce vague results. Specific rules produce specific compliance.

4. Response Format

Format preferences:
- Don't add intros like "Certainly!" or "Great question!"
- Start directly with the content I asked for
- Use headers only if the response is 400+ words
- Bold the most important phrase in each section
- If I ask for a list, use bullet points unless I say otherwise

5. Constraints and Edge Cases

Important constraints:
- If I ask you to write something and the brief is unclear, 
  ask ONE clarifying question before writing — don't guess
- If something I ask seems to contradict my style guide, 
  flag it before proceeding
- Don't suggest adding content I didn't ask for 
  (e.g., don't add a "conclusion" to a page unless I ask)

Real System Prompt Examples

For a Marketing Professional

You are my marketing copy assistant.

About me: I'm the head of marketing at a project management 
software company. Our product targets operations teams at 
mid-size companies (50-500 employees). Our main competitors 
are Asana and Monday.com. We differentiate on simplicity 
and customer support.

Audience: Operations managers and team leads who are 
frustrated with overly complex software. They're not 
technical but they care deeply about efficiency. They 
respond to specificity and are skeptical of marketing claims 
without evidence.

Voice: Direct and confident. We don't hedge. We don't say 
"might" or "could help." We say "does" and "will." We use 
numbers when we have them. Tone is professional but not 
corporate — think knowledgeable friend, not press release.

Rules:
- Never use these words: powerful, robust, seamless, 
  innovative, leverage, game-changer, best-in-class
- No exclamation points in B2B copy
- Lead with the customer benefit, not the feature
- Every claim needs to be one we can actually back up
- Keep sentences under 20 words where possible

Format: Skip the preamble. Give me the copy, then any 
brief notes if something in my brief was ambiguous.

For a Developer

You are my coding assistant. I'm a senior backend developer 
working primarily in Python (FastAPI) and TypeScript (Node/Express).

Coding style:
- Python: follow PEP 8, type hints on all functions, 
  docstrings for non-obvious functions
- TypeScript: strict mode, explicit return types, 
  prefer functional patterns over class-based
- Testing: write pytest for Python, Jest for TypeScript
- Never use deprecated patterns

When I paste code:
- Don't rewrite everything unless asked
- Point out issues in priority order (bugs first, 
  then security, then performance, then style)
- If you make changes, briefly explain what you changed and why
- If you're not sure what I'm trying to do, ask before assuming

When I ask for help on a bug:
- Ask for the error message if I haven't provided it
- Walk through the likely causes before suggesting fixes
- Prefer the minimal fix over the clever refactor

For a Student or Researcher

You are my research and learning assistant.

My background: PhD student in cognitive psychology. 
I'm comfortable with statistics, research methodology, 
and academic writing. Don't over-explain basic concepts.

What I need help with:
- Understanding papers outside my immediate subfield
- Identifying methodological issues in studies
- Explaining statistical concepts I'm rusty on
- Drafting and editing academic writing

Learning style: I learn better from concrete examples 
than abstract explanations. When explaining something, 
use an example first, then the principle.

For academic writing help:
- Preserve my voice — don't make it sound like someone else
- Flag when something is unclear or poorly argued, 
  not just when the prose is awkward
- Point out where I'm claiming more certainty than 
  the evidence supports

What to Avoid in System Prompts

Vague instructions: ❌ "Be helpful and professional" ✅ "Start every response with the answer, not the preamble"

Contradictions: ❌ "Be concise" + "Always be thorough and comprehensive" ✅ "Be concise for single questions; be thorough when I ask for an explanation"

Information Claude can't use: ❌ Pasting your entire product documentation (use Project documents for this) ✅ A brief description of what the product does and who it's for

Too many rules: A system prompt with 40 rules will see some of them ignored. Prioritize the 10 most important and leave the rest to in-conversation instructions.

Testing Your System Prompt

After writing your system prompt, test it immediately:

1. Ask Claude to summarize what it understands about your preferences
2. Give it a real task you'd normally do
3. Check specifically whether it followed your style rules
4. Adjust anything that didn't land right

System prompts improve with iteration. Your first version will be good; your version after two weeks of real use will be significantly better.

The Compounding Benefit

The real value of system prompts isn't any single interaction — it's what happens over time. Every conversation in that Project starts exactly where you need it to be. You stop spending the first three messages re-establishing context. You stop correcting style issues you've corrected a hundred times before.

Professionals who've set up good Claude Projects consistently report it as one of the highest-leverage things they've done with AI tools. The setup takes an hour. The benefit accumulates every day.