How to Write a System Prompt That Keeps Claude On-Brand

Most Claude Outputs Sound Generic Because the Setup Is Generic

Eighty percent of Claude users never touch the system prompt. They type a question, get an answer, and wonder why it doesn't quite sound like them. That's not Claude's fault. That's a configuration problem.

The system prompt is the most powerful lever you have. It runs silently before every conversation. It shapes tone, sets limits, and tells Claude who it's supposed to be. If yours is blank — or a single vague sentence — you're leaving serious quality on the table.

Here's how to write one that actually works.

Start With a Role, Not a Request

Most people treat the system prompt like a first message. Don't. It's an identity brief, not a task description.

The first thing Claude needs is a clear role. Not "you are a helpful assistant." Something like: "You are a sharp, no-fluff copywriter for a DTC skincare brand targeting women 28-42. You write with warmth and confidence, never with hype or hollow buzzwords."

Specific beats vague. Every time.

Then add what Anthropic calls a "values layer" — a short list of non-negotiables. What should Claude never do? Never use passive voice? Never suggest discounts unprompted? Never break the fourth wall? Write those down. Three to five hard rules, maximum.

The Four Components Every Strong System Prompt Needs

There's no official template from Anthropic, but after working with the Claude API across a dozen client projects, four components consistently produce tight, on-brand outputs.

1. Role definition. Who is Claude in this context? Give it a job title, an audience, and a personality in two to three sentences. The more specific, the better the output.
2. Tone and style rules. List three to six adjectives that describe your voice. Then give one or two concrete examples: "Write like a knowledgeable friend, not a corporate manual. Short paragraphs. Active verbs."
3. Hard constraints. What is off-limits? Competitor mentions, legal disclaimers, pricing, certain topics — write them out explicitly. Claude respects clearly stated limits well.
4. Output format defaults. Should responses always end with a call-to-action? Should they use bullet lists or paragraphs? Should headers be used at all? Tell Claude upfront so you don't have to correct it after every run.

That's it. Four components, written cleanly, will outperform a three-page prompt stuffed with edge cases.

Tone Calibration: The Part Most People Skip

Defining a role gets you 60% of the way there. Tone calibration gets you the rest.

Claude is naturally diplomatic and thorough. If your brand is punchy and direct, you need to explicitly push against those defaults. Try adding a line like: "Prioritize brevity over completeness. If a sentence can be cut, cut it."

Contrast words help too. Instead of saying "be casual," say "be casual, not sloppy." Instead of "be professional," say "be professional, not stiff." The contrast gives Claude a range to work within rather than a single ambiguous target.

Testing and Iterating Without Wasting Hours

Write your first draft. Then run the same five prompts through it that represent your most common use cases. Look for three things: tone consistency, format compliance, and any place Claude added something you didn't want.

If Claude goes off-script, don't assume the instruction failed. Check whether the instruction was actually present or just implied. Implied rules get ignored. Explicit rules get followed.