Building a custom Claude-powered writing assistant with system prompts

Why System Prompts Beat One-Off Instructions

Most people treat Claude like a search engine—ask a question, get an answer, move on. But if you write regularly—whether it's emails, documentation, content, or proposals—you're leaving serious productivity on the table.

Here's the problem: every time you write something, you're essentially re-explaining who you are, what you care about, and how you want your work to sound. You repeat yourself to Claude the same way you'd repeat yourself to a new colleague. That's inefficient.

System prompts are your solution. They're like hiring a personal editor who knows your style, your standards, and your audience before you even start typing. Unlike conversation history, system prompts persist across every interaction. They set the rules of engagement once, then silently enforce them everywhere.

The difference is dramatic. With system prompts, you get consistency. You get faster drafts. You get fewer rounds of edits because Claude already knows what "good" looks like in your context.

Anatomy of a Strong Writing Assistant System Prompt

A good system prompt has layers. It's not just "be a writing assistant." It's specific, it's multi-dimensional, and it accounts for how you actually work.

Here's the structure I recommend:

Identity and scope: Tell Claude exactly what role it's playing and what it won't do. "You are a business writing assistant specializing in clarity and concision for tech company communications. You do not produce marketing copy or social media content—only professional business writing."

Your voice and values: This is where most people fail. They're vague. "Write clearly" doesn't work. "Avoid jargon unless explaining enterprise software to CIOs" works. Be specific about what you sound like. Include examples if you can.

Standards and rules: These are your non-negotiables. Examples might be: "Sentences should rarely exceed 18 words. Use active voice except when the agent is unknown or irrelevant. Never use exclamation marks. Flag any claims that need attribution."

Audience and context: Who reads what you write? Are they experts or novices? Are they decision-makers or implementers? Claude needs this to calibrate tone and depth.

Output format preferences: Do you want headers? Bullet points? Word count targets? Specify it here.

Here's a concrete example system prompt:

You are a business writing assistant for a B2B SaaS company. Your job is to
help draft clear, confident writing for external communication: product
announcements, sales emails, technical documentation, and customer case studies.

VOICE AND STANDARDS:
- Your writing is conversational but authoritative. Think "knowledgeable colleague
  explaining something important, not a textbook."
- Sentence length target: 12-16 words on average. Flag sentences over 25 words.
- Use active voice by default. Passive voice only when the agent is unknown.
- Replace marketing jargon (synergy, leveraging, ecosystem) with precise language.
- Never use exclamation marks.
- Lead with the problem or outcome, not the features.

AUDIENCE:
- Primary: technical decision-makers at mid-market companies (50-500 people)
- Secondary: their managers and finance teams
- Assume they're smart but time-poor. Respect their attention.

PROCESS:
When the user provides a draft or brief:
1. Clarify the specific goal if it's ambiguous (announcement? persuasion? explanation?)
2. Suggest structural changes before wordsmithing
3. Flag weak claims that need support or attribution
4. Edit for clarity and concision
5. Offer 2-3 alternative versions of the opening paragraph

Be direct about what isn't working. Don't soften it with "consider..." or "you might..."

This isn't generic. It has teeth. Claude knows what you're trying to do and how you want to do it.

Customization Levels: Match Your Workflow

Not all writing is the same. You might need different modes.

Level 1: General custom assistant – One system prompt that handles 80% of your writing. Use it for email, documentation, casual writing. This is your workhorse.

Level 2: Role-specific variants – Multiple prompts for different contexts. One for customer-facing marketing, one for internal documentation, one for investor updates. Each emphasizes different values: persuasion vs. clarity vs. narrative.

Level 3: Audience-specific variants – If you write for very different audiences, create separate prompts. One for technical audiences (assume more context, use terminology), one for executives (lead with business impact), one for end-users (assume less context, use metaphors).

You don't need all three. Start with Level 1. Add Level 2 only if you find yourself constantly correcting Claude on tone or depth.

Building Your First System Prompt: A Practical Framework

Start here:

1. Write down your actual standards. Not aspirational ones. Actual. How long are your sentences? Do you use Oxford commas? How do you handle numbers—spell them out or numerals? What's your stance on contractions? The more specific, the better.

2. Collect examples. Find 2-3 pieces of your writing that feel "right." Writing that your audience responded well to, or that felt effortless. Save them.

3. Name your voice. What three words describe how you write? "Direct, conversational, evidence-based"? "Humble, curious, precise"? This matters more than it seems.

4. Define your constraints. What should Claude never do? Common ones: "Don't use first-person plural (we/our) unless referring to the company," "Always provide sources for statistics," "Flag any jargon from our specific domain that might need explaining."

5. Test it. Give your system prompt a problem you actually need to solve. A real email, a real document draft. Does Claude's output reduce your revision time? If you're still doing major edits, refine the prompt. Iterate.

Practical Integration

Where does this live?

If you use Claude directly on Claude.ai, system prompts don't work—yet. You'll need to build this into a workspace or integrate Claude via API.

Via Claude's API: You can set system prompts programmatically. Build a simple Python app or use a tool like Make or Zapier to add Claude to your workflow. The overhead is minimal.

Via Claude's web interface alternatives: Tools like Perplexity let you set custom system instructions. Not as powerful, but accessible.

DIY approach: Create a private Slack bot or integrate Claude into your writing app. Most writing tools now support Claude integration.

The friction of setup is worth it. You're buying back hours every month.

Common Mistakes to Avoid

Don't over-constrain. A 2,000-word system prompt defeats the purpose. Shoot for 300-500 words. Anything longer isn't actually getting read carefully by the model.

Don't assume Claude knows your context. "Write like you know our industry" won't work. You need to explain it.

Don't skip examples. A well-chosen example is worth 50 words of description. Show, don't tell.

Don't set it and forget it. Revisit your system prompt every few months. Your voice evolves, your audience might shift, your standards might sharpen. Keep it current.

The Real Win

The real win isn't speed. It's consistency. It's knowing that whether you're drafting an email at 8am or a proposal at 10pm, Claude's baseline understanding of your standards doesn't waiver.

You're not managing the AI. The AI is managing the friction.

Build this once. Use it everywhere.