How to Set Up Claude System Prompts for Your Agency

Most people set up a Claude system prompt once, get mediocre results, and blame the AI. I did the same thing in early 2023. I typed a two-sentence instruction, watched Claude produce generic property descriptions that sounded like every other listing on the internet, and almost went back to writing everything by hand. The problem wasn’t Claude. The problem was that I had no idea how system prompts actually work — and nobody had explained the setup in practical terms for someone running a real client-facing operation.

If you run an agency — or a one-person consulting business like mine — a well-built system prompt is the difference between an AI assistant that sounds like you and one that sounds like a chatbot. Here’s exactly how I do it, step by step, after three years of testing and refining in my real estate business in Madeira.

What a Claude System Prompt Actually Does (And Why Most Are Wrong)

A system prompt is the instruction set Claude reads before it sees your first message. It sets the persona, the constraints, the format rules, and the context. Think of it as the briefing you’d give a new hire before they start talking to your clients.

Most agency system prompts fail for the same three reasons: they’re too vague (“be helpful and professional”), they don’t include format instructions, and they have no context about who the end client actually is. Claude is a capable model — it needs specifics, not corporate-speak.

You can access system prompts in Claude’s API, in the Claude.ai Projects feature (which lets you set a persistent system prompt for a project), and inside any tool that connects to Claude — like Make.com workflows or custom GPT-style setups built on the API. All the steps below apply across those contexts.

Step 1: Define the Role Before You Write a Single Word

Before you open Claude, write down three things on paper or in a notes app: who is Claude playing, who is it talking to, and what is the single most important output quality you need.

For my real estate work, the answers were: a bilingual property consultant based in Madeira (Claude’s role), British and Northern European buyers relocating to Portugal (the audience), and accuracy over enthusiasm (the output quality). That 30-second exercise shaped every system prompt I’ve written since.

For an agency, your role definition might look like this: “You are a senior copywriter at [Agency Name], specializing in [niche]. You write for [client type]. Your tone is [adjective], [adjective], and never [adjective to avoid].”

Be specific about what to avoid. “Never use exclamation marks in body copy” is more useful than “maintain a professional tone.” The more concrete the constraint, the more consistent the output.

Step 2: Write the Context Block — This Is Where Most Agencies Skip

Claude performs significantly better when it understands the business context it’s operating in. This is the section most people skip because it feels like extra work. It’s not — it’s the highest-return 10 minutes you’ll spend on this setup.

Your context block should include:

• Business description: What the agency does, how long it’s been operating, core services
• Client profile: Who the end clients are — industry, typical concerns, level of sophistication
• Competitors or alternatives Claude should be aware of (so it doesn’t recommend them)
• Geographic or regulatory context if relevant (critical for my real estate work — Portuguese property law, NHR tax rules, Madeira’s specific market conditions)

Here’s a shortened example from my own property description prompt:

You are a property consultant assistant working for a boutique real estate consultancy in Madeira, Portugal. The consultancy serves international buyers, primarily from the UK, Germany, and Scandinavia, looking for residential property between €300,000 and €1.5M. Madeira is an autonomous region with specific ownership rules and tax incentives. Never describe any property as "cheap" or use high-pressure sales language. Buyers are educated and do their own research.

That context block alone cut my editing time by roughly 60% compared to prompts without it, because Claude stopped producing output I had to remove.

Step 3: Set Format Rules Explicitly

Claude will default to whatever format it thinks works best. Sometimes that’s great. For agency work, where you need consistent output across dozens of deliverables or multiple team members, “whatever Claude thinks works” is a problem.

Your format block should specify:

• Output length (e.g., “property descriptions must be 120–160 words”)
• Structure (e.g., “always open with location, then property highlights, then lifestyle benefit, then a single call to action”)
• What to avoid formatting-wise (e.g., “never use bullet points in client-facing copy,” “no markdown in email outputs”)
• Language register (formal, conversational, technical)

One format instruction I added after a painful week of inconsistent outputs: “Do not use the phrase ‘nestled in’ in any property description.” You’d be surprised how often Claude reaches for that phrase. Now it never does.

Step 4: Add Output Examples Directly Into the Prompt

This is the single most underused technique in system prompt design. Claude responds extremely well to examples of the output you actually want. Not descriptions of what you want — actual samples.

In my property description prompt, I included two example listings I’d written myself — one for a Funchal apartment, one for a Calheta villa. I labeled them clearly:

EXAMPLE OUTPUT 1 (apartment, urban):
[my own 140-word description]

EXAMPLE OUTPUT 2 (villa, coastal):
[my own 145-word description]

All outputs should match the tone, structure, and length of these examples.

After adding examples, I ran 20 consecutive listings through the prompt. I edited fewer than 4 of them for tone. Before examples were included, I was editing 14 or 15 out of every 20. That’s a concrete improvement you can measure — and in an agency setting, it directly translates to fewer revision rounds with clients.

Step 5: Include Behavioral Guardrails for Client-Facing Work

This step matters especially for agencies where Claude’s outputs go directly to clients or get published without heavy review. Guardrails are explicit rules about what Claude must never do.

Mine include:

• “Never state specific investment returns or yield projections — this is regulated financial advice in Portugal.”
• “Do not quote specific legal timelines for property transfer — these vary by case.”
• “If asked a question outside your knowledge, say so clearly rather than speculating.”

For a marketing agency, your guardrails might include: “Never make comparative claims against named competitors,” “Do not include pricing unless the brief explicitly provides it,” or “Always flag when a claim needs a source before publishing.”

The key is writing these as direct instructions, not suggestions. “Avoid speculating” is weaker than “Do not speculate. If you are uncertain, write: ‘I’d recommend confirming this detail with the client.'”

Step 6: Test With 10 Real Inputs Before You Deploy

Never deploy a system prompt after one or two test runs. I made that mistake in my first six months and ended up with a lead follow-up sequence that used the word “absolutely” in every third sentence — because one test message happened to trigger that pattern and I missed it.

Run at least 10 realistic inputs through your prompt before it goes live. For an agency, that means 10 actual client briefs, 10 different property specs, 10 different email requests — whatever the prompt is designed to handle.

Track failure patterns: What does Claude get wrong consistently? Those consistent failures tell you what’s missing from the prompt. Add specificity to address each one. Then test again.

I keep a simple log — a notes file with the date, the prompt version, the failure pattern, and the fix I made. After two or three iterations, most prompts stabilize. If a prompt still has major failures after five iterations, the underlying role definition (Step 1) is usually the problem.

My Real-World Experience Running Claude Prompts in a One-Person Real Estate Operation

In January 2026, I had 17 listings to write descriptions for in a two-week window. A local developer had approached me to handle the English-language marketing for a new boutique complex in the Palheiro area of Madeira — a mix of one, two, and three-bedroom units at varying price points, each with slightly different views, finishes, and target buyer profiles. In previous years, that kind of volume would have taken me the better part of a week, spread across evenings and weekends.

I ran every listing through a Claude Project I’d set up specifically for this developer, with a system prompt that incorporated all six steps above. The context block included the developer’s brand guidelines, the complex’s USPs (Palheiro’s microclimate, proximity to the golf course, panoramic Atlantic views), and specific information about the target buyers — primarily retirees and semi-retirees from Northern Europe and the UK looking for a permanent or part-time residence.

The format rules required 150–180 words per description, a three-part structure (setting, property, lifestyle), and no use of the phrases “stunning views,” “dream home,” or any variation of “don’t miss this opportunity.” The example outputs I’d included were two descriptions I’d personally written and was happy with as reference points for tone.

Results: 17 descriptions written in 94 minutes total, including the time I spent reviewing and editing. I made substantive edits to 5 of them — mostly to add or adjust a specific detail the developer sent me mid-way through the batch. The other 12 went out essentially as produced. The developer signed off without a single revision request on those 12.

For context, my previous approach — writing each description from scratch with light AI assistance — took me around 25–30 minutes per listing for a project like this, where the properties were similar enough that I kept rewriting the same phrases. That would have been 7–8 hours minimum. I got it done in under 2 hours.

The system prompt I used for this project took me about 45 minutes to build properly, including the test runs. That 45 minutes paid back immediately on the first batch. I’ve since reused the same structure — with minor modifications — for three other developer projects. The time investment is front-loaded and the return compounds.

One genuine limitation I’ve run into: Claude does not handle highly technical or legally specific content well, even with detailed guardrails. When I tried to include a section in the system prompt for generating Portuguese property law summaries — basic explanations of CPCV contracts, IMT tax, and the habitation license requirements — the outputs were too often slightly wrong in ways that sounded plausible. I tested this over about two weeks and 30+ outputs before concluding that legal content is not where I trust this setup. I now use Claude for everything marketing and communications related, and I write legal explanations myself or have them reviewed by a lawyer before they go to clients.

System Prompt Structure at a Glance

Here’s a quick reference for how I structure every agency-level system prompt:

Section	What It Contains	Approximate Length
Role Definition	Who Claude is, who it’s serving, core quality priority	2–4 sentences
Context Block	Business background, client profile, geographic/regulatory context	100–200 words
Format Rules	Length, structure, language register, markdown rules	5–10 bullet points
Example Outputs	1–3 real examples of ideal output	Varies — use actual samples
Behavioral Guardrails	What Claude must never do or say	3–8 explicit rules

Pro Tips From 3 Years of Real Use

Version your prompts. Save each iteration with a date and a one-line note on what changed. I use a simple Notion table. When something breaks after an update, you need to be able to roll back.

Shorter is often better than longer — up to a point. I went through a phase of writing 800-word system prompts. They didn’t outperform well-written 300-word prompts. The quality of each instruction matters more than the quantity. Every line should earn its place.

Use Claude Projects for persistent context. If you’re using Claude.ai rather than the API, the Projects feature lets you save a system prompt that applies to every conversation in that project. I have separate projects for property descriptions, client emails, social content, and market reports. Each has its own tailored prompt.

Revisit your prompts every 90 days. Your business changes. Your clients change. A prompt you built in March may be producing subtly outdated outputs by June. I schedule a prompt review into my calendar every quarter — it takes 30 minutes and has caught two significant misalignments in the past year.

Summary: The 6-Step Setup That Actually Works

Setting up Claude system prompts for agency work isn’t complicated, but it requires more upfront thought than most people give it. Here’s the sequence that consistently produces good results:

1. Define the role clearly — who Claude is, who it’s serving, and what matters most
2. Write a real context block — business background, client profile, relevant constraints
3. Set explicit format rules — length, structure, language register, what to avoid
4. Include actual example outputs you’ve written yourself
5. Add behavioral guardrails for anything client-facing or regulated
6. Test with 10 real inputs and iterate before deploying

The 45 minutes you spend building this properly will save you hours every week. I can put an exact number on that for my own business: roughly 6 hours a month recovered, based on the difference between my old editing time and current editing time across all the content types I now run through Claude Projects.

That time doesn’t vanish — it goes back into client work, into prospecting, into the parts of the job that actually require a human. Which is where your time should be going anyway.

If you want to go deeper on how I use Claude’s Projects feature specifically for different client types, I’ve written a separate walkthrough on using Claude Artifacts for client deliverables — worth reading alongside this guide.