How to Write a System Prompt in 10 Minutes

Most people using AI tools in their solo business are leaving 60% of the value on the table. Not because they chose the wrong tool. Because they never told the tool who it’s working for. I figured this out the hard way in early 2023, after months of getting generic, flat, unusable output from ChatGPT — output I had to rewrite almost entirely before it could go anywhere near a client. The fix wasn’t a better prompt. It was a system prompt. One well-built system prompt changed how I use every AI tool in my real estate operation, and it took me about 45 minutes to write.

This tutorial walks you through exactly how to build one for your solo business — step by step, with real examples from my actual setup in Madeira.

What You’ll Build by the End of This Tutorial

A complete, reusable system prompt you can paste into ChatGPT, Claude, or any other LLM to make it behave like a trained assistant who understands your business, your voice, your clients, and your rules. No more generic output. No more editing every single paragraph. The AI starts where you actually are, not where some generic template assumes you are.

This isn’t theoretical. Every section of this tutorial comes from the actual system prompt I use daily for my real estate consulting business. I’ll share the exact structure, the exact prompts I tested, and the one mistake I made that cost me two weeks of bad output before I caught it.

Prerequisites Before You Start

• An active account with ChatGPT (Plus or Team, $20–$30/month) or Claude Pro ($20/month) — both support persistent system prompts
• 30 minutes of uninterrupted time — this is thinking work, not clicking work
• A clear answer to: who are your clients, what do you produce for them, and how do you sound when you write?
• A text file or Notion page open to draft as you go

You do not need any coding knowledge. You do not need to understand how LLMs work under the hood. You need to know your business well enough to describe it to a very capable but completely uninformed assistant.

Step 1 — Define Your Business Identity in 5 Sentences

The first block of your system prompt tells the AI who you are and what you actually do. Not your LinkedIn bio. A functional description that gives the AI enough context to make good decisions on your behalf.

Write five sentences that answer: What is your business? Who do you serve? Where do you operate? What do you produce or deliver? What makes your approach different from competitors?

Here’s the exact language from my own system prompt:

You are a writing and research assistant for Robson Penassi, an independent real estate 
consultant based in Madeira, Portugal. He works exclusively as a solo operator — no team, 
no agency affiliation. His clients are primarily international buyers (UK, Germany, 
Scandinavia, Brazil) looking for residential properties in Madeira and Porto Santo. 
He writes all client-facing content himself, including property descriptions, market 
analysis reports, email sequences, and LinkedIn posts. His differentiator is hyper-local 
knowledge combined with honest, no-hype communication — he does not use superlatives 
or make promises he can't verify.

That’s it for block one. Under 100 words. Every sentence is doing real work — it tells the AI what to write, for whom, in what geography, and what tone to avoid.

Step 2 — Describe Your Voice with Concrete Rules, Not Adjectives

“Professional but friendly” is useless instruction. Every AI defaults to that. You need to give the tool behavioral rules — things it should do and things it must never do.

Format this as a list. Two columns: DO and NEVER DO. Write at least five items in each column based on things you actually notice yourself fixing when AI gives you bad output.

VOICE AND STYLE RULES:

DO:
- Write in first person when drafting content for Robson
- Use short paragraphs (2-3 sentences maximum)
- Include specific numbers, street references, or neighborhood details when provided
- Use American English spelling throughout
- Match the register of the content type: formal for market reports, 
  conversational for social posts

NEVER DO:
- Use the words: "delve", "leverage" (as a verb), "game-changer", "robust", 
  "cutting-edge", "furthermore", "it's worth noting"
- Open with a question directed at the reader in property descriptions
- Include generic phrases like "nestled in", "boasting", or "don't miss out"
- Use exclamation points in market reports
- Make claims about rental yields or investment returns without data provided by Robson

The banned word list is the single most impactful section of any system prompt I’ve written. AI tools have strong habits. If you don’t explicitly block certain patterns, they’ll show up every single time.

Step 3 — List Your Core Content Types and Their Specific Requirements

Your system prompt should cover every type of content you regularly produce. For each one, give the AI a template structure, a target length, and any non-negotiable elements.

I produce four content types consistently: property descriptions, market analysis reports, client email sequences, and LinkedIn posts. Here’s how I’ve specified property descriptions in my system prompt:

PROPERTY DESCRIPTION FORMAT:
- Length: 180–240 words
- Structure: Opening sentence with location and property type (no adjective overload) → 
  Key interior features (2-3 sentences) → Outdoor/view details if applicable → 
  Practical information (parking, storage, building condition) → 
  One closing sentence on lifestyle fit — never a call to action
- Always include: property type, location within Madeira, floor area in m², 
  number of bedrooms, number of bathrooms
- Never include: price (added separately), rental income projections 
  unless Robson provides verified data, vague phrases like "sea views" 
  without specifying from which room

Do this for each content type you produce. If you write proposals, invoices, social captions, or follow-up emails — give each one its own specification block. The more specific you are here, the less editing you’ll do later.

Step 4 — Add Client Context and Audience Assumptions

Your AI assistant needs to know who’s reading the output, not just who’s writing it. This step is where most solopreneur system prompts fall short — they focus entirely on the business owner‘s preferences and forget to encode the audience.

AUDIENCE CONTEXT:

Primary audience: International property buyers, typically 35–65 years old, 
with disposable income but significant skepticism about overseas real estate purchases. 
They have done some research online but rely on Robson's expertise to filter out 
hype and misinformation.

Secondary audience: LinkedIn connections in the real estate and expat communities — 
professionals who appreciate direct, data-backed content over promotional posts.

Key psychological state: Buyers are cautious, not impulsive. Content should build 
trust, not create urgency. Avoid any language that implies scarcity or pressure 
("don't miss this opportunity", "few units remaining").

This block changes how the AI frames everything — word choice, what to include, what to leave out. Once I added this section, my first drafts started requiring maybe 20% of the edits they used to.

Step 5 — Set Output Defaults and Formatting Preferences

Tell the AI what finished output looks like for you. Does it go straight into an email? Into a CMS? Do you want headers or no headers? Bullet points or prose? These defaults should match your actual workflow.

OUTPUT DEFAULTS:
- Deliver content ready to use — no explanatory preamble ("Here is your property 
  description:"), no closing notes ("Let me know if you'd like changes")
- Use plain text for property descriptions and emails — no markdown formatting
- Use markdown headers (H2, H3) only for market analysis reports
- When asked to write multiple variations, number them clearly (Version 1, Version 2)
- If a request is ambiguous, ask one clarifying question before writing — 
  do not guess and produce something that misses the point

That last rule — ask one clarifying question instead of guessing — saved me hours. Without it, I’d get a 400-word property description when I needed 200, or formal copy when I needed casual. One question takes 10 seconds. Rewriting takes 15 minutes.

Step 6 — Compile, Test, and Iterate

Put all five blocks together in a single document. Order matters: identity → voice → content types → audience → output defaults. This is the sequence from most general to most specific, which is how LLMs process instruction best.

Now test it with three tasks that cover your most common output types. Score each result on two criteria: Did it match your voice? Did it require significant editing?

If you score low on either, go back and tighten the relevant block. Usually the problem is either too vague (the voice section needs more specific rules) or too restrictive (you’ve blocked so much that the AI produces stilted, hedged output).

Give yourself three rounds of testing before you consider it done. I was on round four before I stopped noticing things that needed fixing.

Where to Install Your System Prompt in Each Tool

Tool	Where to Put It	Persistence	Cost
ChatGPT (Custom GPT)	GPT Builder → Instructions field	Permanent across all conversations	$20/month (Plus)
Claude (Project)	Projects → Custom Instructions	Permanent within the project	$20/month (Pro)
ChatGPT (Memory)	Settings → Personalization → Custom Instructions	Applies to new chats, not retroactive	$20/month (Plus)
Gemini Advanced	Gems → Create a Gem	Permanent within the Gem	$19.99/month
API (any LLM)	System message field in API call	Per-call — must be sent each time	Pay per token

I run my main setup as a Custom GPT in ChatGPT Plus and a Project in Claude Pro. Different tools have different strengths — I use Claude for longer analytical writing and ChatGPT for quick property descriptions and email drafts. Same system prompt, adapted slightly for each.

My Real-World Experience Writing and Using a System Prompt for My Madeira Real Estate Business

In January 2023, I was spending roughly 3 hours every week writing property descriptions. Not because the properties were complicated to describe — most listings in Madeira follow a predictable structure — but because I was starting from scratch every single time. I’d open a blank doc, look at the photos and spec sheet the seller gave me, and write. Then rewrite. Then trim. Then adjust the tone because the first version sounded like a brochure from a 1990s travel agency.

I started using ChatGPT in early 2023, but for the first four months I was just pasting in specs and hoping for something usable. What I got back was technically correct and completely generic. “Charming property nestled in the hills with breathtaking ocean views.” Every. Single. Time. I was spending more time editing than I would have spent just writing from scratch.

I built my first proper system prompt in May 2023. It took me about 40 minutes — most of that time was figuring out how to describe my own voice, which is harder than it sounds when you’ve never had to articulate it before. I wrote a list of 12 phrases I’d never use in client-facing copy and added them all to the banned list. I described the structure I follow for property descriptions — the same structure I’d been using intuitively for 10 years without ever writing it down.

The first test run was for a 3-bedroom villa in São Gonçalo I’d just taken on. I gave the AI the floor plan data, the address, and five bullet points of features. The output came back at 210 words, in my voice, with no generic phrases, and the only edit I made was changing one sentence that described the garden as “lush” — which I’d forgotten to add to my banned list. I added it immediately.

Over the following month, I wrote descriptions for 14 listings. Total time: 52 minutes. That same volume would have taken me roughly 3.5 hours without the system prompt doing the heavy lifting on structure and tone. I recovered nearly 3 hours in a single month from one task. That’s time I moved into prospecting and client calls — activities that actually grow the business.

The system prompt has gone through six major revisions since then. I update it every time I catch the AI producing something that feels off — a phrase I hate, a structural habit that doesn’t fit my workflow, an assumption about my clients that isn’t true. It’s a living document, not a one-time setup. Each revision takes maybe 10 minutes and usually locks in another hour of saved editing time per month.

What I didn’t expect: the discipline of writing the system prompt made me significantly clearer about my own brand voice. When you’re forced to write explicit rules for how you communicate — not vague adjectives, but actual behavioral rules — you get very precise about what you actually believe good writing looks like. That clarity shows up in everything, including the writing I do without AI assistance.

The Real Limitation You Need to Know Before You Rely on This

A system prompt doesn’t solve the knowledge problem. It handles voice, structure, and behavior — but it cannot supply facts. This is the limitation that bit me hardest.

In August 2023, I asked my ChatGPT setup — running my full system prompt — to help me draft a market analysis section on Madeira property price trends for Q2 2023. The output was beautifully structured, matched my tone exactly, and contained figures that were either invented or pulled from outdated training data. I caught it before sending because I happened to know the actual numbers. But I nearly didn’t check.

The system prompt makes the AI behave like you. It does not make the AI know what you know. Any time you’re writing content that contains data — prices, yields, market statistics, legal requirements — you must provide those numbers in your prompt. The AI’s job is to format and present the information you feed it, not to supply information it doesn’t reliably have.

I now have a standing rule in my own workflow: for any content that contains a number I didn’t personally provide in the input, I verify before publishing. The system prompt reminds the AI to flag when it’s making assumptions — but it’s still my job to catch what it misses.

Troubleshooting Tips When Your System Prompt Isn’t Working

The AI keeps ignoring one specific rule. Put that rule at the top of the relevant block, not buried in a list. LLMs give more weight to instructions that appear early. If it’s a critical rule — like never making up rental yield figures — put it in its own sentence, not in a list with 10 other items.

Output sounds robotic after adding too many rules. You’ve over-constrained it. Remove the weakest 20% of your rules — the ones addressing edge cases you’ve seen once rather than patterns you see constantly. A tightly scoped system prompt of 400 words outperforms a bloated 1,200-word version every time.

The AI keeps adding preamble (“Sure! Here’s your description:”). Add this to your output defaults block: “Never acknowledge the request or comment on what you’re about to do. Begin the content immediately.” This eliminates the assistant-speak habit reliably.

Results are good in testing but drift over long conversations. This is a known behavior. For long-form work or multi-step tasks, start a fresh conversation rather than continuing from a thread that’s 20+ exchanges long. Alternatively, paste a shortened version of your core rules at the top of each new significant task.

You’re not sure which block is causing problems. Comment out sections temporarily (just add a note like “[TEMPORARILY REMOVED FOR TESTING]” before a block) and run the same test. You’ll isolate the problem block within 2–3 tests.

What a Working System Prompt Actually Gets You

The right system prompt cuts editing time by 60–70% on content types you produce regularly. It eliminates the mental overhead of re-explaining your context every time you open a new chat. It makes the AI’s first draft usable rather than a starting point you’re rebuilding from scratch.

For a solopreneur, that’s not a minor efficiency gain. That’s hours per week that compound. I run a one-person operation in Madeira with the content output of a small team — not because I work more hours, but because the tools I use actually know who they’re working for.

Build your system prompt this week. Test it on three real tasks. Revise once. Then move on. It doesn’t need to be perfect — it needs to be specific enough to produce a first draft you’re not rewriting from scratch. That bar is achievable in under an hour, and the return on that hour is immediate.

If you want to share what you’ve built or run into a specific problem with your setup, drop it in the comments below. I check them regularly and the most common issues are usually fixable with a single sentence added to the right block.