I wasted six months writing Claude prompts that kind of worked. They produced decent output, but I kept rewriting them every single time I sat down to draft a property description or send a follow-up email to a buyer lead. No consistency. No real time savings. Just me, typing variations of the same instructions over and over like I hadn’t learned anything.
The fix wasn’t a better AI tool. It was building proper custom prompts — structured, reusable, niche-specific templates that Claude could run with immediately. Once I cracked that, my workflow changed completely. I’m a one-person real estate consulting operation in Madeira, Portugal. Every hour I save is an hour I can put toward actual client work or, honestly, not working. This tutorial shows you exactly how I build Claude custom prompts for my niche, step by step.
What You’ll Build by the End of This Tutorial
By the time you finish this, you’ll have a working library of Claude custom prompts tailored to your specific niche — not generic AI output, but prompts that produce on-brand, context-aware content you can actually use without heavy editing. For solopreneurs, that distinction is everything.
Specifically, you’ll build three core prompt types:
- A Role + Context prompt that tells Claude who it is and what it knows about your business
- A Task prompt with output structure baked in
- A Variable prompt that lets you swap in new details without rewriting from scratch
Prerequisites Before You Start
You need a Claude account. The free tier works for testing your prompts, but Claude Pro ($20/month as of 2026) gives you access to Claude 3.5 Sonnet and Claude 3 Opus, which handle nuanced, long-form niche content far better than the base model. I use Pro and have since early 2024 — it’s worth it if you’re running a business.
You also need to know three things about your own niche before writing a single prompt:
- Your typical client profile (who are you writing for or selling to?)
- Your output format (email, listing description, social post, report?)
- Your non-negotiable tone requirements (formal, conversational, Portuguese market, UK buyers, etc.)
Don’t skip this prep. Vague prompts produce vague output. The specificity you put in is the specificity you get back.
Step 1 — Write Your Role + Context Block
This is the foundation of every custom prompt. Claude needs to know who it’s playing and what it already “knows” about your business. Think of it as a briefing document you hand to a new contractor on day one.
Here’s the exact structure I use:
You are [ROLE], working for [BUSINESS NAME], a [BUSINESS TYPE] based in [LOCATION].
Your audience is [CLIENT PROFILE].
Key facts about this business:
- [Fact 1: specialization or niche focus]
- [Fact 2: tone or communication style]
- [Fact 3: a market-specific detail your audience cares about]
- [Fact 4: any legal, regulatory, or format constraints]
Always write in [LANGUAGE]. Never use [PROHIBITED PHRASES OR FORMATS].
Here’s my actual version, filled in for my Madeira real estate business:
You are a professional real estate copywriter working for Robson Penassi Consulting, a solo real estate consultancy based in Madeira, Portugal, operating since 2012.
Your audience is international buyers — primarily from the UK, Germany, and Scandinavia — looking for residential properties or investment opportunities in Madeira, including the NHR tax regime.
Key facts about this business:
- Focus on mid-to-high-end residential properties and Golden Visa-adjacent investments
- Tone is warm but professional; avoid overselling and corporate jargon
- Madeira's key selling points include year-round mild climate, EU residency pathways, and low crime rates
- All property descriptions must comply with Portuguese real estate advertising standards; do not include projected rental yield figures unless explicitly provided
Always write in American English. Never use "paradise," "hidden gem," or "once-in-a-lifetime opportunity."
Save this block somewhere permanent — a Notion page, a text file, wherever you store working documents. You’ll paste it at the top of every new Claude prompt you build.
Step 2 — Build Your Task Prompt with Output Structure
The Role + Context block tells Claude who it is. The Task prompt tells it exactly what to produce. Most people stop at “write me a property description” and then complain the output needs too much editing. That’s a task prompt problem.
A strong task prompt has four parts:
- The action — what Claude should do
- The output structure — what sections or format you want
- The length constraint — exact word count or paragraph count
- The quality filter — what makes this output successful
Template:
Task: Write a [OUTPUT TYPE] for [SPECIFIC PURPOSE].
Structure:
- [Section 1 name]: [what goes here, how long]
- [Section 2 name]: [what goes here, how long]
- [Section 3 name]: [what goes here, how long]
Total length: [word count or character limit]
Quality criteria:
- [Criterion 1: specific thing this must do]
- [Criterion 2: specific thing this must avoid]
- [Criterion 3: tone or style requirement]
My property description task prompt looks like this:
Task: Write a property listing description for publication on real estate portals and my website.
Structure:
- Opening hook (1 sentence, 20-30 words): Lead with the strongest feature or the lifestyle this property enables. No generic openings.
- Property overview (2-3 sentences): Key specs — bedrooms, bathrooms, size in m², location within Madeira, and standout architectural or interior features.
- Living experience (2 sentences): What it actually feels like to live here. Sensory, concrete, honest.
- Location context (2 sentences): Proximity to key amenities, views, neighbourhood character.
- Call to action (1 sentence): Invite serious enquiries. Keep it professional.
Total length: 180-220 words.
Quality criteria:
- Must feel like something a real person wrote, not a template
- No superlatives without supporting detail ("stunning" only if paired with a specific feature)
- Avoid future-tense projections about value or rental income
Step 3 — Add the Variable Layer
This is where the prompt becomes genuinely reusable. Instead of rewriting the whole thing for every new listing or email, you add a clearly marked variable section at the bottom. Claude reads it as the specific inputs for this particular run.
I use capital letters in brackets to mark variables so they’re impossible to miss:
---
PROPERTY DETAILS FOR THIS LISTING:
Property type: [APARTMENT / VILLA / TOWNHOUSE / LAND]
Location: [AREA IN MADEIRA, e.g., Funchal city center, Calheta, Ponta do Sol]
Bedrooms: [NUMBER]
Bathrooms: [NUMBER]
Size: [M² INTERIOR] interior, [M² EXTERIOR OR TERRACE] terrace/garden
Key features: [LIST 3-5 STANDOUT FEATURES, e.g., sea view from living room, private heated pool, recently renovated kitchen]
Price: [PRICE IN EUR or "price on request"]
Special notes: [ANYTHING SPECIFIC — gated community, pet-friendly, short-term rental license held, etc.]
---
When I have a new listing, I paste the full prompt (Role block + Task block + Variable section), fill in the variables, and hit send. The output I get needs maybe 2-3 minutes of light editing, not a full rewrite. That’s the point.
Step 4 — Test, Break, and Refine
Don’t trust your first version. Run the same prompt with three different sets of variables and read the outputs side by side. Look for:
- Places where Claude defaulted to generic language despite your instructions
- Sections that came out too long or too short
- Tone drift — where it started professional and got overly casual (or vice versa)
- Banned phrases slipping through
When you find a problem, fix the prompt, not just that one output. Add a line to your quality criteria, tighten a structure instruction, or add an explicit example of what good looks like.
I keep a running notes section at the bottom of my prompt document that tracks every fix I’ve made and why. After six months of iteration, my property description prompt has 11 refinements logged. Each one came from a real output failure.
Step 5 — Build a Prompt Library, Not Just One Prompt
One good prompt is useful. A library of 6-8 interconnected prompts covering your full workflow is a real operational system. Here are the prompts I currently maintain for my real estate business:
| Prompt Name | Output Type | Use Frequency | Time Saved per Use |
|---|---|---|---|
| Property Description | Portal listing copy | 8-12x/month | ~20 min/listing |
| Buyer Inquiry Reply | Email response | Daily | 10-15 min/email |
| Lead Follow-Up Sequence | 3-email nurture sequence | 2-3x/month | ~45 min/sequence |
| Market Update Report | Client-facing PDF section | Monthly | ~90 min/report |
| Instagram Caption | Social post with hashtags | 3-4x/week | ~8 min/post |
| Seller Pitch Deck Outline | Slide structure + key points | 1-2x/month | ~60 min/deck |
| NHR Regime Explainer | FAQ-style client document | On demand | ~30 min/document |
Each of these has its own Role + Context block (slightly adapted), its own Task prompt, and its own variable section. I store all of them in one Notion database with a tag system so I can filter by output type.
My Real-World Experience Building These Prompts for Madeira Real Estate
Let me give you a specific example of what this system actually changed for me.
In March 2026, I had 14 new listings come in within the same two-week window. That’s unusual — typically I’m managing 6-8 active listings at once, not 14. The seller side was a combination of an estate situation, two developers clearing inventory before summer, and a few private owners who’d been sitting on the fence. All at once. Which meant I needed 14 property descriptions, translated context documents for my German-speaking clients, and social media content for the best 6 of those listings — all within about 10 days.
Before I had the prompt library, a single property description took me roughly 25-35 minutes. Good ones took longer because I’d research comparables, think through the buyer angle, and rewrite the opening three or four times. For 14 listings that’s potentially 8 hours of writing work on top of the actual client coordination and site visits.
With my property description prompt fully built and tested, I processed all 14 listings in one afternoon. I spent about 5 minutes per listing filling in the variables — pulling specs from my notes, choosing which features to highlight — then Claude generated the draft. I edited each one for 3-5 minutes. Total time: just under 2 hours for all 14 descriptions. That’s a reduction from roughly 7-8 hours down to 2. I had time that evening to draft the social posts for the 6 featured properties, which would normally be a next-day task.
What made it work wasn’t Claude being magic. It was the specificity of the prompt. Because I’d already told Claude to avoid superlatives without supporting detail, not to mention rental yield projections, and to lead with lifestyle over specs, I didn’t have to correct those things in 14 drafts. The instructions were already there. The outputs were consistent enough that I could edit in a focused way rather than rebuilding each one from scratch.
The social posts were a different story. My Instagram caption prompt was less mature at that point — I’d only used it maybe 20 times — and Claude kept defaulting to a slightly too-casual tone that doesn’t match how my international buyer audience expects a consultant to communicate. I had to do more editing there. That prompt has since been updated with a specific line: “Match the tone of a knowledgeable friend with professional standards, not a lifestyle brand account.” That one instruction cut my editing time on social posts by about half.
The honest truth is that building the prompt library took time I didn’t want to spend upfront. Writing, testing, breaking, refining — I estimate I put in about 6-7 hours across two months building the core six prompts. But I recovered that investment within the first month of regular use. At this point I’m probably saving 8-10 hours per month compared to writing from scratch. For a solo operator, that’s not a small number.
What Claude Custom Prompts Do NOT Handle Well
I want to be direct about the limits I’ve actually hit, not theoretical ones.
Local market nuance doesn’t come free. Claude doesn’t know that a property in Calheta commands a different buyer profile than one in Funchal city center, or that certain streets in the old town have noise issues during summer festivals. That context has to be in your variable section every time. If you forget to include it, the output will be accurate but generic. You’re always the knowledge layer — Claude is the writing layer.
Long conversation threads degrade prompt adherence. If you’re working inside a long Claude session and you use your custom prompt mid-conversation, sometimes the earlier context of the conversation bleeds in and Claude starts drifting from your instructions. I’ve had it drop my banned phrase list after a long back-and-forth. My fix: for prompt-based production work, I always start a fresh conversation and paste the full prompt fresh. Don’t rely on Claude “remembering” your standards from three sessions ago.
Legal and regulatory content needs human review every time. My NHR regime explainer prompt produces good structural drafts, but Portuguese tax rules change, and Claude’s training data has a cutoff. I review every output against current official sources before it goes to a client. I’d never send AI-generated tax or legal content without that check.
Troubleshooting: When Your Prompt Produces Bad Output
Here are the five most common failures I’ve seen and how to fix them:
- Output is too generic — Your Role block isn’t specific enough. Add one more concrete fact about your niche or your audience. Generic inputs, generic outputs.
- Claude ignores your length constraint — Add the constraint to both the Task block and the variable section. Repetition works. “Total length: 200 words. Do not exceed 220 words under any circumstances.”
- Banned phrases keep appearing — Add a line: “Before finalizing, check your output and remove any instance of [phrase]. Replace with [alternative].” Making it a self-check instruction helps.
- Tone is inconsistent across outputs — Add a concrete tone example. Paste two or three sentences that represent the exact voice you want and label them “Write in this voice.”
- Structure isn’t being followed — Number your structure sections and tell Claude to label each section in the output. You can remove the labels before publishing, but having them visible makes it easy to see what Claude is actually doing with each part.
Practical Summary: What to Do This Week
If you’re a solopreneur and you’re not already running a prompt library, here’s the minimum viable version to build this week:
- Write your Role + Context block for your niche. Keep it under 200 words. Be specific about who your clients are and what you never want Claude to say.
- Pick your highest-volume output type — the thing you write most often. Build one Task prompt for that output with explicit structure and length constraints.
- Add a variable section with clearly marked inputs in brackets.
- Test it three times with real data. Note every fix you need to make.
- Refine the prompt, not the output. Save the updated version as your new standard.
Robson Penassi
Real estate consultant in Madeira, Portugal. Solopreneur since 2012. Testing AI tools since 2023 to automate his one-person business. Writes about what actually works — and what does not.
More articles by Robson →