Claude Multi-Agent Workflow Guide From Scratch

Most solo operators think multi-agent AI is something that only engineering teams at big tech companies build. I thought the same thing until March 2026, when I accidentally set one up in Claude to handle a property valuation request — and it finished in 4 minutes what used to take me the better part of a morning. I’m a one-person real estate consultancy in Madeira. I don’t write code. If I can build a working Claude multi-agent workflow, you can too.

This tutorial walks you through exactly what I built, step by step, with the exact prompts I use and the real limitations I’ve hit along the way. No theory. No fluff about “the future of AI.” Just a practical setup you can replicate today.

What You’ll Build — And What It Actually Does

By the end of this tutorial, you’ll have a Claude multi-agent workflow with three specialized “agents” that pass work to each other inside a single conversation thread. Each agent has a defined role, a constrained set of instructions, and a clear output it hands off to the next one.

The example I’ll use is what I run in my own business: a property listing pipeline. Agent 1 researches and pulls together raw property facts. Agent 2 writes the listing copy. Agent 3 reviews and optimizes it for SEO and tone. You can adapt the same structure to almost any sequential task — client onboarding, market reports, social media content chains.

This is not about Claude’s API or anything that requires a developer. Everything here runs inside Claude.ai on the Pro plan ($20/month as of 2026) using Projects and extended thinking prompts.

Prerequisites: What You Need Before You Start

• A Claude Pro account ($20/month at claude.ai) — the free tier won’t give you the context window or Projects feature you need
• Access to Claude Projects — available on Pro; this is where you store each agent’s system prompt
• About 45 minutes for the initial setup
• A sequential task in your business that has at least 2–3 distinct stages (research → draft → review works perfectly)
• No coding required. Seriously.

Step-by-Step: Building Your First Claude Multi-Agent Workflow

Step 1 — Map Your Workflow Before Touching Claude

The biggest mistake beginners make is jumping straight into prompting without mapping the task first. Spend 10 minutes on paper (or a notes app) before you open Claude.

Ask yourself three questions: What are the distinct stages of this task? What does each stage need as input? What does it produce as output? Write it down.

For my listing pipeline it looks like this:

Agent	Role	Input	Output
Agent 1 — Researcher	Extract and organize raw property data	Raw notes, area, specs	Structured fact sheet
Agent 2 — Copywriter	Write full listing description	Fact sheet from Agent 1	Draft listing copy (400–500 words)
Agent 3 — Editor	Review for tone, SEO, accuracy	Draft copy from Agent 2	Final polished listing

If your workflow has more than five stages, split it into two separate workflows. Longer chains lose coherence fast.

Step 2 — Create a Separate Claude Project for Each Agent

In Claude.ai, click Projects in the left sidebar, then New Project. Name it clearly. I use: “Agent 1 — Property Researcher,” “Agent 2 — Listing Copywriter,” “Agent 3 — Listing Editor.”

Each Project has a Project Instructions field — this is your system prompt. This is what makes each project behave like a specialized agent. Open Agent 1’s Project Instructions and paste in your first system prompt.

Step 3 — Write the System Prompt for Agent 1 (The Researcher)

The system prompt defines the agent’s entire personality, constraints, and output format. Be specific. Vague instructions produce vague outputs.

You are a real estate research assistant specializing in properties in Madeira, Portugal.

Your job is to take raw property notes — which may be messy, incomplete, or in mixed languages — and turn them into a clean, structured fact sheet.

OUTPUT FORMAT (always use this exact structure):
- Property Type:
- Location (neighborhood + municipality):
- Total Area (m²):
- Bedrooms / Bathrooms:
- Key Features (bullet list, max 8 items):
- Nearest Amenities:
- Asking Price (if provided):
- Condition / Year Built (if known):
- One-sentence summary of unique selling point:

RULES:
- Do not invent facts. If a detail is missing, write "Not specified."
- Do not write any marketing language. Facts only.
- Flag any contradictions in the source notes with [CHECK THIS].
- Always respond in English unless instructed otherwise.

When you have completed the fact sheet, end with: "FACT SHEET COMPLETE — ready for Agent 2."

That last line matters. It creates a clear handoff signal you’ll use to know when to move to the next project.

Step 4 — Write the System Prompt for Agent 2 (The Copywriter)

Open your second Project. Paste this into Project Instructions:

You are a real estate copywriter specializing in luxury and mid-market properties in Madeira, Portugal.

Your job is to write compelling property listing descriptions based on a structured fact sheet provided by a research agent.

OUTPUT FORMAT:
- Headline (max 12 words, no question marks, no exclamation points)
- Opening paragraph (2–3 sentences, lead with the strongest selling point)
- Body (3–4 paragraphs covering location, interior, lifestyle, and practicalities)
- Closing paragraph (1–2 sentences, subtle call to action)
- Total word count: 400–500 words

TONE: Warm, confident, informative. Do NOT use: "stunning," "breathtaking," "don't miss out," "rare opportunity," "nestled," "boasting."

RULES:
- Every claim must appear in the fact sheet. Do not invent features.
- Write for an international English-speaking buyer audience.
- If the fact sheet contains a [CHECK THIS] flag, note it at the bottom as: "Editor note: [CHECK THIS] flagged by researcher — verify before publishing."

When you have completed the listing copy, end with: "DRAFT COMPLETE — ready for Agent 3."

Step 5 — Write the System Prompt for Agent 3 (The Editor)

You are a senior real estate content editor with expertise in SEO and international property marketing.

Your job is to review and improve listing copy produced by a copywriting agent. You are the final stage before publication.

YOUR REVIEW CHECKLIST:
1. SEO: Does the copy include the property type, location (Madeira), and key feature naturally in the first 100 words?
2. Tone: Is it consistent throughout? Flag any sudden shifts.
3. Accuracy: Does anything contradict the fact sheet? If yes, flag it.
4. Forbidden words check: Remove "stunning," "breathtaking," "nestled," "boasting," "rare opportunity," "don't miss out."
5. Headline: Is it specific and compelling without being sensational?
6. Length: Is it 400–500 words? If not, adjust.

OUTPUT:
- Revised headline (if changed, explain why in one line)
- Full revised listing copy
- Editor's notes (max 5 bullet points summarizing what you changed and why)

When you have completed the review, end with: "FINAL COPY READY — approved for publication."

Step 6 — Run Your First Full Pipeline

Now open a new chat inside your Agent 1 Project. Paste in your raw property notes. They can be messy — mine usually are. I dump in a mix of Portuguese notes from property owners, photos I’ve described in text, and area details I know from memory.

Claude will produce the structured fact sheet and end with “FACT SHEET COMPLETE — ready for Agent 2.” Copy the entire fact sheet output.

Open a new chat in your Agent 2 Project. Start the message with:

The following fact sheet was produced by the research agent. Please write the listing copy now.

[PASTE FACT SHEET HERE]

Copy the draft copy output. Open a new chat in your Agent 3 Project and paste:

The following listing copy was produced by the copywriting agent. The original fact sheet is included below for reference. Please review and produce the final version.

DRAFT COPY:
[PASTE DRAFT COPY]

ORIGINAL FACT SHEET:
[PASTE FACT SHEET]

Agent 3 will return the final polished copy plus editor notes. That’s your complete output, ready to publish.

Step 7 — Save Your Handoff Template

Create a simple text file with your three handoff prompts saved as templates. I keep mine in Notion. Every time I run a new listing, I open the template, swap in the new content, and run the chain. The whole process takes about 12 minutes once the workflow is set up.

My Real-World Experience Running This in Madeira

I want to be honest about how this actually started — it wasn’t planned. In late February 2026, I had 9 listings to write in one week. A colleague who’d been helping me with translations had moved to Lisbon, and I was on my own. Nine listings in five days, each needing English copy for international buyers and a Portuguese version for local portals.

Before I built this workflow, my process was chaotic. I’d open Claude, dump in everything I knew about a property, and basically write a long prompt begging it to produce good copy. Sometimes it worked. More often I’d get something generic that didn’t capture what made the property interesting — a quinta in Câmara de Lobos with a century-old wine press, for example, getting described in the same flat language as a new-build apartment in Funchal. I was spending about 45–55 minutes per listing, including rewrites.

I built the three-agent chain over one evening, testing and adjusting the system prompts across about 6 iterations. The next morning I ran all 9 listings through it. Total time for all 9: 2 hours and 20 minutes, including my own final read-through and minor edits. That’s an average of just over 15 minutes per listing, down from roughly 50 minutes. I recovered about 5.5 hours in a single week.

The quality difference surprised me more than the speed. Because Agent 1 is constrained to facts only, and Agent 2 is constrained to what Agent 1 produced, the copy stopped inventing features. That wine press in Câmara de Lobos? It showed up in the fact sheet, got flagged as a key feature, and Agent 2 built the opening paragraph around it. The editor tightened the language. The final copy needed three small edits from me. That was it.

By mid-March I’d processed 23 listings through the workflow. I’ve since adapted the same three-agent structure for market analysis reports — Agent 1 organizes raw data, Agent 2 drafts the narrative, Agent 3 formats it for client delivery. The time saving there is even larger because those reports used to take me 3–4 hours each.

The Pro plan costs me $20/month. At my consulting rate, recovering 5.5 hours in a single week more than pays for a year of that subscription.

Genuine Limitations I’ve Hit in Daily Use

This setup has real constraints and I won’t pretend otherwise.

The biggest one: you’re doing the handoffs manually. Copy, paste, open new tab, paste again. It takes about 2 minutes per transition and if you’re running 10 listings it adds up. A true automated multi-agent system built on Claude’s API with something like Make.com would handle this automatically — but that requires either coding knowledge or a willingness to learn Make, which is a separate learning curve. What I’ve described here is a manual workflow that mimics multi-agent behavior. It is not a fully automated pipeline.

Second limitation: context doesn’t carry between Projects. Each agent starts fresh. If Agent 1 made an assumption or judgment call that isn’t visible in its output, Agents 2 and 3 won’t know about it. This is why the handoff prompts matter — you have to pass everything forward explicitly.

Third: the system prompts need maintenance. I’ve rewritten mine four times since March 2026. As I add new property types to my portfolio (I recently started working with commercial units in Funchal), I need to update the researcher agent’s instructions. It’s low effort, but it’s not a set-it-and-forget-it situation.

Troubleshooting the Most Common Problems

Agent Output Is Too Generic

Your system prompt is probably too short. Add a “DO NOT” list of generic phrases specific to your industry. For real estate I ban “nestled,” “stunning views,” and “perfect for families.” Constraints sharpen outputs dramatically.

Agent 2 Is Inventing Facts

Add this line to Agent 2’s system prompt: “If a feature is not listed in the fact sheet, do not include it. If you feel a feature is missing, add an editor note at the bottom: ‘Suggested addition — verify with agent: [feature].'” This stops hallucination without killing creative output entirely.

Agent 3 Is Making Too Many Changes

Add a constraint: “Make only changes that are necessary for accuracy, SEO, or tone consistency. Do not rewrite for style preference. If Agent 2’s copy is already strong, say so and return it with minimal edits.” Without this, the editor agent can go overboard.

The Workflow Feels Too Long for Simple Tasks

For shorter tasks, collapse it to two agents: Drafter and Editor. I use a two-agent chain for social media captions — one agent writes three variations, the second picks the strongest and refines it. That takes under 5 minutes end to end.

Quick Reference: Agent Setup Comparison

Agent	System Prompt Length	Key Constraint	Handoff Signal	Avg Output Time
Researcher	~150 words	Facts only, no marketing language	“FACT SHEET COMPLETE”	~3 min
Copywriter	~180 words	Must use fact sheet only	“DRAFT COMPLETE”	~4 min
Editor	~200 words	Structured checklist, minimal rewrites	“FINAL COPY READY”	~3 min

Practical Summary and Next Steps

Here’s what to take away from this tutorial. Multi-agent workflows in Claude don’t require code, an API, or a technical background. They require clear thinking about your task structure and specific system prompts that constrain each agent’s role. The manual handoffs are a real limitation — but for a solo operator running 10–30 tasks a month, the time saving is significant enough that the copy-paste overhead doesn’t matter.

Start with three agents. Map your task on paper first. Write system prompts that are at least 100 words each with explicit output formats and a short DO NOT list. Run one full test before you rely on it for client-facing work. Adjust the prompts after the first real run — you’ll spot gaps immediately.

If this setup works for you and you want to take it further, the logical next step is connecting it to Make.com to automate the handoffs. That turns this manual workflow into a true automated pipeline. I’m currently building that version and will document it once it’s been running reliably for a full month.

Want the exact system prompt templates I use for my Madeira property listings? I’ve put them together as a free download on the tools page — grab them, adapt the location and property type references to your own market, and you’ll have a working workflow in under an hour.