How to test your content before people see it using AI
Hey, it's Cindy 🌱 You commented CROWD, so here is the full MiroFish setup: what it is, the two keys you need and the easiest way to install it.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
the idea
What MiroFish actually does
MiroFish is a free, open-source tool on GitHub (over 80,000 stars). You give it something you are about to put out, and it builds a whole crowd of AI people from it. Its README says it creates "thousands of intelligent agents with independent personalities, long-term memory, and behavioral logic". They react to your content and to each other, then a report agent writes up what happened.
Why a crowd beats one chat: if you paste your landing page into one AI and ask "is this good?", you get one opinion, and it tends to be nice to you. A crowd gives you lots of different people with different backgrounds, so you can see who leans in, who scrolls past, and why.
You give it Your launch plan, landing page copy, ad script, or social post, saved as a document.
It builds A simulated crowd with their own backgrounds and memory, reacting over a number of rounds.
You get back A written report, plus a live world where you can chat with any single person in the crowd.
It runs on Your own computer with your own keys, so you control the setup and the spend.
Think of it as a dress rehearsal. The project calls itself "a rehearsal laboratory for decision-makers", so use it to spot weak spots early, polish them, then launch to real people with more confidence.
before you start
What it costs and what you need
The code is free. MiroFish is open source under the AGPL-3.0 licence. You download it and run it on your own computer.
An AI key (you pay for your own usage). It works with any AI provider that uses the OpenAI format, so an OpenAI key works, and the project recommends Alibaba's Qwen-plus. The README says runs are "high consumption", so start small (below).
A free Zep key. Zep stores the crowd's memory. Its free plan is "Start free with 10,000 credits a month", which the README says "is sufficient for simple usage". Sign up at app.getzep.com.
A computer that can run it. Either Docker Desktop (easiest), or Node.js 18+, Python 3.11 or 3.12 and uv.
get set up
Install it, the easy way first
No coding background? Let an AI do the setup. If you have Claude Code (or any coding agent), open it in an empty folder and paste the prompt below. It downloads MiroFish, sets up the keys file and starts it for you, asking before each step.
💻 setup prompt for Claude Code
Set up MiroFish for me on this computer. The project is at https://github.com/666ghj/MiroFish.
1. Clone the repo into this folder.
2. Check whether Docker is installed. If it is, use the Docker option from the README. If not, check for Node.js 18 or newer, Python 3.11 or 3.12, and uv, and tell me exactly what is missing before installing anything.
3. Copy .env.example to .env. Ask me for my LLM_API_KEY, LLM_BASE_URL, LLM_MODEL_NAME and ZEP_API_KEY one at a time and fill them in. Never print my keys back to me.
4. Start the app and confirm the frontend opens at http://localhost:3000.
Explain each step in plain English before you run it, and ask me before installing anything.
Prefer to do it by hand? Here is the same thing step by step.
📥 Get the code. On the MiroFish GitHub page, click the green Code button, then Download ZIP, and unzip it. Or, in a terminal: git clone https://github.com/666ghj/MiroFish.git
🔑 Add your keys. In the MiroFish folder, copy the file .env.example and name the copy .env. Open it and fill in LLM_API_KEY, LLM_BASE_URL, LLM_MODEL_NAME and ZEP_API_KEY. For OpenAI, the base URL is https://api.openai.com/v1 and the model name is any model on your account. The README's default is Qwen-plus with https://dashscope.aliyuncs.com/compatible-mode/v1.
🐳 Start it with Docker. Open Docker Desktop, then in a terminal inside the MiroFish folder run docker compose up -d.
🛠️ Or start it from source. Run npm run setup:all once, then npm run dev every time you want to use it.
✅ Open it. Go to http://localhost:3000 in your browser. That is your own private MiroFish, running on your computer.
Use the official repo. Paid "hosted MiroFish" sites exist that are not run by the authors. The free, official project is github.com/666ghj/MiroFish.
do it
Run your first test
📄 Save your content as a document. Paste the thing you want to test into a text or PDF file: your landing page copy, the ad script, the post caption, or a one-page launch plan. Add one short paragraph about who it is for and the price, if there is one. This is what MiroFish calls "seed material".
⬆️ Upload it. In MiroFish, upload that file as the seed material.
✍️ Describe what you want to know. MiroFish asks you to describe your prediction in plain language. Use the prompt below and fill in the brackets.
🔢 Keep the first run small. The README says to "try simulations with fewer than 40 rounds first". A small run is cheaper and quicker, and tells you whether your setup works.
📊 Read the report. When it finishes, the report agent writes up how the crowd reacted.
💡 prediction prompt (paste into MiroFish)
This is a [landing page / ad / social media post / launch plan] for [what you are offering], made for [who it is for, for example busy small business owners in Australia]. The price is [price, or "free"].
Simulate how a realistic mix of people in that audience react when they first see it. Include people who would love it, people who are unsure, and people who would scroll straight past.
I want to know:
1. Who stops and pays attention, and what exactly made them stop.
2. Who scrolls past, and the real reason why.
3. Which words, claims or parts of the price feel confusing or hard to believe.
4. What people say to each other about it.
5. The three changes that would most improve how it lands.
the best part
Chat with anyone in the crowd
After the report, MiroFish lets you "chat with any agent in the simulated world". This is where the gold is: pick someone who scrolled past and ask them why.
💬 questions to ask a person who scrolled past
You saw my [post / ad / landing page] and kept scrolling. Be honest with me:
1. What was the first thing you noticed, and what made you move on?
2. What would it have needed to say for you to stop?
3. Was anything confusing, or did anything feel too good to be true?
4. If a friend sent this to you, what would you text back?
💬 questions to ask a person who loved it
You stopped and paid attention to my [post / ad / landing page]. Tell me:
1. Which exact line or idea made you stop?
2. What almost stopped you from going further?
3. How would you describe this to a friend in one sentence?
Then take it to real people. Use the crowd's answers to polish the weak spots, then post a version and see how your real audience responds. The rehearsal gets you to a stronger first draft, faster.