🎞️ The feedback no one gave
Tools like PostHog and Hotjar offer session replays (session replays): you watch, like a video, as a user navigates your product. Where they clicked, where they paused, where they went back, where they gave up.
💳 A big-company habit
In its early months, the Stripe team watched its customers use the product, one by one, to find every pain point. This isn't just a small-company habit; it's what successful companies did at the beginning.
Many bugs only show up this way: a button that fails on a double-click, a form that disappears on a specific phone, a user trying something you never imagined.
✓ What to look for in recordings
- ✓Repeated clicks in the same spot (something didn't respond).
- ✓Scrolling up and down (the person is lost).
- ✓Leaving right after a specific step.
✗ Cautions
- ✗Recording sensitive fields without masking them.
- ✗Not disclosing data collection in the privacy policy.
- ✗Watching one recording and generalizing to everyone.
💡 Practical tip
Building your own AI recording tool is possible, but established tools are still more reliable and have free plans. Save your energy for your product.
⚙️ The /product-feedback command
Today, instead of watching every recording yourself, a fast multimodal model (such as Gemini Flash versions) can watch them all and tell you where the user journey is confusing and where people fail to reach the desired click in under ten seconds. Then, a coding agent like Claude Code turns that diagnosis into a fix plan.
Copy and run
Create the /product-feedback skill in Claude Code. Save it as .claude/skills/product-feedback/SKILL.md in the project folder.
---
name: product-feedback
description: Analyzes recent session recordings and generates a user journey fix plan.
---
# /product-feedback
Goal: find where users get stuck, even when no one has complained.
1. Download the <N, e.g., 30> most recent session recordings from <PostHog | Hotjar>
using <the tool's API/CLI, with the key in .env>. Skip sessions shorter than 10 s.
2. For each recording, ask the video model <e.g., Gemini Flash> for a summary with:
- the user's apparent goal;
- where they hesitated (repeated clicks, searching by scrolling, long pauses);
- whether they reached the main outcome in under 10 seconds;
- where they left.
3. Group the problems by screen and sort by the number of affected sessions.
4. Read the code for the most problematic screens in <front-end folder> and write
PLANO-FEEDBACK.md with: problem, evidence (number of sessions), likely cause,
proposed change, file to update, how to measure whether it improved.
5. Do NOT change any code. Stop and show me the plan for approval.✓ Let AI handle it
- ✓Watching and summarizing dozens of recordings.
- ✓Grouping problems by screen.
- ✓Proposing changes and identifying files.
✗ Keep these with you
- ✗Approving changes to the product.
- ✗Deciding what matters most to the business.
- ✗Checking that the recordings respect privacy.
⚠️ Cost and privacy
Sending videos to an external model has a per-use cost and sends browsing data outside your organization. Use recordings with masked fields, limit the number of sessions per run, and confirm that your terms of use allow this processing.
🧭 Copy the onboarding of products that already make money
Onboarding is how a user enters your product: from their first visit to the moment they understand it and start using it. Products that make money have refined this entry process through many tests. Instead of starting from scratch, study what they do.
How to study an onboarding flow
Find products with real revenue
Some sites list products with revenue verified through payment provider integrations. Revenue is a sign that the onboarding flow works.
Go through the entire flow
Create an account and try it for the first time. Or ask an agent with computer use (Codex has a built-in browser) to do it and record each screen.
Note what’s different
The order of the questions, what’s asked and when, button text, and what appears right after signup.
Adapt, don’t clone
Take the principle, not the design. What works for their audience may need adjusting for yours.
⚠️ Why not just ask "make a good onboarding flow"
If you ask the model for an onboarding flow without a reference, it gives you the average of what it saw in training. The average isn’t what converts best; it’s just the most common. Real references raise the result above average.
💡 Practical tip
Be careful with public metrics platforms: exposing your revenue on an open site also alerts copycats. Study others, but think carefully before exposing your numbers.
🖌️ Design without bloating your setup
A common recommendation is "install these seven design skills so your app doesn’t look AI-made." The result often looks... AI-made. Worse, each skill brings a different style, and when the agent uses three at once, the design paradigms clash. The setup gets bloated, and the visuals get confusing.
21st.dev
A gallery of UI components: backgrounds, borders, cards, buttons. Useful for seeing visual patterns.
whatships
Product launch videos. A reference for your site’s marketing video.
Awwwards
Award-winning sites of the month and year. Typography, layout, and interaction above average.
Copy and run
Ask an agent with a browser (Codex with computer use, or Claude with a browser) to study references and propose a visual guide for your product.
Use the browser to study design references. Do NOT install skills, plugins, or packages.
My product: <one-line positioning statement>
Audience: <who uses it>
Desired feel: <e.g., trustworthy and calm | fast and bold>
1. Open https://21st.dev and scroll through the components (backgrounds, borders, cards, buttons).
Note 5 that fit the product, and why.
2. Open https://whatships.com, watch 3 launch videos for similar products,
and describe frame by frame what makes each one clear in under 2 minutes.
3. Open https://www.awwwards.com (sites of the month/year), browse 3 sites to the end and note:
typography, button placement, scrolling rhythm, and what is unusual.
4. Write GUIA-VISUAL.md with: color palette, 2 fonts, 5 selected components,
homepage structure, and a 60-second video script.
5. Justify each choice based on the audience and desired feel.💡 Value per scroll
When reviewing your page, ask at every screenful: "What did the person learn here?" Minimalist doesn’t mean empty; each scroll should deliver value.
👥 Simulated users with subagents
A subagent is an AI agent that the main agent creates to handle a separate task, with its own instructions. You can create several, each with a persona: a user profile with a different age, routine, patience level, and experience.
Copy and run
Run a panel of 4 simulated users in Claude Code on your page or app.
Create 4 subagents, one for each persona below. Each one should independently analyze
<Page URL | project folder> from their own point of view.
Personas:
1. <Name>, 55 years old, <profession>, uses their phone for almost everything, is wary of paying online.
2. <Name>, 20 years old, ADHD, tested 100 products in the past week, decides in 10 seconds.
3. <Name>, a manager who approves purchases: wants clear costs, security, and returns.
4. <Name>, a skeptical technical expert: looks for flaws and exaggerated claims.
Each subagent answers:
- In 10 seconds, did I understand what this is? (yes/no + what I thought it was)
- What would make me leave the page?
- What would make me pay?
- The exact sentence I would change.
Then consolidate the results in PAINEL-PERSONAS.md: points where 3 or more agree,
important disagreements, and the 5 highest-impact changes, in order.⚠️ Simulation is not validation
Simulated personas help you find obvious problems before launch. They do not replace real people paying. Use them to improve the product, not to conclude that it will succeed.
🌐 Synthetic crowds and judgment
There are tools, such as Mirofish, that create hundreds or thousands of agents, each with a short system prompt describing who they are. Together, they simulate buyers in a market and tell you what they would think of your product, page, or ad.
📊 The test that builds confidence: the backtest
People have been using these tools for a simple test: take the last 10 posts from an account and ask a synthetic crowd modeled on that account’s audience to say which one would perform best. In many cases, it picks the post that actually went viral.
If the simulation gets the past right, you gain some confidence in using it for future decisions: which headline to test, which page to publish, which offer to make.
✓ Use simulation to
- ✓Choose between headline or page versions.
- ✓Find objections before spending on ads.
- ✓See how different audiences react.
✗ Don’t use simulation to
- ✗Conclude that the product will sell.
- ✗Skip talking to real customers.
- ✗Justify a decision you had already made.
🧭 Closing: judgment is the game
Everything this course has shown is within anyone’s reach today: building, designing, simulating, analyzing. That’s exactly why none of it is a differentiator on its own.
What sets you apart is judgment: knowing whether something is worth building, for whom, for how long, and how to reach those people. In a world where you can build anything, that judgment is the whole game.
💡 Practical next step
Take the product or idea you’re working on and do these in order: the one-line statement (3-2), the persona panel (3-3), and the distribution plan for the first 100 people (3-1). In one afternoon, you’ll have all three answers written down.
🧪 Quick module quiz
Three questions. Click an option to see the answer.
1. Why are session recordings valuable?
2. What’s the problem with installing several design skills at the same time?
3. What is the only definitive validation of a product?