If you want to know how to automate user feedback processing with AI agents, the answer is a three-part pipeline: collect rich feedback from real users, run a scheduled AI agent to read and classify it, then let another AI agent fix the issues in bulk and open a pull request — all without you touching a line of code. After one week of running this system with Fable AI and Cursor, the developer behind the Kora project woke up to a merged PR and a teammate saying the design looked slick. He had been asleep the whole time.
How Do You Automate User Feedback With AI Agents?
The workflow has three distinct layers that connect together in a loop. Here is how each one works in practice.
Layer 1 — Collect feedback in a Slack channel
Internal testers use a dedicated Slack channel, in this case a v2 alpha channel, to drop screenshots, written notes, and recorded sessions. Every piece of feedback lands in one place, making it trivially easy for an agent to find it later. Feedback items get tagged with emoji to track status: an eye emoji means someone is looking at it, a checkmark means it is resolved. This low-tech signaling system keeps the whole team aligned without any extra tooling.
Layer 2 — A scheduled agent reads and structures everything
A scheduled routine inside Cloth Co-Work runs twice a day — once in the morning, once in the evening. The agent connects to Slack via the Slack MCP, reads every message in the alpha channel, and classifies each one as a bug, a feature request, or a note. If any message contains a recording or video file, the agent downloads it automatically so it can be analyzed. The structured output gets written into YAML and Markdown files and committed to a new branch as a pull request. By the time you open your laptop, there is already a PR waiting with a full inventory of everything users reported.
Layer 3 — Cursor fixes all the items in one pass
With the PR branch checked out, you drop a single prompt into Cursor using the LFG flow from Compound Engineering: fix all the items in this PR, leave notes for anything that needs human input, and consult the strategy document before making any big decisions. Cursor then works through every line item, applies fixes, and generates a video walkthrough of every change it made. The whole batch runs in two to four hours. You review the walkthrough, check the app using the slash f polish command to get a live preview, and decide whether to merge.
Why Is Fable AI Changing the Way Devs Ship Features?
The creator describes Fable as feeling like a big shift — the moment when stuff that used to not work suddenly started working. After one week, it became his favorite model ever. The key difference is reliability at scale. Running an agent across 17 separate feedback items in one session is a serious ask. Previous models would lose context, make inconsistent edits, or fail partway through. Fable holds the whole batch together and ships coherent changes across the codebase. For a solo developer or small team trying to move fast, that reliability is everything. The speed at which you can now ship features is, by his own description, insane.
How to Collect Rich User Feedback From a React App
One of the most interesting pieces of this system is how feedback gets captured in the first place. The developer built and open-sourced a library called Rifreck — a lightweight wrapper you drop around your React app. Once installed, it adds a button that testers can click to start a recording session. Here is what Rifreck captures during a session:
- Click events — exactly what the user tapped or clicked on
- Voice narration — the user can talk through what they are experiencing
- Network requests — every API call made during the session
- Errors — any JavaScript errors that fired in the console
When the user hits stop, Rifreck saves everything into a single file that can be shared directly to Slack. The result is richer than a screen recording because it bundles the technical context — the network layer, the error stack — alongside what the user actually saw and said. The scheduled agent then downloads these files and analyzes them as part of its classification pass. If you do not want to build a custom tool, plain video recordings work too, but the structured file format makes the agent's job significantly easier.
What Is Slack MCP and How Does It Fit the Workflow?
MCP stands for Model Context Protocol — a standard that lets AI agents connect to external services and take actions on your behalf. The Slack MCP gives the scheduled agent permission to read messages, fetch file attachments, and navigate channel history. Without it, you would have to manually copy feedback into a document before any AI could process it. With it, the agent goes straight to the source, pulls everything that happened since the last run, and structures it automatically. This is the piece that makes the twice-daily cadence possible with zero human effort in between runs.
What Is Compound Engineering and Why Does It Matter?
Compound Engineering is a framework — available inside Cursor and other editors — that introduces a learning loop into agentic workflows. The LFG flow is one of its core patterns: a structured prompt sequence that tells the agent how to approach a large batch of tasks systematically. The compound aspect is the important one: if the agent makes a mistake in one session, that mistake gets logged. In the next session, the agent checks its own history and avoids repeating the same error. Over time the system gets smarter about your specific codebase, your patterns, and the kinds of fixes your app tends to need. It is compounding returns applied to AI-assisted development.
Why Batching Bug Fixes With AI Is a Game-Changer
Think about what the alternative looks like. Seventeen feedback items means seventeen separate context switches — reading the report, finding the relevant code, making the fix, testing it, opening a PR. Even at thirty minutes per item, that is over eight hours of work. With this pipeline, all seventeen items go into one agent session. You review one pull request instead of seventeen. The cognitive overhead collapses. The creator makes the point explicitly: imagine having to review 17 separate pull requests. That is a lot of work. Batching makes the volume manageable and keeps your review time short enough that you can actually do it every day.
How to Auto-Merge PRs While You Sleep
The moment that crystallized the value of this entire system happened on a single evening. After reviewing a batch of fixes and feeling confident in the direction, the developer added one final instruction: if everything looks good and the CI pipeline is green, go ahead and merge it. He went to sleep. The agent finished its work, the tests passed, and the PR merged automatically. The next morning, a teammate messaged saying the design looked great — things felt slick and polished. That is the end goal of this kind of workflow. You are not just saving time during working hours. You are turning idle hours into productive shipping cycles. The factory runs while you rest.
How to Start Building This Workflow Yourself
You do not need to implement all of this at once. Here is a practical starting order:
- Start with feedback collection. Get Rifreck or any screen recording tool in front of your real users and create a Slack channel for them to post in.
- Set up the Slack MCP. Connect it to a Claude or Fable-powered agent and test a simple read-and-summarize task on your channel history.
- Build the scheduled classification agent. Even a daily run at midnight is enough to start. Have it output a simple Markdown list of issues.
- Bring in Cursor with the LFG flow. Point it at your issue list and let it attempt fixes on a feature branch.
- Add the auto-merge step last. Only once you trust the agent's output quality should you let it merge without your sign-off.
The whole thing is still evolving. The creator is looking at ways to automate the middle step — the manual cursor prompt — so the pipeline runs fully end-to-end without any human input at all. If you are building something similar or have ideas for making this more robust, the community conversation is very much open.
