How to Add Human Feedback to Your AI Coding Workflow
By Testbay
Testbay connects your coding workflow to people who try your website or iOS app and answer a focused brief. Through its MCP integration, your AI assistant can prepare an unpaid test draft and read your results and feedback. You review, fund and publish on Testbay, then bring what people found back into your next coding session.
If you have been vibe coding for a while, you probably know the feeling. You describe a feature, your agent builds it, and suddenly you are working on the next one. It feels productive. But somewhere between the first working version and your tenth new idea, a question gets lost: does anyone actually want to use this?
Human feedback gives that workflow a point where you stop guessing. Here is how to make it part of development without turning every small release into a research project.
Start with one decision you need to make
“Get feedback on my SaaS” is too broad to guide a useful session. “Can a first-time customer create a project without help?” is something you can investigate and act on.
Imagine you have built an invoicing tool. You could ask people to explore everything, but the result may be a collection of unrelated opinions. Instead, ask them to create a draft invoice for a fictional customer and explain what, if anything, made them hesitate. Now your next change has a clear purpose.
Before opening your agent, write down three things:
- Who should be able to complete the task.
- What they are trying to achieve, expressed in their own terms.
- Which product decision you will make after reading the feedback.
Connect your assistant to Testbay
Testbay provides a hosted MCP server at https://testbay.dev/mcp. Add it as a remote HTTP MCP connection in a supported assistant, then sign in to Testbay and approve access. The MCP setup guide has connection instructions for different assistants. There is also a CLI if you prefer your terminal.
The available tools can list your tests, prepare cash-reward website or public-link iOS TestFlight drafts, check participation and read submitted answers and design proposal comments. Your assistant can work with feedback available to your connected account. It does not gain access to everybody else's private tests.
Creating a draft does not publish a listing or spend money. Funding and publication remain on the Testbay website. Subscription rewards and more detailed setup can be configured there after you open the draft.
Ask your agent to prepare a focused brief
Give the agent the product context you already have in your coding session. Include the deployed URL or public TestFlight invitation link and explain the task you want people to attempt.
Create an unpaid Testbay draft for my invoicing website. Recruit five people, with a $5 cash reward each and an estimated 10-minute session. Ask them to create a draft invoice using fictional customer details. Ask what they expected to happen, where they hesitated, and whether the final result was clear. Do not publish it.
Those participant counts, rewards and durations are an example, not a recommended budget for every product. Match the compensation to the effort, including installation, account setup and written answers.
Review the draft yourself. Remove instructions that give away the route through the interface. “Create an invoice” can reveal a navigation problem; “click the blue button in the top right” may hide it.
Choose the evidence you actually need
In Testbay, review participant selection, feedback questions, reward and deadlines before funding. If you need website activity checks, install and verify the project-specific tracker. It reports active time and required page visits. You can also explicitly skip tracking and collect feedback under the remaining saved requirements.
These choices matter. A page visit does not prove an invoice was created successfully. An automatically approved answer does not independently prove product use. Ask about the outcome you care about and use activity evidence as context.
For native apps, read the iOS testing guide before choosing between App Store Connect and manual TestFlight setup. They collect different kinds of evidence.
Bring the feedback back into your coding session
Once participants have submitted feedback, ask the assistant to read it. A useful prompt keeps observations, interpretations and proposed work separate:
Read the feedback for this Testbay project. Group recurring problems, identify the participant comments behind each finding, and distinguish observed problems from feature requests. Suggest the smallest change worth testing next. Do not edit the code yet.
Keep the original comments nearby. If the assistant says “users find onboarding confusing,” check who said what and at which step. One person's preference should not silently become a claim about your whole audience.
MCP feedback retrieval provides written answers and design proposal comments. Do not assume it retrieves every screenshot, Apple statistic or activity report shown in the dashboard. Open Testbay when you need that additional context.
Ship one improvement, then check it
Suppose several relevant participants cannot tell whether their invoice is saved. The next useful change might be a clearer saved state, not a new dashboard. Ask the agent to implement that change, deploy it, then give the same task to a fresh set of relevant people.
The practical loop is short: build a working flow, recruit people who might use it, read what happened, and improve the part that got in their way. AI helps you move quickly. Human feedback helps you decide where that speed should go.
Start in the Testbay MCP guide, or create your first brief directly.
Put the feedback to work
Choose one product question, write a focused brief and offer people a reason to help you answer it.
Create a brief