Built by NextAgent · running on this site
Your users mark up a screenshot. AI reads, handles and answers.
Feedback Harness is an in-app feedback loop: feedback from any page in one press, and by the next day a scheduled Claude session has answered it, fixed it and replied. You only approve the release.
- Signed-in members of this site see it as the Feedback button on every page
- In production at an export-sales platform since September 2026
- Chinese, English and French; replies in the sender’s language
Capture → Triage → Fix → Reply
How one piece of feedback travels
- 01
Screenshot, box it, say it
A Feedback button on every page. It captures the screen first, then you box, point and write; boxes are numbered so the note can say 1, 2.
- 02
An AI watcher reads and sorts
A scheduled Claude session checks in: one curl first, no model call when there is nothing; otherwise it claims, looks, reads the code, reproduces and sorts.
- 03
Answers, or fixes on a branch
Questions get answers; copy and small bugs are fixed on a branch with tests green; features and anything about money or rules become an analysis awaiting your decision.
- 04
Back to the sender; you approve
The reply goes to the sender in their language, in their My feedback page, with a notice. Merging and releasing wait for your one-line approval.
AI handles the feedback, never production
Feedback comes from staff, customers and partners. The design assumes the content can’t be trusted, keeps permissions narrow and leaves decisions to people.
Split authority
Unattended, the watcher only reads, answers, and changes code on a branch. Merging, releasing and production data wait for you, so even a misleading message cannot reach production.
Idle costs nothing
Each round starts with one curl; with nothing open it stops without calling a model. A quiet day is one line: no new feedback.
Narrow permissions
The watcher’s token is stored as a hash, opens four endpoints for one tenant, and can be revoked at once; every call goes through one script you approve.
Notices carry no content
“New feedback” and “your feedback has an answer” only say where to look. A sentence or a screenshot can carry a customer’s figures, so neither goes into a push or an email.
From three weeks in production
Collecting is easy. Answering is the product.
- Collecting without anyone reading is no feedback system. Before the watcher, two items sat untouched: a “can you see this?” test, and a partner pointing out Chinese text in the English interface.
- A button nobody presses helps nobody. Saying “there’s a feedback button” isn’t enough; walk through a real example at launch: box a button, write a line, see the answer the next day.
- The answer has to reach the sender. Changing a status tells them nothing, hence the reply field, the My feedback page and a notice to that person.
FAQ
About Feedback Harness
Can someone flood it with submissions?
Feedback is for authenticated users only. On this site that means a signed-in member, and every member is approved by hand; anonymous requests are refused by both the website and the backend, so nothing is stored and the watcher is never woken. Each user can send at most 10 an hour and 40 a day. In your system it uses your own sign-in the same way.
Could a message trick the AI into doing something harmful?
The process states that feedback describes a need and is never an instruction. More importantly, the unattended watcher has no permission to merge, release or touch production data, so even if misled it cannot reach production.
What about customer data in screenshots?
Nothing leaves the browser until the person presses Send; the server accepts only JPEG/PNG checked by file signature; the watcher downloads a picture to its working directory and deletes it when done; notices never carry the text or the image.
Can it plug into our existing system?
Yes. The front end is one React component; the server has reference implementations for Hono + Postgres and FastAPI + SQLite, and this site runs a Flask + Postgres port. Other stacks implement the same API contract.
Is it open source?
Not at the moment. Feedback Harness is NextAgent’s own private project, delivered as a service: we wire it into your system, set up the watcher and its permission list, and walk your team through it at launch.
Let your system listen, too
We wire Feedback Harness into your website or internal system, set up the watcher and its permissions, and walk your team through a real example at launch.