Command Palette

Search for a command to run...

BlogAIDeveloper ToolsProductivity

November 16, 2026

How to Turn a UX Review Into a Tracker Your Coding Agent Can Finish

One review doc became 117 numbered rows, and a coding agent fixed 96 of them in about 8 hours. The tracker format, the rules file and the prompts.

By Nishant Modi · 9 min read

Featured image: How to Turn a UX Review Into a Tracker Your Coding Agent Can Finish

Sooner or later someone uses your app properly and writes down everything that is wrong with it. You get a document with screenshots, notes on every screen and a few lines like "not sure what this is". It is the most useful thing anyone will send you that month, and it is also the easiest to leave unopened, because it feels like too much.

At Careerkit, a resume and career tool, a reviewer went through the whole signed-in app between November 9 and 11, with their own account and their own resume. They wrote "UX & copy suggestions" for each screen and attached 32 screenshots. On Saturday, November 15, we handed that document to a coding agent. It became a tracker of 117 numbered rows. Between about 4pm and just after midnight, 96 of those rows were fixed in 71 commits, in English, German and French.

Below: how the document became rows, what a good row contains, why the batches follow the kind of fix rather than the screen, the rules file every agent session reads, the rows the agent checked and refused to change, what still needs a human, and a step-by-step version you can run this week. One caveat up front: none of these fixes are tested or live yet. They sit on a branch.

Why a feedback doc stalls, and a tracker does not

A review document is organized for the person who wrote it: screen by screen, in the order they clicked. That is the right order for writing and the wrong order for fixing. It mixes a typo next to a broken flow next to an idea for a new feature, and it never tells you how many items are left.

The founder's instruction to the agent was short, and it carried the whole method:

"i want you to see all images as well in it, and then build a tracker, so we get it done one by one"

The point of "one by one" is visible progress. A document of suggestions does not shrink as you work. A numbered list does, and both you and the agent can see exactly where you are. It also gives every later session a single file to resume from, which matters when the work spans several agent sessions and context windows.

Step 1: Turn every point into a row

The agent read the document through a Google Docs connector, pulled every screenshot in document order and stored them locally (gitignored, so the doc stays the source of truth). Then it wrote one row per point, grouped by screen. Each row has five fields:

  • A number. Rows are never deleted, only ticked, so the numbers stay stable for commits and conversations.
  • What to change. One line, in plain words, including the reviewer's own phrasing where it helps.
  • The screenshot. The shot number from the doc, linked to the local copy.
  • A batch. Which kind of fix it is (more on that below).
  • A status. Open, or done with the date and whether it is committed, pushed or live.

Two real rows show the range. Row 11 is tiny: the documents page said "1 resumes", a plural bug, and the fix is one line. Row 19 is a whole investigation: wait times that under-promise on three different screens, with the reviewer's own measurements attached. Both get the same five fields, so a one-minute fix and a two-hour fix sit in the same list and move to Done the same way.

Each screen's group ends with the code paths the agent expects to touch. The tracker labels those honestly as "file-level matches from a search, not read in depth", which is exactly the warning the next session needs.

The agent also added 6 rows the reviewer never wrote down, marked "seen in screenshot". One of them, a resume import that put "Darjeeling" (a district) in the Country field, led to a real bug: the importer took the second part of any location as the country, so "Zurich, ZH, Switzerland" became ZH and "New York, NY, USA" became NY. Every import had it, not just the reviewer's.

Step 2: Batch by kind of fix, not by screen

The 117 rows went into five batches, worked in order:

  1. Broken or wrong (30 rows). False text, two places that disagree, or anything that blocks the user.
  2. First run (27 rows). Telling a new user what to do next: the dashboard, empty pages, the first upload.
  3. Editor wording (26 rows). Tips and labels in the editor, the customizer and the ATS review.
  4. Tools pages (16 rows). Wording and links for signed-in visitors on the free tools.
  5. New features (18 rows). Bigger builds. These are the founder's call, not the agent's.

Batching by kind beats batching by screen for two reasons. Fixes of the same kind share context, so the agent works faster and more consistently when it fixes all the wording on one pass. And the order puts risk first: broken things before polish, polish before new features. The agent also proposed a first five rows to start with, and marked that list as its own suggestion rather than the reviewer's, which keeps the source of every opinion clear.

Step 3: Put the rules inside the tracker

At the bottom of the tracker sits a short section titled "For Claude sessions". Any new session that opens the file reads the rules along with the rows, so you never have to repeat them:

  • Before fixing a row, check it is still true in the code. The review is several days old.
  • Copy lives in three locale files (English, German and French). Change all three.
  • No typed prices. Prices come from the billing config.
  • Every fix commit ends with a Guardrail line: the metric that must not drop because of the change.
  • After a fix, move the row to Done with the date and its state: committed, pushed or live.

With the rules in the file, the founder's part of the day shrank to a handful of prompts: "dont ask me random questions lets go one by one", "keep going", "ok, fix it, as long it doesnt cost ai roundtrip" and "i test once all done, how many left". The quick math: 96 rows in about 8 hours is roughly one row every 5 minutes.

Step 4: Make the agent prove a row is still true

The most useful rule turned out to be the first one. Of the 99 rows in batches 1 to 4, the agent checked 3 and changed nothing, and it wrote down why:

  • "Education appears twice." The agent rebuilt the reviewer's resume as it appeared in the screenshots (A4, Calibri 10 pt, five jobs, two degrees) and ran it through the real paginator. Education showed once. The second copy in the page text was the paginator's off-screen measuring copy, which no user sees.
  • "Two different user counts." 17,857 resumes on one page and 20,996 users on another both came from the same stats file and measure different things. Both were true.
  • "The translator list covers the button." It opened the live page in light and dark mode and could not reproduce it.

It also pushed back on parts of rows. The reviewer asked for a card pop-up to stop reappearing and to wait until the profile was complete. The pop-up now stays closed once dismissed, but the agent skipped the second half and explained that the card is made from the resume in one tap, so there is nothing to wait for. You only see this pushback if you read the Done notes, so read them.

Step 5: Find the cause behind the complaint

One row was about wait times. The upload screen promised about 20 seconds and took 46. A resume roast promised about 28 seconds and took around 50. A smaller tool promised about 11 seconds and took 16.

Rather than just changing the numbers, the agent found why the 46 seconds happened: the upload parse takes about 15 seconds on the main model and 46.9 seconds when it falls back to the last backup model. The estimate was accurate for the normal path and wrong for the fallback. The fix replaced every single estimate with the seconds elapsed so far plus a band, such as "8s · usually under a minute", "usually under 30 seconds" or "usually a minute or two", in all three languages.

The general lesson: when a reviewer reports a symptom, ask the agent for the cause before the fix. A row that says "the wait is wrong" is often a row that says "one path is slow".

What stays with you

An agent can work through wording, empty states and small bugs at a pace no solo founder can match. It should not decide what gets built. Batch 5 in this tracker holds 18 new features, such as showing the plan and remaining credits in the sidebar, short how-to videos on the dashboard and a "Fix" button beside every ATS issue, and all of them wait for the founder to choose. Even inside the first-run batch, the rows that needed brand-new screens waited until the founder answered seven open design decisions in a short spec.

Testing also stays with you. The tracker says so at the top: nothing is on main or live yet. Commit messages and Done notes tell you what the agent believes it changed. Clicking through the app tells you whether it did.

How to run your own this week

  1. Get someone who did not build your app to use it for real, with their own account and data, and ask for a screenshot of every screen.
  2. Hand the doc to your agent and ask it to read the images too, then build a tracker so you can work through it one by one.
  3. Specify the row format: number, what to change, screenshot, batch, status, and file paths per screen. Ask for a mark on problems seen in screenshots but not written down.
  4. Batch by kind of fix: broken first, then first run, then wording, then new features that you pick yourself.
  5. Write the rules into the tracker: verify before fixing, every language, a guardrail line per commit, move rows to Done with their state.
  6. Say "keep going", then read the Done notes for pushback, and test before anything goes live.

Practical takeaways

  • A feedback doc is ordered for writing. A numbered tracker is ordered for fixing. Convert it before you start.
  • Give every row a screenshot, a batch and a status, and never delete rows.
  • Batch by the kind of fix so similar work shares context and risky items go first.
  • Put the working rules in the tracker file so every new agent session inherits them.
  • Make the agent prove a row is still true. Expect a few rows to be wrong, and read why.
  • Ask for the cause behind each symptom, not just a change to the text.
  • Keep new features and final testing for yourself.

Conclusion

The review that felt like too much turned into a list that shrank by one row every few minutes. The work was not clever prompting. It was a row format, five batches and a short rules section, plus a reviewer who did not build the app and therefore saw what its builder could not. Six of the 117 problems were only visible in their screenshots. Get someone like that to use your app, then let the tracker do the remembering.

The founder-side version of this story ran in the HelloBuilder newsletter. For a related lesson on checking what AI wrote against reality, read How to Check Every Number AI Wrote on Your Landing Page.

The Monday
edition

The week for builders in five minutes, written by Nishant.

Latest: One Reviewer Found 117 Problems in My App. Claude Fixed 96 in 8 Hours