# Claude Code had worked on my codebase for weeks. Then I asked it to use the product like a buyer and it found bugs it had never noticed. — retained record

> Retained reddit record from reddit:ClaudeCode.

Canonical: https://highsignal.app/data/records/927ddfbaa666413b

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:ClaudeCode)  published 2026-09-14 08:51 ingested 2026-09-15 02:40

[original source ↗](https://www.reddit.com/r/ClaudeCode/comments/1wfy0ea/claude_code_had_worked_on_my_codebase_for_weeks/)

## retained text



```
I've been using Claude Code heavily on a FastAPI + Next.js project, so by the time I was preparing a release it knew the repo extremely well.

Tests passed, type checks passed, lint passed, build passed.

I then gave Claude what I thought was basically a documentation/marketing task: make the product walkthrough from a fresh generated project.

The instructions were roughly: start clean, follow the setup, boot the services, sign in through the real UI, exercise the main flows and record it.

It got stuck on the first real login.

Clerk redirected to routes that didn't exist.

Fixed that, tried again.

Then billing didn't work from the browser. Backend was fine, frontend was fine, but the shipped local CORS config didn't include the frontend origin.

Another thing none of the existing checks had exposed.

What I found interesting wasn't really the two bugs. Claude had already worked on this codebase a lot. It could inspect both of these areas whenever it wanted.

The only thing I'd really changed was the objective.

Normally I ask it things like:

* implement X
* review Y
* run the tests
* check whether Z is correct

This time the task was basically "use the software successfully from the beginning."

Same agent, same repo, different role.

It still had repo access, so I wouldn't call this a proper black-box test. But in a way that made the result more interesting to me. Apparently just forcing the agent through the user path was enough to expose things it hadn't noticed while reasoning about the code directly.

Since then I've been experimenting with this distinction:

coding agent: repo -> implementation

user agent: goal -> product -> result

I'm now wondering whether I should make the second one an explicit release step and restrict it harder, maybe browser-only with no repo access.

Someone in another discussion mentioned doing exactly that with persona-based browser agents: fresh user, impatient user, non-linear navigator, etc. That sounds pretty compelling.

For people doing this already, how are you separating implementation agents from acceptance/QA agents?

Same Claude session with a different prompt?

Separate subagent?

Fresh context?

Browser-only?

And do you let the QA agent see the repo at all?
```



document fetched 2026-09-15 00:42 · canonical https://reddit.com/r/ClaudeCode/comments/1wfy0ea/claude_code_had_worked_on_my_codebase_for_weeks
