Skip to content
Published 5 min read

Claude couldn't fix your website: retry, switch, or stop?

Claude says the website is fixed, but the menu still won't open on a phone. Before you retry or switch models, name the failing visitor action, protect the live site, and judge every attempt by that same check.

A hand taps the menu icon on a phone showing a simple website, with three cards behind it for retrying, asking for a second opinion, and stopping.

Claude says the mobile menu is fixed. You open the site on your phone, tap the menu icon, and nothing happens. (This is an illustrative example, not a recorded incident, but it is a familiar shape.)

Write down the visitor action that still fails, keep the live site untouched, and retry with the evidence Claude was missing. If the task is well defined and still stuck, a bounded second opinion from another model is reasonable. If the blocker is access, payments, sensitive data, a destructive change or a result nobody can check, stop and hand it to a person. Whichever you pick, judge it by the same check you wrote before the attempt, which is what MDN's testing guide asks for once a fix lands [4]. A model switch is another attempt, not evidence that the website is fixed.

Write down the visitor action that still fails

A vague request can send the next retry in the wrong direction. Claude Code's own guidance asks you to supply what it cannot find in the repository and to give it a check with an outcome it can read, such as a test or a command's output [1]. MDN's testing guide makes the same move from the other side: collect the platform, device and browser, then narrow the failure down before you change anything [4].

"The menu is broken" meets neither standard. An agent cannot see your phone, and it cannot tell which tap you mean. Write one sentence for what a visitor should see and one for what they see now, then add the device and browser:

On an iPhone in Safari, tapping the menu icon should show five links. It shows nothing. No error message appears on the page.

That sentence is illustrative, but its parts are not optional: between them, the expected result, the observed result and the setup cover what those two guides ask for.

Paste the error text, a screenshot or the steps that reproduce the problem. Leave out passwords, API keys and customer records. None of them help diagnose a menu, and an agent's context is the wrong place to keep them.

Protect the live version before the next attempt

Do the next attempt somewhere a mistake cannot reach visitors. Claude Code's checkpoints help, but its documentation says plainly that they do not replace version control [2]. Rewinding can restore files Claude edited and the conversation around them. Changes made through shell commands or outside the session are normally not captured, with exceptions for files the session had already edited, so read a rewind as a convenience rather than a backup.

A rewind also will not undo an email that was sent, a record that was written or a setting changed on a live server. If your setup lets you isolate a candidate change from the live site, use that, so the retry can fail without anyone noticing but you.

Retry, switch or stop: sort the blocker first

Now the real fork. Several different blockers can produce the same "it didn't work", and by the reasoning in this guide (an inference, not a measured result), a different model plausibly helps with only one of them. Claude Code's permission settings decide what an agent is allowed to touch [3], and picking a different model does not change them.

What you seeLikely blockerNext moveDoes switching models help?Check that must pass
The fix touched the wrong element, or Claude asked what you meantUnclear requestRewrite the expected and observed result, retryNoThe written visitor action
Claude guessed at a cause it could not see (device, browser, error text)Missing evidenceAdd the device, steps and error output, retryNoThe written visitor action
A clear request, full evidence, careful clarified attempts, still failingBounded reasoning problemAsk a second model for a diagnosis, scoped to this one actionPossiblyThe same visitor action, plus neighbours
The fix needs a password, a payment setting, DNS or live dataMissing authorityStop and hand to the person who holds that accessNoTheir sign-off
Nobody can see the result, such as an enquiry that should arrive in an inboxMissing verificationStop until a test destination or owner can confirm itNoA confirmed delivery, or a written gap
The rows are an editorial decision guide built from the sources in this article, not a measured ranking. Notice how few rows end in a switch: changing models does not create missing permissions, credentials, data or a working way to check the result.

If the request lacks the device, the steps or the error, close that gap before any switch. Without them, the most capable model available receives the same missing evidence and is still guessing.

When is switching from Opus 5.5 to Fable 5.2 worth it?

Consider a bounded second opinion when the task is clear, the evidence and access are available, and careful attempts still fail. Outside those conditions, a switch hands the same missing information to a different model.

For context, here is what Anthropic says, dated to 16 November 2026, because model roles change. Its announcement recommends Claude Opus 5.5 for most coding and knowledge work and positions Claude Fable 5.2 for more ambitious, harder technical work. It also says medium is Fable 5.2's default effort in the Claude apps and Claude Code [6]. The API documentation lists a different default, high for Fable 5.2 and medium for Opus 5.5 [7]. So "the default" depends on where you are working.

Third-party coverage reads this the same way: Opus as the everyday starting point, Fable as the advisor you bring in [8].

The strongest case for switching is real. Anthropic positions Fable 5.2 for harder technical work [6], so on a genuinely hard bug a second model may notice something the earlier attempts missed. That is the third row of the table: the request is clear, the evidence is attached, the access is already in place, and careful attempts still fail. In that case, ask Fable 5.2 for a diagnosis of this one action, keep the scope narrow, and keep the same check.

Then treat its answer as a proposal. One Hacker News commenter described having Fable 5.2 review an Opus 5.5 project and said they did not know whether its findings were true [9]. That is one person's account, not a measurement, but it captures the right posture. A second opinion tells you where to look. It does not tell you the menu opens.

For the other rows, a different model meets the same gap. The missing information or authority is still missing, whichever model reads the request.

Did the new attempt fix it? Run the same check again

Repeat the exact action, on the same device and browser, before you look at anything else. MDN's guide calls for testing again after the fix and then checking other browsers and places for new breakage [4]. Playwright's guidance points the same way: test what the visitor sees and does, not how the code is arranged [5].

For the menu example, a pass looks like this:

  • On the iPhone in Safari, tapping the icon shows the five links.
  • Tapping "Contact" inside the menu reaches the contact page.
  • The desktop navigation still looks and works as before.
  • The menu also opens in one other mobile browser.

If the first line fails, the attempt failed, whatever the summary says. If the first passes and a later one fails, the fix moved the problem rather than removing it.

Does a successful build mean you can publish?

No. A successful build and a working visitor journey answer different questions.

Claude Code's guidance lists build exit codes as one useful check, then separately recommends running the app and looking at it, because some problems only show up in use [1]. A build tells you the code compiles. A screenshot tells you how one screen looked at one moment. Only the visitor action tells you the visitor can do the thing.

Forms are where this bites. A thank-you message on screen does not prove an enquiry was stored or delivered. If you can, send a test to a destination you control and confirm it arrived. If nobody can check, write that down as a gap instead of calling it done.

When to stop and hand the change to a person

Stop when the blocker is authority or verification. That covers payments, sensitive records, account access, DNS changes, anything that deletes content, a live outage, and any result nobody can check. Claude Code's permissions decide what the agent is allowed to touch [3]. A model switch does not change who is authorised to approve a payment setting or a domain change. That is a person's call, and the useful move is to name the owner of the next decision.

Give that person a short note, so they start from the same check you used. Six lines are enough:

text
Expected:   iPhone, Safari: tapping the menu icon shows five links
Observed:   nothing opens, no error on the page
Candidate:  change to the menu script, not yet on the live site
Blocker:    bounded reasoning (clarified attempts still fail)
Check:      same tap, same phone, then Contact link and desktop nav
Owner:      whoever approves changes to the live site

Keep that note beside your next request, and send it to whoever approves changes to your site. The next time a fix comes back "done", you will already know which tap decides it.

References

  1. Anthropic, "Best practices for Claude Code," Claude Code documentation. https://code.claude.com/docs/en/best-practices. Accessed 16 November 2026.
  2. Anthropic, "Checkpointing," Claude Code documentation. https://code.claude.com/docs/en/checkpointing. Accessed 16 November 2026.
  3. Anthropic, "Configure permissions," Claude Code documentation. https://code.claude.com/docs/en/permissions. Accessed 16 November 2026.
  4. MDN Web Docs, "Introduction to cross-browser testing," last modified 4 September 2026. https://developer.mozilla.org/en-US/docs/Learn_web_development/Extensions/Testing/Introduction. Accessed 16 November 2026.
  5. Playwright, "Best Practices." https://playwright.dev/docs/best-practices. Accessed 16 November 2026.
  6. Anthropic, "Claude Fable 5.2 and Mythos 5.2," published 9 November 2026. https://www.anthropic.com/claude-fable-and-mythos-5-2. Accessed 16 November 2026.
  7. Anthropic, "Claude Fable 5.2 overview," Claude Platform documentation. https://platform.claude.com/docs/en/models/fable-5-2/overview. Accessed 16 November 2026.
  8. Tom's Guide, launch coverage of Claude Fable 5.2 and its positioning against Opus 5.5, 10 November 2026. Third-party coverage. https://www.tomsguide.com/ai/anthropic-just-launched-its-most-powerful-claude-ahead-of-its-ipo-heres-why-its-telling-users-to-stick-with-opus. Accessed 16 November 2026.
  9. Hacker News comment thread, item 50436100. Anecdotal user report. https://news.ycombinator.com/item?id=50436100. Accessed 16 November 2026.