Break it on purpose
Most AI content shows the highlight reel. This page shows the errors — and the exact language that got out of them. The agent is interchangeable. The pattern is not.
Why document the failures?
Errors are the process
Grok, Codex, Claude Code — any of them will break a layout while fixing a button. That is expected. It is not a sign you picked the wrong religion.
The fix is a second brain
When one agent loops, stop punching the same bruise. Screenshot. Paste into a different AI. Bring the diagnosis back. Not “Claude builds, ChatGPT debugs” as law — whoever is sharp today.
Pattern recognition is the skill
After enough sessions you will know the class of bug before you finish reading the red text. This page is that practice, written down.
Common failures and the language that works
These patterns showed up in real builds. The prompt is the part to steal.
The circular fix loop
Type: Infinite loopThe agent fixes a CSS layout bug. The fix breaks the mobile nav. It fixes the nav. The original layout breaks. Repeat.
Context fills with failed attempts. Each new try is reasoned from a history of wrong turns. The model cannot see that it is circling.
Stop that agent. Take the error to a different AI:
"I'm getting [describe the error]. I've tried [brief description of fixes]. Each fix creates a new version of the same problem. What is the most likely root cause I'm not seeing? Give me 2–3 probable causes."
Then back to the builder: "Ignore your previous fixes. Based on fresh analysis, the root cause is [second AI's answer]. Approach it from that angle."
CSS doesn't apply / styles ignored
Type: StylingYou ask for a button color change. The CSS file is edited. The browser still shows the old color. The agent insists it made the change. It did. The browser does not care.
Cache. Specificity. Or the right file and the wrong selector.
Hard refresh: Ctrl+Shift+R (Windows) or Cmd+Shift+R (Mac). If that fails:
"The style change isn't showing even after hard refresh. Check that you edited the correct selector in the correct file. Check for competing rules with higher specificity."
Deploy fails (Firebase or otherwise)
Type: DeploymentA deploy command throws auth, permission, or “project not found.” The site is stuck on your machine.
Not logged in. Wrong project selected. Config file points at a different ID. CLI version mismatch.
Deploy is on the never-without-approval list. You still have to say yes. Then:
"Deploy is failing with this error: [paste the full message]. Check the project config, confirm I am logged into the right account, and show me the exact command before you run it."
The agent edits the wrong file
Type: Scope overreachYou ask to fix the homepage. It “helpfully” also rewrites three other pages. Working styles die.
Agents like to be useful. Adjacent files look related. Helpfulness without a leash is a bug.
"Fix only index.html. Do not touch any other files. Do not 'improve' anything I haven't asked about. Only make the specific change I describe."
If it already happened: "Revert any changes you made to files other than index.html."
Mobile layout breaks after a desktop fix
Type: ResponsiveDesktop looks fine. Phone: cards overlap, text clipped, hamburger gone.
A global display, flex, or position change stomped a breakpoint. The agent did not look at 320px.
"The desktop fix broke mobile. Check the @media (max-width: 768px) rules. Do not hide overflow with overflow-x: hidden on the page. Fix mobile without breaking desktop. Test at 320."
The formula
1. Stop and reset
“Stop your current approach.” / “Ignore the previous attempts.” Do not let it continue from a broken foundation.
2. State the specific error
Paste it. Screenshot it. In Grok Build on Windows, image paste is often Alt+V, not Ctrl+V — environment dependent.
3. Bring a new cause
From a second AI, or from you. That is the new starting point. Then you still approve the next edit.
The debug move, in context
Same project file. Different instrument when one loops. You keep the consequential buttons.