Recovery toolkit

Agents go wrong. The expensive part is not the mistake, but noticing it late and feeding the same broken chat for thirty more turns. This page shows the signs and four moves to recover.

Signs a chat is going wrong

• "Let me try a different approach", three times in a row. Each new approach is weaker than the last.
• "This should work", without a run that shows it. "Should" means it is guessing.
• The same files, read again and again. It has lost the thread.
• You correct the same mistake twice. It forgot what you said.

Move 1 · Reset

When: the agent goes in circles, but the task itself is fine.

How: start a fresh chat. Say the task in one sentence, plus what you learned: "Fix the card spacing in site/assets/styles.css. A grid on #projects did not work; put it on .projects." Leave out the failed attempts.

Cost: about two minutes. The cheapest move: use it earlier than feels comfortable.

Move 2 · Catch up with a checkpoint

When: in the workshop, the Block's time is up and your site is not there yet.

How: run the checkpoint of the Block that has just ended, in the terminal of Antigravity IDE. For the Developer that is:

Terminal · same on Mac, Windows and Linux

node tools/checkpoint.mjs developer

Cost: seconds. Your content file, design brief and spec stay yours. Each Role's page shows the command, in its card Fell behind?

Move 3 · A bigger model

When: the small model failed at the same task twice, and a reset did not help.

How: pick a bigger model under the chat box, start a fresh chat and give the same task. Switch back to Gemini 3.8 Flash afterwards.

Cost: quota. Worth it once you are stuck; wasteful as a habit.

Move 4 · Do it yourself

When: the fix is quick for you and slow for the agent, such as a typo in your content.

How: open site/content.json in Antigravity IDE and change the text yourself. Then let the agent check your work:

Agent chat

I changed site/content.json by hand. Run node tools/check-content.mjs and tell me if it is still valid.

Cost: a minute of typing. The agent is good at checking, even where it struggles to write the fix.

Which move when

1. Is the task still the right one? If not, go back to the spec, not to a recovery move.
2. Is the agent going in circles? Reset.
3. Is the room starting the next Block? Catch up with the checkpoint.
4. Did the small model fail twice at the same task? Try a bigger model.
5. Would you fix it faster by hand? Do it yourself.