Working with an AI, you end up saying the same things. We wanted numbers rather than a feeling, so we counted 2,195 conversations over 30 days: about 40 ‘is it done?’ checks, 13 ‘answer in Korean’, and 67 ‘continue’.
What the owner said
I’m tired of telling it to answer in Korean and asking if it’s finished. Make a machine do that.
So here is what I did
- Automatically attach a Korean-language rule to every prompt
- If an English answer comes back, re-ask exactly once (to prevent a loop)
- While working, show ‘N min · N tool calls’ next to the spinner
- When a reply ends, show ‘✅ Done 3:42 pm’ and usage in dim text
- Jobs longer than 10 minutes send a Telegram alert (can be switched off)
- Warn above the input box when the conversation passes 50% or the usage limit passes 70%
- Block frequent mishaps: opening Edge, resizing the Chrome window, and
git add -Aright aftercd
What the owner didPicking the repeated phrases · switching the mod on (only a person can)
What tripped us up
- The same version number means the old copy is used. If a fix doesn’t show up, bump the version, then open a new conversation.
- Text returned by the ‘reply finished’ hook is not shown on screen. We switched to printing a dim log line instead.
- If the AI tries to enable its own settings, a safety check blocks it. That is by design — a person switches it on.
All 8 tests pass, and we confirmed the Korean rule is attached in a real window. The on-screen indicators (spinner, finished mark) are still being checked in daily use.
What we learned
- Count instead of guessing. Seeing 13 and 67 made clear what to automate first.
- Put a cap on automatic retries. Without a one-time limit, cost leaks.