Stephan Schmidt - March 22, 2026

Claude just got a new superpower

Claude got a superpower and no one recognized it.

It’s /loop.

You now can run a prompt every 5 minutes.

You might think, “So what?” a way to repeat prompts. Nothing earth shattering.

I look at this differently.

The one thing AIs are bad at is determinism. LLMs are probability machines.

If you ask an AI to search for bugs you get a list of bugs. If you ask the AI again, you get a different list. Doesn’t the AI know what they do? No they don’t - they are not thinking and planning engineers - they are probability machines.

This can be fought with /loop.

Let’s let Claude run a bug searching prompt every 10 minutes.

/loop 10min "deep search src/ for bugs. add to potential_bugs.md"

First run:

• Done. Created potential_bugs.md with 12 unique potential bugs found

Second run:

• Found 5 new bugs (1 was a duplicate of an existing entry). Let me append them to potential_bugs.md.

Third run:

• The search is converging — most code paths have now been covered across 3 scans.

Fifth run:

• No new bugs found this scan — all 3 results were duplicates of already-known issues (#4, #17, #18). potential_bugs.md remains at 22 bugs. The codebase has been thoroughly covered across 5 scans. Further scans are unlikely to surface new findings without code changes.

Then:

• No new bugs found. After 6 thorough scans covering all .go files in src/ and its subdirectories, the search has converged.

This feels a lot like the casino method of algorithms.

The results have been converging after five runs.

With the list, finally one would run a prompt to evaluate and verify the bugs, rank them, write tests and fix them.

Incidentally there is an informal ‘Rule of 5’. Xuezhi Wang and the team at Google Research wrote in their 2022 paper on ‘Self-Consistency,’ running a prompt multiple times leads to better results.

This goes way beyond bug hunting. Every analysis query that you use an AI for, finding architecture problems, security flaws, requirements from a codebase now can easily be made much more powerful.

Just /loop !