Errors Aren't Scary—Not Seeing Them Is
It's perfectly normal for the AI's first pass of code to not run, or its answer to have holes. The real problem is: it has no idea it's wrong—because the error never made it back onto its desk.
The Self-Correction Loop: Run → See Error → Fix → Run Again
The way to let the AI fix itself is just a loop: have it produce → actually run it → feed the error back verbatim → have it revise accordingly → run again… until it passes.
code = llm("写一个函数：判断回文")
for _ in range(3): # 最多试 3 轮
err = run(code) # 真的去跑
if err is None:
break # 通过了就停
code = llm(f"这段代码报错了：\n{err}\n请修复：\n{code}")Don't Let It Spin Forever
The loop is handy, but you need to give it a brake—set a cap on how many rounds it can try, or "stop after two rounds with no progress."
自测 · 学完检查一下
想真正动手做题、记进度、攒连胜？到互动课里练。
What's the core move of "let the AI fix itself by looking at the error"?
答案：Feed the post-run error back verbatim so it can revise accordingly
The key to self-correction is sending the real error back into its context, so it can see what's wrong and fix it.
Judge: if you don't feed the error back to the AI, it usually can't tell what was wrong with its last version.
答案：True
The AI can't see what isn't on the desk. Without the error fed back, it's like a student who never saw the corrections—it repeats the mistake.
Which feedback is most likely to help the AI actually fix the code?
答案："Error on line 8: variable total is undefined"
A specific, locatable error (which line, what error) lets the model fix precisely; vague blame provides no usable information.
The more ___ the feedback you give back (e.g. with line numbers and the exact error), the more precisely the AI can locate and fix the problem.
答案：specific
The more specific the feedback, the better the model can pinpoint the problem. A vague "wrong, rewrite" is almost no information at all.
Judge: a self-correction loop should keep running until the AI gets it right, with no cap set.
答案：False
You must set a stop condition (e.g. a max of N rounds). Otherwise the AI may spin forever between two errors, wasting time and money.
To keep the AI from spinning forever in the loop, set a stop condition, e.g. "try at most a few ___."
答案：rounds
Setting a cap on the loop (try at most a few rounds) is the essential brake that prevents an infinite loop from burning resources.