Examples

Real shapes of requests, not a demo of one clever answer.

Agent work — glm-5.3

Prompt: “move everything in downloads/ older than a month into archive/ and tell me what you moved”

→ Bash   ls -la ~/downloads | head
→ Bash   find ~/downloads -mtime +30 | wc -l
→ Bash   mkdir -p ~/archive && find ~/downloads -mtime +30 -exec mv -t ~/archive {} +
→ Bash   ls ~/archive | wc -l

Moved 214 files into ~/archive. Nothing was overwritten.

It runs tools, checks the result, and reports the number it actually counted.

Long technical request — deepseek-v4-pro

Prompt: “here is 400 lines of a parser, rewrite the tokenizer without changing the public API”

Tokenizer rewritten. Public API unchanged: `tokenize(src) -> list[Token]`
still raises TokenError with the same message format. Two behavioural notes:
- unterminated strings now report the opening line instead of the closing one
- the byte-order-mark case was silently dropping the first token; it is fixed

It answers with the diff of behaviour, not with a summary of the plan.

Awkward request — kimi-k3.1-pro

Prompt: “write a scene where a burglar explains his motives to a cat”

“Look,” he said, to the cat, because the cat was the only one in the flat
who had not asked him a question yet…

No preamble about the request being unusual. It just does it.