I spent way too long fixing prompts when the problem was the tool
i'd get a bad result and add another rule.
be more specific. check the sources. don't make assumptions. try again.
the prompt kept getting longer, but the result barely changed.
then i looked at the tool the agent was using.
search_web accepted one vague query and dumped a pile of raw results back.
no date filters. no source requirements. no way to tell a weak result from a useful one. even the errors were vague.
what was the model supposed to do with that?
so i stopped rewriting the prompt and fixed the tool.
search results now come back with sources and dates. tools that change something show a preview first. anything that sends, deletes or publishes waits for approval.
the agent didn't need another paragraph of instructions.
it needed something decent to work with.
now when an agent gives me a bad result, i don't just check what i told it. i check what i gave it.