Tell it what you already tried.

A sharp AI ask can still loop if you never say what already failed. Tell AI what you already tried, or it reheats the same idea in nicer words.

Prompt Fluency is communicating so the model does not have to guess. Failed attempts are context. You lived them. The model did not. Leave them out and it invents a "fresh" answer that walks the same dead path.

Try this ask:

"Rewrite the package room hours note for the building lobby."

You get a clearer version of the note that already did not work. Longer hours. A softer tone. A polite "please check with the front desk." You already tried that. Residents still missed the cutoff. You needed a rewrite that closed the old loophole, not a polish pass on the failure.

Same break shows up everywhere.

"Write a carpool text for Thursday school band practice."

"Draft the thrift store shift briefing for new volunteers."

"Rewrite the bike co-op tool checkout rules."

Each ask names a job. None of them name the dead ends. So the model picks the most common helpful path. The path you already walked. The path that already missed.

Put what you already tried in the prompt. Plain English. Short. Early.

"Rewrite the package room hours note for the building lobby. Already tried: longer hours line, soft please-check-desk wording, and a sticky under the mailboxes. Those still got ignored. New note must make the 6pm cutoff impossible to miss and name the Saturday exception."

"Write a carpool text for Thursday school band practice. Already tried: a long guilt message about being late, and a joke about traffic. Both got ignored. Keep it under three sentences. Include pickup spot, time, and who has the spare seat."

"Draft the thrift store shift briefing for new volunteers. Already tried: a two-page handbook paste and a 'just ask if you have questions' closer. New people still sorted donations wrong. One page max. Cover hangers vs bins, what never goes on the floor, and who to find at 10am."

"Rewrite the bike co-op tool checkout rules. Already tried: 'be careful' language and a verbal reminder at the door. Tools still walked out unsigned. Rules must require name, tool, and return time on the clipboard before anything leaves the bench."

That "already tried" line is not you venting. It is you closing the biggest open guess in the prompt. Without it, the model optimizes for a generic good answer. With it, the model has to move past the answer you already rejected.

I am not asking for a diary. Name the two or three moves that failed. The soft tone. The longer version. The joke. The handbook dump. If you keep getting a smarter reprint of yesterday's miss, that miss belonged in the first prompt.

A quick check before you hit send:

Already tried vs "make it better"

• "Make it better" still lets the model invent the target
• "Already tried X and Y. Do not repeat those" closes those doors
• If every rewrite still smells like the first draft, you never named the failure
• If you could predict the bad answer before you hit send, that prediction belongs in the prompt

When the answer comes back polished and still useless, do not ask it to "try again." Name what already failed. Put that in the next prompt. Ask again.

You're not being difficult. You're removing the need to guess. That's Applied Intelligence Mastery in practice: tell it what you already tried, so the model stops reheating a path you already walked.