Before the prompt, fix the file.
When an AI draft misses the mark, it is tempting to rewrite the prompt. First, I want to know what the model was given.
Start with what is missing.
Consider a reply to a property owner. The draft may sound professional and still miss the point. Perhaps the owner already answered the question yesterday. Perhaps the number in the draft came from an old listing. Perhaps someone on the team promised to call instead.
The model cannot recover a conversation it never received. Before adjusting tone, check whether the system has the correct thread, property, current facts, and next decision.
Keep the source beside the claim.
An asking price supplied by an owner is different from an estimated value. A current requirement is different from an assumption about what a buyer might want. Those distinctions need to survive every summary and every handoff.
I want to be able to ask where a fact came from and get a useful answer. That means storing the source with the fact, preserving uncertainty, and making corrections part of the record.
Then work on the writing.
Once the context is right, the writing task becomes smaller. The model can draft a response to a specific situation, using facts someone can check. A reviewer can spend their time on judgment instead of reconstructing the entire conversation.
That is the sequence I care about: establish the record, identify the decision, then write.