I can't hand off AI work without it falling apart

I start the morning with a familiar request: a stakeholder needs the monthly competitor summary updated before the afternoon meeting. I know the AI chat that worked last time, so I find it, copy the prompt, swap in the latest source notes, and get a draft that is close enough to be useful.

The small win is real. In twenty minutes I have a better first draft than I would have produced from scratch. But then the hidden work starts. I check which numbers came from which source, rewrite the caveats, remove a confident claim that feels too strong, and adjust the tone so it sounds like something our team would actually send.

After lunch, a team member asks if they can take this task next month. I send them the prompt and talk through the basics, but I can feel the gaps as I explain it. I have not written down which sources count, what to do when the AI guesses, which format the stakeholder expects, or the standard I use before I call the answer good.

By late afternoon they try a version and send it back. It is not careless. It just misses the judgment I never gave them. I spend another half hour rewriting it, and the task quietly comes back to me.

What I wish I had is not another prompt library. I want a simple way to turn the good chat into a handoff pack: the inputs, one accepted example, the checks I use, and a small test that proves someone else can run it before I let go.