Your first version of any prompt will produce 70%-good output. Your fifth version will produce 95%-good output. The difference is iteration discipline. Properties that "tried AI and it didn't work" almost always stopped at version one.
Step 1: run and judge
Run the prompt with three different inputs that represent the range of cases you actually deal with — a five-star review, a one-star review, a mixed review. Look at the output. Ask: would I send this exactly as written? If yes, the prompt is shippable. If no, identify specifically what is wrong: too long, wrong tone, missed a key point in the input, generic-sounding, made up a fact.
Be specific. "The output is bad" is not feedback. "The output uses the word 'fantastic' three times in a single paragraph" is feedback the next iteration can act on.
Step 2: edit the prompt, not the output
The instinct is to edit the AI's output until it is shippable. That works once. The discipline is to edit the PROMPT so that the next run produces shippable output without manual editing. If the response was too long, add "maximum 90 words." If the response missed the guest's specific concern, add "address the guest's specific concern in the first sentence." If the response used corporate clichés, add "never use the words crafted, curated, fresh, journey, or experience."
Each edit is small. Three to five edits compounded across iterations is the difference between a prompt that requires manual touching every time and a prompt that produces shippable output 9 times out of 10.
Step 3: ship and capture
Once the prompt produces shippable output reliably, save it in your template library with a name (e.g., "review_response_v3"), a date, and a one-sentence description of what use case it handles. When a team member needs to respond to a review, they pull "review_response_v3" from the library instead of writing a new prompt.
Every quarter, audit the templates. Three of every five will get further refined as you learn what the AI does well and badly. The library becomes a property asset — the same way a SOP library is a property asset. New hires get trained on the library; departing employees do not take their personal prompts with them.
AI gets useful when you stop treating each prompt as a one-off conversation and start treating prompts as reusable artifacts that the team improves together.