Before you begin
Open the existing project so your instruction is grounded in the current app.
Write the instruction from the user’s perspective, then state the exact result you want to inspect.
Write the instruction
A good prompt explains the desired result.
A production prompt also explains the unacceptable result.
Add constraints such as
Never invent a source or citation.
Never approve access when required information is missing.
Never send a message without showing the user a preview.
Never make a payment or account change without human approval.
Never use information outside the approved knowledge base.
Never expose one customer’s data to another customer.
Never hide uncertainty. Flag it.
Then describe the fallback
“If the system cannot complete the task confidently, show the missing information and route the case to [role].”
The fallback is what makes an AI workflow usable when the world does not match the happy path.
Common mistakes
Writing broad cautions such as “be safe.”
Failing to define the fallback when a constraint is triggered.
Allowing irreversible actions without approval.
Frequently asked questions
How long should a prompt be?
Long enough to define the user, job, input, rules, output, and important constraints. Clear structure matters more than word count.
Can I make several changes in one prompt?
Small related changes can be combined, but changes to different systems are easier to test and recover when requested separately.