Most people use AI to look up answers. Ask a question, get back a vague paragraph, move on.

I think the real value is in treating it like a coworker. The kind who asks good questions, and helps you figure out what you missed.

But a coworker is only useful if you brief them properly. If you sent a new hire a spreadsheet with no context, you'll get a useless answer too. AI is exactly the same. The quality of what comes back is set almost entirely by the quality of what you put in.

The single biggest upgrade to your results is front-loading context before you ask for anything.

Try opening with the situation. Who you are, what the document is, what standard it has to meet, and what you actually want back. Then attach the file so it can read it before responding.

I want to illustrate this point with two examples from my own world.

A pFMEA is a risk-mitigation tool used in manufacturing to identify and correct potential failure modes in a process.

Here’s an example of a weak prompt: "Review this pFMEA."

It doesn't know your process, your scoring scale, or what good looks like, so you get generic filler.

Here’s a stronger prompt: "You're a senior quality engineer in medical device manufacturing. Attached is a pFMEA for production of a PCBA. Our severity, occurrence, and detection scales are 1 to 10, and we action anything with an RPN above 60. Review it for three things: failure modes I haven't considered, any detection controls that look too weak for their severity, and rows where the scoring looks inconsistent with the rest. Point to specific line items."

Now it has a role, the standard, the threshold, and a clear job. The output looks more like an actual review and less like generic slop.

Corrective and preventive action (CAPA) is a regulated tool used to eliminate defects or undesirable situations in a manufacturing facility. They are used frequently, but require careful review given their importance.

Weak prompt: "Is this CAPA okay?"

Strong prompt: "You're an auditor reviewing this CAPA before an FDA inspection. Read the attached. Tell me if the root cause analysis is thin, whether the corrective actions actually address the stated root cause or just the symptom, and whether the effectiveness check would genuinely prove the problem is gone. Be skeptical. Flag anything an auditor could catch or question."

"Be skeptical" matters. Left alone, these tools tend to agree with you. Tell it to find holes and it becomes useful instead of flattering.

As a simple rule of thumb, give it a role, describe the context, define the goal, and name the expected output. Role, context, goal, task.

If you aren’t getting back what you expect, try giving more context and defining the goal more clearly. It can be frustrating, but AI can be useful with the right setup.

Got a second? Give some feedback on today’s article so we can keep making improvements to The Manifold.

PS — I find the usefulness of AI to be largely determined by how clear I am with my request. If I make generic requests, I get generic results.