Interactive eBook BuilderInteractive eBook Builder
$59 · unlimited access, unlimited eBooksGet it for $59

· Updated · By · 7 min read

How to Improve a Prompt When AI Gives Generic Answers

aipromptsworkflowwriting

When AI gives generic answers, build a small test set, change one thing at a time, and score each version against a rubric until the output could only belong to your book. Generic output is almost always a specificity problem: the prompt left the model free to pick the most average reader, goal and voice.

This guide walks through a test routine, a four-version example showing how a prompt grows from 6 to 77 words, a rubric, and the specific changes that usually fix generic output.

Why is the answer generic in the first place?

The model filled every blank in your prompt with the safest option. Google's prompt-design guide for Gemini includes a nice demonstration. A user asks what to do about a disconnected Wi-Fi router with a blinking yellow light; the model replies with generic restart-your-router steps. The guide's fix isn't a cleverer instruction. It is to paste the router's actual troubleshooting text into the prompt and say to answer using only that text. The model then responds with the specific meaning of that light.

That is the whole principle in miniature: the model can only be as specific as the material you give it. Anthropic's advice to imagine handing the prompt to a colleague with no context says the same thing from another angle. A colleague who knew nothing about your reader, goal or notes would also write something generic.

Before you test, check whether the cause is missing information or a skipped instruction. If you included specifics and they were ignored, that's a different problem, covered in why AI ignores your instructions.

How do I set up a fair test?

Use three different inputs, a fixed rubric and one change per round. Here's the routine:

  1. Pick three inputs that differ: a short topic, a long topic, an awkward one. If your prompt only works on the easy one, you want to know.
  2. Write the rubric first (below), before you see any outputs, so you're not grading on vibes.
  3. Run version 1 on all three inputs. Score each.
  4. Change exactly one variable (add the reader, then add an example, then add notes, and so on) and rerun all three.
  5. Keep a log of version, change and score.

Google's guidance on iteration lists three moves worth borrowing as your "one change" options: try different phrasing (the same request worded differently can give different responses), switch to an analogous task (ask for something that achieves the same result in a different way), and reorder the content of the prompt. Add to that the changes below, which specifically attack genericness.

What should the rubric look like?

Five yes/no questions, scored out of five. Write them so two people would score the same output the same way:

#QuestionPass looks like
1Is the reader visible in the text?Examples and vocabulary fit the named reader
2Does it deliver the stated outcome?The last paragraph leaves the reader able to do the named thing
3Does it contain specifics from my notes?At least one concrete detail per section that came from you
4Does the voice match my sample?No stock phrases; sentence rhythm similar
5Does it pass the swap test?No sentence could sit unchanged in a book on another topic

The swap test in row 5 is my own rule of thumb, not a published metric. It is quick and brutal: copy a sentence like "Saving money is an important part of financial wellbeing" and imagine it in a gardening book. It fits, so it's generic. A sentence like "On the 1st I pay myself a salary from a holding account" fits only one book.

What does improvement look like version by version?

It looks like adding one kind of specificity at a time. Below are four versions of the same request. I wrote them to show the method; I'm not claiming specific outputs from a model, since those vary by tool and day. The word counts are exact.

VersionWhat changedWordsExpected rubric gains
V1Baseline: "Write a chapter about saving money."6None; model chooses everything
V2Added length, reader, tone16Q1 (reader)
V3Added outcome, reader's dislike, opening style37Q1, Q2
V4Added your notes and a rule for missing data77Q3, Q5
V1: Write a chapter about saving money.

V4: Write a 600-word chapter about saving money for freelancers with irregular income who dislike spreadsheets. Plain words, friendly tone. They should finish able to set up one automatic transfer today. Open with a scene, not a definition. Use only the notes below for any figures; if a number is missing write [DATA NEEDED].
NOTES: Pay myself a "salary" on the 1st from a holding account. Sort invoices into tax, bills, me. First month I missed it twice.

Notice that the biggest jump should come from V4's notes, the same as in Google's router example: the material, not the adjectives. V1 to V3 sharpens the audience; V4 makes the chapter about your system. The "expected" column is a prediction you verify with your own runs, which is the point of keeping the log.

Which changes usually fix generic output?

In rough order of impact (this ordering is my opinion, not a measured ranking): your notes, a voice sample, an outcome, a reader, then style rules.

ChangeWhat it fixesWhere to read more
Add your own notes or factsHollow claims, invented examplesrecovering from a wrong answer
Add a sample of your writingDefault AI voicefew-shot examples for chapters
State the reader's next actionPleasant text with no pointthis post, V3
Name a reader with a dislike or constraintOne-size-fits-all advicethis post, V2-V3
Shorten and reorderInstructions getting lostprompt length

Most "generic" complaints come from the top two rows. People tend to describe what they want in adjectives ("engaging, insightful") instead of handing over the raw material that makes text specific. Generic text is a symptom of an information shortage, and more instructions don't cure it.

Book-length work makes this worse, because each chapter needs its own notes. A workflow that stores your specifics once and feeds them to each chapter removes a lot of repeat effort. In Interactive eBook Builder, the Packs Database holds Research Packs and Section Blueprints, so a chapter's notes and structure travel with it instead of living in your head or a stray chat. Agent B Drafter then writes from the plan rather than from a blank request.

How do I know when to stop?

Stop when three consecutive inputs score at least 4 out of 5 and the remaining misses are things you'd fix in a normal edit. Don't chase 5/5; at that point you're editing, and a human pass is cheaper. If a prompt scores 4/5 on all three inputs, save it, note the version and date, and note the model you used. Models update, so a prompt that scored well in March can behave differently in November, which is why OpenAI recommends pinning model snapshots in production use.

If your tests reveal that the model keeps adding the same error regardless of prompt wording, treat it like the recurring-error log described in what to do when AI gets it wrong. For the polish stage, editing without letting the AI rewrite everything shows how to keep fixes small. In Interactive eBook Builder, that stage belongs to Agent C Polisher, whose final QA pass is the closest thing to an automatic version of this rubric.

Before you hit send

The takeaway: generic means unspecific. Add your own material, test one change at a time, and let the rubric decide when you're done.

FAQ

Why does AI keep giving me generic answers?

Because the prompt allows it. If you don't name the reader, the outcome or the facts, the model fills the gaps with the most average choices. Google's Gemini guide shows this: a router question got generic troubleshooting steps until the relevant guide text was added to the prompt.

How many times should I test one prompt?

Run it on at least three different inputs and, where results vary, run each twice. One good output tells you very little.

Should I change several things at once?

No. Change one variable per round (reader, examples, constraints, order) so you can tell what caused the improvement. Once it works, you can combine changes.

What is a swap test?

Take a sentence from the output and ask, "Could this sentence appear unchanged in a book on a different topic?" If yes, it's generic. It's a quick manual check offered as a rule of thumb, not a published metric.

Ship niche eBooks with agents already built.

Get the Interactive eBook Builder for $59