How to get honest AI feedback on your novel: AI beta reader prompts that skip the flattery

Getting AI feedback on a novel means pasting your manuscript into an AI chat like ChatGPT, Gemini, or Claude and asking what works, what doesn't, and what to fix. Send just “give me feedback” and you'll usually get praise first and a generous score, so decide whose eyes it should read with, what it should look for, and what shape the answer should take. Then you get notes you can actually revise from.

This post covers why AI feedback leans toward praise, what to decide before you ask, six copy-and-paste prompts that turn an AI into a usable beta reader, and how to sort the notes you get back. The example is chapter 1 of They Say I'm a Dark Mage, but All I'm Good At Is Digging Dirt, the story from TypeTak's course Write a Web Novel, Step by Step, a Korean web novel written in Korean.

At a glance

AI chats tend to tell the person they're talking to what that person wants to hear, so an open question gets praise first. When you ask for feedback, pick whose eyes it reads with (a reader, an editor, a continuity checker) and make it quote your actual sentence for every note. Positions like “where I stopped reading” are easier to revise from than scores. Run the same prompt again in a fresh chat and start with the notes that show up both times.

Why AI feedback leans toward praise

AI chats are tuned to give more of the answers people rate highly, and that pushes them toward agreeing with you. The companies building them treat this as a problem too. In April 2025, OpenAI rolled back a ChatGPT update that had made it overly flattering and published an explanation (OpenAI). Hand an AI your chapter and ask “what do you think?” and it leads with what the author hopes to hear.

Scores are just as shaky. Paste the same chapter into a new chat and the number can change, and unless you define what each score means, there's no telling what separates a 72 from an 85. The AI also doesn't know what readers on your platform like, or what a scene will set up fifty chapters later.

AI feedback is still worth having. It'll read your draft as many times as you like, as if for the first time, and it's good at catching the repetitions and continuity slips you've stopped seeing. Know which way it leans and build your questions around that.

Three things to decide before you ask

One more helps a lot: what you did on purpose. In chapter 1 of They Say I'm a Dark Mage, but All I'm Good At Is Digging Dirt, an old judging stone misreads “earth” as “dark,” and that's the premise of the whole story. (In Korean, “earth magic” and “dark magic” sound the same, and that's the whole joke.) Leave it out and the AI may list “a judging stone misreading an affinity is implausible” as a weakness. Say what's intentional up front and the AI takes the premise as given and reads the rest.

Six AI beta reader prompts

Paste one chapter at a time. Several chapters at once tend to get blurred into a handful of broad notes. Number your paragraphs so you can find what each note points to.

1. A first-time reader's reactions

You read a lot of (fantasy web serials). Read the chapter below as if for the first time and log your reactions as you go.

- Where you wanted to know what happens next
- Where you got bored and wanted to skip
- Where you had to reread to understand
- Where you wanted to stop reading (if anywhere)

For each, give the paragraph number and quote one sentence. One line on why.
No praise and no overall review.

--- Chapter ---
(chapter)

This asks for reactions instead of a verdict. “I got bored here” points to the spot to fix more precisely than “good” or “bad.” If a “wanted to stop reading” shows up, start revising from the paragraph before it.

2. Three weaknesses, no praise

Name only the three weaknesses in the chapter below that most need fixing first. No praise.

For each:
- Quote the sentence from the chapter, verbatim
- Why it's a weakness, in two lines or fewer
- The direction of the fix, in one line (no rewritten sentences)

Don't count anything I did on purpose as a weakness:
(e.g., the judging stone misreading “earth” as “dark” is the premise of the story)

--- Chapter ---
(chapter)

Capping it at three makes the AI pick the big problems instead of listing small ones first. Leaving out rewrites keeps the direction of the fix in your hands. When you do want replacement sentences, follow up with the revision prompt in AI Writing Prompts for Fiction.

3. Web serial chapter 1 check

Below is chapter 1 of a web serial. Check it through the eyes of a reader deciding from chapter 1 alone whether to follow the series.

1. Does the opening make clear who the protagonist is and what the problem is? In which paragraph?
2. What question makes the reader want to know more, and in which paragraph does it first appear?
3. Does the last scene cut off in the middle of the action, or does it wrap up?
4. Does any explanation run longer than three paragraphs?

Answer each with “yes / no / unclear” first, then quote the sentence that shows it.

--- Chapter ---
(chapter)

Fixed questions keep the AI from drifting into compliments about the prose. This story's chapter 1 cuts off at the moment Clay, looked down on as a commoner, is taken for a dark mage, on a line that opens the next chapter's misunderstanding. Question 3 checks that a cliffhanger like that is where it should be.

4. Conflicts with your lore

Compare the story summary and the chapter below and find only the places where the chapter conflicts with the summary. Don't judge the prose or how fun it is.

For each conflict:
- The sentence from the chapter
- The lore it conflicts with
- “certain” or “possible”

If there are none, answer “none” and nothing else.

--- Story summary ---
(characters' voices, powers, relationships; the lore to keep)

--- Chapter ---
(chapter)

This one earns its keep as the chapters add up. It catches a character in chapter 30 using a different form of address than in chapter 3, or doing something you ruled out. The “answer none” line keeps the AI from forcing a finding. The story summary template is in AI Writing Prompts for Fiction.

5. Which of two versions

I wrote the same scene two ways. Pick the better one through the eyes of (a reader).

- Choose one first
- Give your reasons in three lines or fewer, quoting the text
- If the other version has one sentence worth keeping, name it
No scores.

--- A ---
(version A)

--- B ---
(version B)

Asked to judge one draft, the AI leans toward praise. Asked to choose between two, one of them has to lose, which makes this the prompt for when you've written a scene twice and can't decide. Swap the order of A and B and run it again: if it picks the same one, trust it more.

6. When you really want a score

Score the chapter below on these criteria. For each, quote the evidence first and give the score last.

[Criteria]
Opening: 5 = the protagonist and the problem are clear on the first screen / 3 = clear, but late / 1 = never clear
Characters: 5 = you can tell who's speaking from the line alone / 3 = only the protagonist is distinct / 1 = no one is distinct
Ending: 5 = cuts off mid-action / 3 = cuts off but adds a preview / 1 = wraps up

If the case for a higher score is unclear, pick the lower one.

--- Chapter ---
(chapter)

A score only means something with criteria behind it. Write down what 1, 3, and 5 mean, and you can score the revised chapter against the same bar and see whether it improved. Change the criteria to suit your story.

How to sort the notes you get back

Take every note an AI gives you and your draft drifts toward the AI's taste. Sort them in this order.

1. Set aside notes with no quoted sentence. “The tension feels low overall” doesn't tell you where to fix anything.
2. Run it again in a fresh chat. Same chapter, same prompt. Start with the notes that come up both times; a note that shows up once may just be that run's answer.
3. Check whether it's something you meant. If a note targets a deliberate choice, whether to change it is your call. Next time, put that intention in the prompt.
4. Compare with human readers. Where the AI's notes overlap with what a critique group or your comment section said, fix those first.

Feedback in TypeTak

TypeTak is an AI writing editor where you write alongside an agent. It does what the prompts in this guide were for, hearing what to fix instead of praise, with one button.

• What to fix, criterion by criterion. Press Feedback on the editor toolbar and it reads the passage on screen and answers in this order: Overall impression, Evaluation by aspect, Strengths, Areas to improve. Each criterion in the evaluation gets a star rating and its reason.
• No lore summary to paste. What prompt 4 above does is the job of the Consistency card. It checks the passage against the lore you've written and your Writing Style, with conflicts at the top and what you got right below them.
• Sentence habits shown on the manuscript. The Style card highlights repeated words and sentence habits right in the text.

The full walkthrough is in the Feedback guide.

FAQ