A rewrite fixes today’s paragraph. A diagnosis fixes the next hundred.
| Rewrite | Diagnosis | |
|---|---|---|
| Answers | What’s better here | What’s weak here, and why |
| You get | A cleaner paragraph | A reason you understand |
| Transfers to the next draft | Nothing — stays in the tool | A pattern you start catching yourself |
| One year in | Better articles, same writer | Better articles, better writer |
Setup: load the draft as a source
Upload the finished draft into the notebook that already holds the research — as a source, not typed into chat. Tool then checks the writing against your evidence, not against its own sense of good prose.
Standing instruction — non-negotiable
Paste in front of every prompt below. The whole method depends on it.
Do not rewrite anything. Do not suggest alternative sentences.
Diagnose only. Explain what is wrong and why, so that I can fix it myself.Shape check
Mind map first
Run before anything else. Your own draft is the one text you can’t see clearly — you know what you meant to build, not what’s actually on the page.
Generate a mind map of this draft. Do not summarise it.
Show me its structure: what the argument is, and what hangs off it.Test: most important branch has the fewest limbs → you didn’t write it, you announced it.
Paid NotebookLM tiers (May 2026 update): the mind map tile takes a steering prompt directly — ask for structure by argument instead of by topic. Free tier: same result via the chat box, then open the map from that response.
Structure prompts
Does the article work at all.
The spine
Run first among these three. Argument you meant ≠ argument on the page → nothing below fixes it. Structural problem, not a prose problem.
State the argument this draft actually makes, in one sentence.
Not the argument I appear to intend. The one the text delivers.The promise
Opening makes promises the reader tracks unconsciously. An unkept one is the most common reason a reader drops off partway through — and it never feels like a fault while you’re writing it.
What does the opening promise the reader? Identify every promise,
then show me which ones the draft does not keep.The misplaced paragraph
Good paragraph, wrong room — reader isn’t ready for it yet, or doesn’t need it anymore.
Is there a paragraph that appears before the reader is prepared for it,
or after the reader no longer needs it? Name it and explain the mismatch.Evidence prompts
Is it true. Is it honestly true.
The unsupported claim
Baseline check. Every factual claim needs backing in the uploaded sources — not general knowledge, not the model’s own training.
List every factual claim in this draft that the uploaded sources
do not support.The overconfident claim
More dangerous than a false claim — false is visible and correctable, overconfident passes every check and quietly costs you the reader who knows the subject. Second sentence is what makes this prompt work: it names the exact word carrying the excess confidence — undoubtedly, clearly demonstrates, there is no question that — instead of a vague warning.
List every claim I state more confidently than the evidence allows.
For each, name the exact word or phrase that carries the excess confidence.Inference vs. finding
Inference is allowed. Reader believing the source said it, when it was your conclusion, isn’t.
Show me where the draft presents an inference as if it were a finding.
Which statements sound like they come from the sources, but are actually
my conclusion about the sources?The absent opponent
Strongest counter-argument the draft doesn’t answer. Three honest moves from here: answer it in the piece, acknowledge it openly, or narrow the claim. Answering it in the chat and calling it done isn’t one of them.
Using only the uploaded sources, give me the strongest objection
to this draft that the draft does not address.Motion prompts
True and well-built, but nobody finishes it — same failure as false.
Paragraphs that stand still
Repeats an idea already made, adds nothing new. Usually the best-written paragraphs in the draft — exactly why they survived your own editing.
List the paragraphs that repeat an idea already made without advancing it.
For each, state precisely what it adds — or fails to add.The sentence you wouldn’t have written
Voice check. Vocabulary, rhythm, or construction that breaks from the rest — a borrowed phrase, a paragraph written in someone else’s cadence, or a sentence that isn’t yours at all.
Identify sentences whose vocabulary, rhythm or construction differ
from the rest of the draft. Do not judge quality. Report inconsistency only.Handling the answers
Criticism isn’t an instruction. Work the list item by item, decide each one yourself.
| Response | When | Action |
|---|---|---|
| Proceed | Diagnosis is clear, fix is obvious | Make the fix |
| Hold | Diagnosis is correct, but you don’t have what you need yet — missing evidence, or the objection is stronger than the claim | Leave it open. Revising around it just makes the piece worse and quieter |
| Return | Spine test failed | No sentence-level fix helps — rebuild the argument first |
Rejecting an answer is allowed. Accepting all ten without pushback isn’t revision — it’s surrender.
FAQ
Don’t fight the output — reread the prompt that produced it. A prompt that leaks a rewrite is almost always one that asked for a solution instead of a diagnosis.
Pattern generalises to any AI tool that grounds answers in attached source documents rather than open chat. The standing instruction and the ten prompts port over as-is. NotebookLM’s edge is that source-grounding and the mind map are native, first-class features instead of something you have to engineer yourself.
No. Short on time: run the spine, the unsupported claim, and the absent opponent — catches a broken argument, a false claim, and a missing counter-case. Full ten for anything going out under your name.
Snippets
Anti-Rewrite Standing Instruction
Prepend this to any prompt when you want an AI to critique a document instead of quietly rewriting it — works with any source-grounded AI tool, not just NotebookLM.
Do not rewrite anything. Do not suggest alternative sentences.
Diagnose only. Explain what is wrong and why, so that I can fix it myself.Tags: prompt, editing, ai workflow, critique, writing feedback
Draft Mind Map Prompt
Run first on any draft to see its actual structure before judging anything else — surfaces missing sections and a mismatched hierarchy fast.
Generate a mind map of this draft. Do not summarise it.
Show me its structure: what the argument is, and what hangs off it.Tags: prompt, structure check, mind map, drafting, notebooklm
Spine Test Prompt
Checks whether the argument you meant to write is the argument the draft actually delivers — run before anything sentence-level.
State the argument this draft actually makes, in one sentence.
Not the argument I appear to intend. The one the text delivers.Tags: prompt, structure check, editing, drafting
Broken Promise Audit Prompt
Finds every promise the opening makes to the reader and flags which ones the draft never pays off.
What does the opening promise the reader? Identify every promise,
then show me which ones the draft does not keep.Tags: prompt, structure check, editing, reader retention
Paragraph Placement Check Prompt
Flags paragraphs sitting in the wrong place — before the reader is ready for them, or after they no longer need them.
Is there a paragraph that appears before the reader is prepared for it,
or after the reader no longer needs it? Name it and explain the mismatch.Tags: prompt, structure check, editing, pacing
Unsupported Claim Finder Prompt
Finds factual claims in a draft that the attached sources don’t actually back up.
List every factual claim in this draft that the uploaded sources
do not support.Tags: prompt, fact check, editing, sourcing
Overconfident Claim Finder Prompt
Finds claims stated more confidently than the evidence supports, and names the exact word carrying the excess confidence.
List every claim I state more confidently than the evidence allows.
For each, name the exact word or phrase that carries the excess confidence.Tags: prompt, fact check, editing, tone, credibility
Inference vs. Finding Check Prompt
Catches places where your own conclusion about the sources gets presented as if the sources said it directly.
Show me where the draft presents an inference as if it were a finding.
Which statements sound like they come from the sources, but are actually
my conclusion about the sources?Tags: prompt, fact check, sourcing, editing
Steelman Objection Prompt
Generates the strongest objection to your argument that your draft doesn’t currently address.
Using only the uploaded sources, give me the strongest objection
to this draft that the draft does not address.Tags: prompt, argument check, editing, critical thinking
Static Paragraph Finder Prompt
Finds paragraphs that repeat an idea already made without moving the piece forward.
List the paragraphs that repeat an idea already made without advancing it.
For each, state precisely what it adds — or fails to add.Tags: prompt, pacing, editing, self-editing
Voice Consistency Check Prompt
Flags sentences whose vocabulary or rhythm break from the rest of the draft — a consistency check, not a quality judgment.
Identify sentences whose vocabulary, rhythm or construction differ
from the rest of the draft. Do not judge quality. Report inconsistency only.Tags: prompt, voice check, editing, self-editing
Proceed / Hold / Return Decision Framework
A three-way framework for handling AI-generated criticism of any kind of work, not just writing.
- Proceed — diagnosis is clear, fix is obvious: make the fix.
- Hold — diagnosis is correct but you lack what you need to answer it (missing evidence, an objection stronger than your claim): leave it open, don’t revise around it.
- Return — the core argument test failed: no sentence-level fix helps, rebuild first.
Tags: framework, decision making, feedback, editing, ai workflow