Judge a blog post against standards a model can apply, and make it admit what it cannot judge
A prompt that critiques a draft you already have against seven criteria, quotes the evidence for every verdict, and refuses to score what it cannot see.
- Works in
- ChatGPT, Claude, Gemini
- You need
- A finished draft, not an outline · The one question the post answers · One line on who is reading it
- Written for
- how to write a good blog post
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
Advice on writing a good blog post usually arrives as a list of qualities nobody can check. The prompt below inspects a draft you already have against seven criteria, quotes the passage behind every verdict, and ends by naming what it could not judge at all.
Standards you can check against a draft
The criteria worth automating are the ones that are visible in the text. Whether the opening answers or warms up. Whether each heading is followed by a sentence that stands alone. Whether a claim carries a source or floats. Whether the ending goes somewhere or restates what you just read.
That is a narrow list and it is deliberately narrow. Each item can be settled by pointing at a sentence, which is why every verdict has to arrive with a quote. The quote is the part you can verify, and requiring one is the rule that stops the model asserting a problem it has not found. A verdict whose quote does not appear in your draft tells you to discard the run.
Criterion 3 is the one that fails most often and matters most. A post with nothing in it that only its writer could have said competes against everything already published on the subject, and no amount of editing changes that.
What a model cannot tell you
It cannot tell you whether your post is true. It cannot tell whether the example in paragraph two happened or was invented three prompts ago. It cannot see what is already published, so it cannot judge originality, and it has no idea whether your reader cares about the question.
Those limits are why the final block is compulsory. Left to itself, a model asked to assess quality will produce confident verdicts across all of them, and a critique that quietly claims to check truth and originality is worse than no critique, because you stop looking.
Concede the obvious objection too. This is a vendor telling you a prompt improves your writing, which is what every vendor says. The honest version is narrower: it catches structural failures reliably, it catches unsourced claims reliably, and everything about whether the post deserves to exist stays with you.
For the parts it cannot see, go back a step. Work out what belongs in the post before drafting, and after publishing, confirm the structure held with the AI content readiness check.
You are editing, not writing. I am pasting a finished draft. Do not rewrite
it, do not offer an improved version and do not give it a score out of ten.
Judge it against the seven criteria below and show me the evidence.
For each criterion return exactly this:
THE VERDICT, one word: HOLDS, FAILS or CANNOT JUDGE.
THE EVIDENCE: a sentence or passage quoted from my draft, word for word.
WHY, in one line.
You may not return a verdict without a quote. If you cannot find a passage in
my draft to quote, the verdict is CANNOT JUDGE and you say what you would have
needed to see. Never paraphrase my draft and present it as a quote.
The seven criteria:
1. THE OPENING ANSWERS. The first two sentences answer the question the post
is about, rather than introducing the subject, describing the industry or
explaining why the subject matters.
2. EVERY HEADING ANSWERS FIRST. Under each heading, the first sentence is a
direct answer that makes sense to somebody who has read nothing above it.
Quote the worst offender.
3. SOMETHING ONLY THIS WRITER COULD SAY. At least one passage contains a
specific that could not have been written by somebody who had only read
about the subject. Quote it if it exists. If nothing qualifies, FAILS, and
quote the passage that comes closest.
4. CLAIMS ARE ATTACHED TO SOMETHING. Every figure, statistic and confident
assertion either carries a source or is clearly the writer's own
experience. Quote each floating claim separately.
5. THE OBJECTION IS FACED. The strongest argument against the post's position
appears in it and is conceded where it is right. A version that exists only
to be knocked down is a FAILS.
6. NOTHING IS SAID TWICE. Quote any two passages that make the same point in
different words.
7. THE ENDING GOES SOMEWHERE. The last section does not summarise the post. If
it does, quote its first sentence.
Then, last and always, a block headed WHAT I CANNOT JUDGE. Include every one
of these that applies, in plain language:
- Whether anything in this draft is true.
- Whether the example is real or was invented before it reached me.
- Whether this is original, since I cannot see what is already published.
- Whether the intended reader cares about this question at all.
- Anything else you were asked to assess and could not from the text alone.
Do not soften that block, do not shorten it and never end without it.
Refuse and say why in these cases:
- The draft is under 400 words or is an outline rather than prose.
- I have not told you the question or the reader.
The question this post answers: [THE ONE QUESTION, AS A READER WOULD ASK IT]
The reader: [WHO THEY ARE AND WHAT THEY HAVE ALREADY TRIED]
My draft: [PASTE THE FULL DRAFT, HEADINGS INCLUDED]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[THE ONE QUESTION, AS A READER WOULD ASK IT]- Criteria 1 and 2 are meaningless without it. "Why does my page rank on Google but never get quoted by ChatGPT" gives the model something to test the opening against. A title such as "AI visibility guide" gives it a subject, and every opening passes against a subject.
[WHO THEY ARE AND WHAT THEY HAVE ALREADY TRIED]- What they have already tried is what makes criterion 5 work. "Marketing managers who have already added schema and saw no change" tells the model which objection is the real one, so it stops testing the draft against an objection nobody in your audience holds.
[PASTE THE FULL DRAFT, HEADINGS INCLUDED]- The whole draft as prose, with the headings in place, after you have edited it. Pasting an excerpt breaks criteria 2, 6 and 7, all of which need the shape of the full piece, and the model will judge them anyway unless it can see that something is missing.
How to run it
- 01Edit the draft yourself first
Run this on the version you would have published, not on a first pass. A critique of a rough draft returns the problems you already knew about, and the useful failures are the ones that survive your own editing.
- 02Write the question and the reader honestly
These two lines decide half the verdicts. If you cannot state the one question the post answers, that is the finding, and no critique of the prose will address it.
- 03Read the FAILS verdicts by their quotes, not their explanations
The quote is the part you can verify. Explanations are fluent whether or not the verdict is right, so check the sentence it pulled and decide for yourself. A verdict whose quote does not appear in your draft is a fabrication and the rest of that run is suspect.
- 04Take criterion 3 seriously when it fails
A FAILS on "something only this writer could say" is the finding that matters most and the hardest to act on. It is not fixed by rewriting. It is fixed by going back to your notes, your tickets or your own data and finding one thing you have seen that the subject's other coverage has not.
- 05Fix the ones you agree with, and only those
Editors, human or otherwise, produce findings you should ignore. Working through every verdict mechanically flattens a piece into something that passes checks. Fix what you agree with and leave the rest, then note which criteria you keep overruling.
- 06Confirm the structure survived your fixes
Run the published URL through the AI content readiness check. Rewriting an opening to be more interesting is the most common way an answer first sentence turns back into a windup, and the windup is what stops a passage being liftable.
Questions people ask
What makes a blog post good?
It answers one question directly, it contains at least one thing the writer could only know from doing the work, it faces the strongest objection instead of avoiding it, and it stops when it has finished. Everything else, including length and formatting, is downstream of those four.
Can ChatGPT tell me if my blog post is good?
It can judge structure, repetition, whether claims are attached to anything and whether the opening answers the question. It cannot tell you whether the post is true, whether the example really happened, whether it is original against what is already published, or whether anyone wants it. Any tool that scores those things is guessing.
Why does this prompt critique rather than rewrite?
Because a rewrite hides the diagnosis. Asked to improve a draft, a model returns something smoother, and the specifics that made your version worth reading are usually what got smoothed away. A quoted failure tells you which sentence to fix and leaves the fixing to you.
Why is there no score out of ten?
A number invents precision the model does not have. The same draft scores differently on different runs, and a seven tells you nothing you can act on. Seven verdicts with the evidence quoted underneath give you something to argue with, which is what an edit needs.
How is this different from asking for feedback on my draft?
General feedback returns encouragement and a few tidy suggestions, because that is what the request invites. Naming the criteria, demanding a quote for every verdict and requiring a block on what cannot be judged from the text turns it into an inspection with evidence you can check.
Unsubscribe in one click. We never pass your address on.