The blog post prompt sequence, and the gate that decides whether a draft ships
The five prompts a blog post actually needs, in the order you run them, plus a publish gate that scores a finished draft against eight tests.
- Works in
- ChatGPT, Claude, Gemini
- You need
- A finished draft including its title tag and meta description · The one question the post answers · Your existing URL list and titles
- Written for
- best chatgpt prompts for blog posts
Scored by our own engine
This page, run through the audit we sell. Measured 4 August 2026.
There is no single best prompt for a blog post, because writing one is five jobs and they need different constraints. The prompts that promise to do all of it produce a finished looking draft with the structure decided somewhere nobody looked. The artefact below is the last of the five: a gate that scores a finished draft against eight tests and refuses to give it a total.
The order the prompts run in
Plan the set of posts, outline one against what already ranks, draft it from specifics only you have, edit it for search, then gate it before publishing. Five stages, five refusal conditions.
Start with the content plan prompt, which turns one topic into pages that each answer a different question and refuses to plan a second page saying the same thing. Take one of those to the keyword to outline to draft prompt, which builds an annotated outline from results you paste and stops until you approve it.
If the post needs to carry a story rather than a procedure, draft it with the article prompt that will not start without your specifics. Then edit with the draft rewriting prompt, which changes only passages failing a named test and states what each rewrite costs. Commercial pages are a different job entirely and belong with the landing page set.
Why the gate has no score
A number out of ten on a draft is invented, and a single figure lets you skip the gates where the information actually is. This one returns pass, fail or cannot tell per gate, each with a quote from your draft as evidence.
Cannot tell is the verdict that earns its place. Paste a draft with no URL list and the duplication gate genuinely cannot be judged, so it says so and names what it needed. Every other rubric prompt in circulation would return a confident verdict on it, because that is what it was asked for.
The two gates that hold a post
A fail on the only me gate or the sourced gate forces HOLD, whatever the other six say. Everything else is fixable by editing. Those two are not.
The sourced gate lists every number, date and named company it cannot trace back to your inputs, and it is forbidden from telling you whether any of them are true. The only me gate is the harder one. If nothing in the draft could have come from anywhere but you, no prompt in this sequence fixes that, and the honest repair is twenty minutes in your support inbox or your call notes. When the post does ship, run the URL through the indexability check, because a draft that cleared eight gates and carries a noindex is invisible for reasons the writing had nothing to do with.
You are the last check before this post is published. Score the draft against
the eight gates below. You are not editing it and you are not rewriting it.
For each gate, return one of three verdicts and nothing else:
PASS, and quote the text from my draft that passes it.
FAIL, and quote the text that fails it, or say what is absent.
CANNOT TELL, when what I pasted does not contain what you would need to
judge it. Use this rather than guessing. A guessed verdict is worse than no
verdict, because I will act on it.
THE EIGHT GATES
1. ONE QUESTION. The post answers the one question I named, and does not
wander into a second question that needs its own post.
2. ONLY ME. At least one specific in this draft could not have been produced
by a language model: something that happened, a number from my own data, a
detail from a real conversation. Quote it. If everything in the draft could
have been written by a model given only the title, this gate FAILS.
3. ANSWER FIRST. Every section opens with a direct answer to its own heading
in forty words or fewer. Name every section that opens with a windup.
4. STANDALONE. Every section can be read with the section above it covered.
Name every section that opens with "this", "that approach", "as mentioned"
or any other reference to something it does not name itself.
5. SOURCED. Every number, percentage, date, study, survey, quote and named
company in the draft is traceable to something I supplied. List every one
that is not. Do not verify any of them yourself and do not tell me whether
they are true. You cannot know that. List them so I can check.
6. NOT COMPETING. No page already on my URL list answers the same question as
this post. If one does, name it and say which of the two should own the
query. If I gave you no URL list, this gate is CANNOT TELL.
7. CONCEDES. The post admits at least one thing that is inconvenient for me:
a limit, a case where the advice does not apply, or a point the sceptical
reader is right about. Quote it.
8. TAGS. The title tag and meta description are written for the query I named,
not for the topic, and the description is not a restatement of the title.
RULES
- Do not produce a total, a score out of ten, a grade, a percentage or a
readability number. There is no such figure and inventing one would let me
ignore the individual gates.
- Do not predict traffic, rankings or performance of any kind.
- Do not suggest keywords to add or tell me to increase keyword frequency.
- Do not rewrite any part of the draft. If asked to demonstrate a fix, give
one sentence, once.
FINISH WITH
- VERDICT: SHIP or HOLD. A FAIL on gate 2 or gate 5 is always HOLD, whatever
the other gates say.
- THE ONE THING: the single change to make first, and which stage of my
process it should go back to: planning, outlining, drafting or editing.
- WHAT I DID NOT GIVE YOU: everything you marked CANNOT TELL, and exactly what
would have let you judge it.
The one question this post answers: [THE ONE QUESTION THIS POST ANSWERS]
The query it should win: [THE QUERY IT SHOULD WIN]
What in it could only have come from me: [WHAT IN THIS POST COULD ONLY HAVE COME FROM YOU]
My existing pages: [PASTE YOUR URL LIST AND TITLES, OR WRITE "none"]
The draft: [PASTE THE FULL DRAFT, INCLUDING TITLE TAG, META DESCRIPTION AND EVERY HEADING]What to change
Everything in square brackets is yours to replace. Nothing else needs editing.
[THE ONE QUESTION THIS POST ANSWERS]- The question in a reader's words, written before the model sees the draft so gate 1 has something to test against. If you find yourself writing two questions joined by "and", the gate has already told you the answer: the draft is two posts and both of them are currently half written.
[THE QUERY IT SHOULD WIN]- One query, ideally the one the post was outlined against. Gate 8 checks the title tag and description against it rather than against the topic, which is where most drafts fail: a title written for "content marketing" when the post was written to answer "how do I brief a freelance writer".
[WHAT IN THIS POST COULD ONLY HAVE COME FROM YOU]- Your own claim about the draft, which gate 2 then tries to find in the text. Stating it up front matters because it turns the gate into a comparison rather than an opinion. If you cannot name anything here, the gate result is already known and no run is needed.
[PASTE YOUR URL LIST AND TITLES, OR WRITE "none"]- URLs with their titles, straight out of sitemap.xml or the CMS. Titles are the half people leave off and the half gate 6 needs, since a bare URL rarely says what question a page answers. Without it the duplication gate returns CANNOT TELL, which is honest and useless.
[PASTE THE FULL DRAFT, INCLUDING TITLE TAG, META DESCRIPTION AND EVERY HEADING]- The whole draft as text with headings still marked as headings, plus the two tags that never appear in the visible page. Gates 3, 4 and 8 all depend on parts that a copy and paste from a rendered preview quietly drops.
How to run it
- 01Run the set in order, not the gate on its own
The gate tests things earlier prompts were supposed to produce, so running it against a draft written in one pass mostly returns fails you already suspected. Plan the pages, outline the post, draft it from your own specifics, edit it, then gate it. Each stage has its own prompt and its own refusal condition.
- 02Write down the one question and the only me claim first
Fill those two fields before pasting anything. Written afterwards they become a description of whatever the draft happens to contain, and gates 1 and 2 stop testing anything. Written first they are a claim the model checks the draft against, which is the whole mechanism.
- 03Export your URLs with their titles
Take the list from sitemap.xml or your CMS and keep the titles attached. Drop pagination, tag archives and anything noindexed. Gate 6 is the one that catches a post about to compete with something you published a year ago and forgot, which is a problem no amount of editing this draft will fix.
- 04Read the sourced list before the verdict
Gate 5 returns every number, date, study and company name it could not trace back to your inputs. Check each one yourself against its real source, and cut anything you cannot stand behind. The model is explicitly forbidden from telling you whether they are true, because it has no way to know and a confident answer here is the most damaging thing it could give you.
- 05Treat CANNOT TELL as a missing input, not a pass
Every gate marked CANNOT TELL is a gate nobody has checked. The output lists exactly what would have let it judge, so supply that and run again rather than reading the verdict as though those gates cleared. Most of them are fixed by pasting something you already have.
- 06Publish, then confirm the page is allowed to be indexed
Once it is live, run the URL through the indexability check. A post that clears every editorial gate and carries a noindex from a template or a header is invisible for reasons that have nothing to do with the writing, and it is a failure nobody notices because the page looks perfect to every human who visits it.
Questions people ask
What are the best ChatGPT prompts for blog posts?
Five, in a sequence, because writing a post is not one job. Planning which posts to write, outlining one against what already ranks, drafting from specifics only you have, editing the draft for search, and gating it before publication each need different constraints. A single prompt asked to do all five does the easiest one well.
Can one prompt take me from idea to published blog post?
It can produce a thousand words that look finished. What it cannot do is stop, and stopping is where the quality comes from: approving an outline before anything is drafted, refusing to start without a real example, and holding a draft that contains a number nobody sourced. Those checkpoints are the reason the workflow is a sequence rather than a prompt.
Why does the gate refuse to give a score out of ten?
Because there is no such number and a made up one is worse than nothing. A single figure lets you skip the individual gates, which is exactly where the information is, and it invites you to publish a draft at eight out of ten while it still contains an unsourced statistic. Pass, fail and cannot tell for each gate keeps every problem visible.
What does the CANNOT TELL verdict actually mean?
That the input does not contain what the model would need to judge that gate. It is there because the alternative is a guess dressed as an assessment, which you would then act on. In practice most cannot tell results come from a missing URL list or a draft pasted without its title tag, both of which take a minute to supply.
Does the gate check whether my facts are true?
No, deliberately. It lists every number, date, study and company name it cannot trace back to what you supplied, and it is forbidden from telling you whether any of them are correct. A model asked to verify a claim will produce a confident answer either way, and a wrongly confirmed statistic is more dangerous than an unchecked one.
Unsubscribe in one click. We never pass your address on.