This is the whole loop in one place. Every other workflow on this site is a close-up of one stage. This one walks the stages in order, from a thought you could barely explain at lunch to a piece with your name on it. Direct, propose, judge, own, repeated at each step.
The running example is a physical therapist who owns a two-person clinic and wants to write a blog post about why she stopped handing patients printed exercise sheets. She has six months of intake logs, two patient comments she has permission to quote, and about three hours across two evenings. That is enough. It is usually enough.
Two sittings matter more than the total hours. Steps 1 through 5 belong in the first sitting, while the idea is warm. Steps 6 through 9 belong in the second, when you can read the draft the way a stranger would.
When to use it
Use it for
- A piece of 600 to 2,500 words where facts, voice, or your reputation are involved: a blog post, a newsletter issue, a personal essay, a talk written out in full
- The first time you try vibe writing end to end and want the stages in order rather than improvised
- A piece you have been putting off because the idea is real but the shape is not
- Any writing that will carry your name somewhere you cannot easily correct later
Not for
- A two-paragraph email or a social post. Pick the single stage you need. The email mega prompt or one bounded edit will do it in ten minutes
- Fiction longer than a short story. The stages still apply, but the outline and draft steps need the fiction path's adjustments
- Work that contains confidential client, patient, or employee data you are not permitted to paste into a third-party tool. Strip or substitute it first, or do not use the tool
What you need before you start
- A rough idea, even one sentence, and an honest answer to who it is for
- Raw material: notes, numbers, a voice memo transcript, quotes with permission, links you have opened yourself
- Three short samples of your own writing, or a filled-in Voice Sheet
- Two sittings on your calendar, the second at least a few hours after the first
- A document outside the chat where the master copy will live
The steps
Each step names who does the work, the decision you make before moving on, and the prompt to use where one helps.
- 01You15 minutes
Set up the session before you ask for anything
Full prompt for this stepOpen a fresh chat and paste the Vibe Writing Mega Prompt with its six blocks filled in. The two blocks that carry the most weight are the raw material and the definition of done. For the therapist, raw material is the intake-log numbers (31 percent of patients reported doing their exercises with printed sheets, 58 percent after she switched to 40-second phone videos), the two approved patient comments, and the fact that she stopped buying toner. Done means: every number traces to the log, the quotes are verbatim, and it reads like her last two posts.
The brief ends by telling the assistant not to draft yet. Let it ask its questions. If one of them exposes a gap in your material, go get the material now, not after a draft has been built around a guess.
- 02You and AI20 minutes
Discover the angle, not just the topic
A topic is 'exercise sheets.' An angle is 'the sheets were never the problem: the sheet was a sign that I had stopped paying attention after the appointment ended.' Before you outline anything, write five angles of your own, one line each. Then run the Discover Writing Ideas prompt on your material to see angles you did not think of.
Judge the assistant's angles by one question: do I have material for this? An angle that needs numbers you do not have is a research project, not a post. The therapist chose the angle her log actually supports, which was the quieter one.
Send this Here are five angles I wrote for this piece, followed by my raw material. Do not write any prose. Ask me the four questions whose answers would tell us which angle has the most real material behind it. Then, and only then, propose three further angles, each tied to a specific detail in my material. Mark any angle that would need facts I have not given you with [VERIFY: what it needs]. No summary after.
- 03You and AI20 minutes
Build the outline and mark the gaps
Full prompt for this stepGive the assistant your one-sentence claim and ask for two candidate structures, not one. Two structures force a comparison. One structure invites you to accept it because it exists. For each section, ask what material from your raw notes would support it.
Then take the outline out of the chat and reorder it by hand. Mark every section that has no material behind it: cut it, gather for it, or tag it [VERIFY]. The therapist cut a section called 'what the research says about adherence' because she had not read the research and did not want to pretend she had.
- 04AI proposes10 minutes, most of it reading
Ask for one draft, nothing else
Full prompt for this stepPaste the locked outline back, say 'keep' for anything you wrote yourself, and ask for one complete draft. Not three versions, not a draft with commentary. One. Options at this stage make you a judge of beauty contests instead of a writer with a plan.
Read the draft once, straight through, without touching it. You are checking one thing: is the spine right? If the argument runs in the order you chose and the facts are the facts you gave, the draft is worth editing. If it quietly reorganized the piece or filled gaps with confident generalities, do not start fixing sentences. Reset with a tighter brief.
- 05You and AI30 to 45 minutes
Judge, then make bounded edits one problem at a time
This is the step most people skip in favor of 'make it better,' which is how a draft drifts away from them. Instead, mark the draft yourself first: keep, cut, fix. For each fix, write one line naming the problem and its effect on the reader. 'Paragraph four sounds certain, but the log only shows two numbers side by side.'
Then run one bounded edit per turn. Name what may change, what must stay, and ask for OLD and NEW on the sentences touched. Paste the accepted result into your master copy after every turn. The chat is a workshop. Your document is the manuscript.
Send this Change only paragraph four. The problem: it says the videos caused the increase, and my log can only show the two numbers side by side. Rewrite it so a careful reader sees the comparison without a causal claim. Keep the two figures exactly, keep the paragraph within 20 words of its current length, and leave every other paragraph untouched. Show OLD and NEW for the sentences you change, then stop.
- 06AI proposes15 minutes
Get a critique in a fresh session, as the reader
Full prompt for this stepOpen a new chat. The one that wrote the draft is a poor critic of it: it will defend its own choices and admire its own sentences. Paste the draft with the reader description and ask for a critique in the reader's voice: where they would stop reading, what they would not believe, what they already knew, what question they would be left with.
Treat the critique as a list of hypotheses, not verdicts. The therapist's critique said the opening was slow, and she agreed. It also said the post needed a 'clear call to action.' She did not want one, and she was right to ignore it. A critique is an AI proposal like any other.
- 07You and AI20 to 40 minutes, depending on how many claims the piece makes
Run the fact-check pass, by hand
Full prompt for this stepAsk the assistant to list every checkable claim in the draft: each number, date, name, quote, and 'studies show.' That part it can do. Then check each one yourself against the source you have in front of you: the log, the email granting permission, the page you opened. The assistant cannot verify anything, and a reply that says 'I have verified' is wrong by design.
Resolve every [VERIFY] tag. Resolution means one of: you found the source and the claim stands, you corrected the claim, or you cut it. There is no fourth option.
- 08You30 minutes
Do the final human edit, out loud
Full prompt for this stepClose the chat. Read the piece aloud from your master copy and change anything that does not sound like you: the sentence you would never say, the rhetorical question, the tidy summary you did not ask for, the three-item list that appeared because models like threes. This pass is yours alone. The Final Human Edit prompt is a checklist for it, not a request for another rewrite.
Check the stance last. Nothing you said carefully should have been strengthened, and nothing you said confidently should have been hedged. The therapist found 'may help some patients' where she had written 'helped most of mine.' She put it back.
- 09You10 minutes
Release it, disclose what you should, and file the ledger
Publish or send from your master copy, never from the chat. If the destination has a disclosure rule (a client, a classroom, a publication with an AI policy), follow it. If it does not, decide what your reader would reasonably want to know and say that much.
Then spend five minutes on the accepted-decisions ledger: the angle, the structure, the voice notes that worked, the critiques you rejected. Next piece, you start from there instead of from nothing.
Where it goes wrong
You finished in one sitting and the piece reads fine but feels like someone else wrote it.
You skipped the gap between drafting and editing. Put the second sitting at least a few hours later, and do step 8 before anything else in it. Your ear needs distance.
Each round of edits fixes one thing and loosens two others.
You are asking for whole-draft rewrites. Return to bounded edits: one problem, OLD and NEW, master copy pasted back each turn. If the assistant keeps regenerating untouched paragraphs, reset the session with the current master copy and a shorter brief.
The draft contains a statistic or a study you never provided.
Ask for a list of every number and where it came from in your material. Delete anything it cannot trace. Then remind the session: replace, do not invent, and use [VERIFY]. This is also why step 7 is by hand.
You chose the most exciting angle and ran out of material halfway through the draft.
Go back to step 2 and pick by material, not by headline. The quieter angle you can support beats the bold one you cannot.
The critique and the fact-check were done in the same chat that wrote the draft, and both came back clean.
Self-review by the same model is weak evidence. Critique in a fresh session, and fact-check against sources you open yourself. 'Everything looks good' is the symptom, not the result.
Before you call it done
- The definition of done was written before the first draft, and the finished piece meets it.
- The angle was chosen for the material behind it, and that material is in the piece.
- The outline was reordered by hand, and every gap was cut, gathered, or flagged before drafting.
- The draft came as one draft, and the spine was judged before any sentence was edited.
- Every edit after the first draft was bounded, and the master copy lives outside the chat.
- The critique came from a fresh session, and you wrote down which points you rejected.
- Every number, name, quote, and date was checked against a source you opened, and no [VERIFY] tags remain.
- You read the final version aloud and could defend every sentence.
- Disclosure was decided before release, and the ledger is updated for next time.
Questions
How long does this really take the first time?
Longer than the estimate, usually four to six hours, because the first session is also where you learn what your raw material is missing. The second piece is faster: the voice notes, the banned list, and the ledger carry over. By the third, two to four hours for 1,000 words is realistic.
Can I skip the outline step if I already know what I want to say?
You can skip the AI part of it, but write the outline yourself and mark the gaps anyway. The gap-marking is what keeps the draft from filling holes with confident generalities. Ten minutes on paper saves an hour of bounded edits.
Do I have to use a different chat for the critique?
Yes, if you want the critique to be worth anything. The session that wrote the draft has every reason to approve it. A fresh session with only the draft and the reader description gives you something closer to a stranger's read.
What if the assistant's draft is better than what I would have written?
Then check whether it is better or just smoother. Smooth is the model's default, and it is not the same as right, specific, or yours. Keep the sentences you can defend, rewrite the ones you cannot, and notice that the parts you like best are usually the ones built from your own material.