Asking an assistant for 'ten blog post ideas about bookkeeping' produces ten blog post ideas about bookkeeping. They will be plausible, evenly spaced, and interchangeable with the ten it gave the last person who asked. That is what happens when you hand over the one part of writing that was supposed to come from your life.
This workflow reverses the order. You write first, badly and fast. Then the assistant works on your list: it asks questions, finds the pattern you could not see, and stretches the list in directions anchored to your details. You choose at the end by the only criterion that predicts a finished piece: what do I actually have material for?
The running example is a bookkeeper with a two-person practice and about 60 small-business clients who wants to start a monthly email to them. She has a calendar, a folder of client questions, and no idea what to write about. Forty-five minutes later she has a first issue she can write tonight.
When to use it
Use it for
- You have a topic, a reader, or a deadline, and no angle
- Your own ideas feel thin and you want to find out whether they are thin or just underdescribed
- You keep writing the same kind of piece and want to see what else your material supports
- A recurring format (newsletter, column, weekly post) where you need a bank of angles, not just one
Not for
- You already know the angle. Go straight to the outline workflow. Brainstorming would only give you reasons to doubt it
- You have no material at all: no experience, notes, numbers, or stories. Brainstorming cannot supply what reading, interviewing, or living will
- The piece is assigned with a fixed angle, such as a brief from a client. Interrogate the brief instead
What you need before you start
- Fifteen quiet minutes before you open any chat
- The reader in one line: who they are and what they are busy with
- Whatever material you have within reach: a calendar, a folder of questions, old emails, a notebook
- A blank page, paper or digital, that is not the chat window
The steps
Each step names who does the work, the decision you make before moving on, and the prompt to use where one helps.
- 01You10 minutes
Write your own ten first, before any AI
Set a timer for ten minutes and write ten ideas in one line each. They can be bad. Half should be. The point is to get your own associations onto the page before a model's smoother ones can crowd them out. Once you have seen a fluent list, your own half-formed ideas look worse than they are, and you will abandon them.
The bookkeeper's ten included 'why your receipts folder is a lie,' 'the three clients who paid late in March' (she will not use names), 'what I do on the first Monday of every month,' and 'quarterly estimates, again.' Four were dull. One was angry. That is a normal ten.
- 02You and AI10 minutes
Ask for questions, not answers
Paste the list and the reader description. Then do the thing most people never do: forbid ideas. Ask only for questions, the ones a good editor would ask to find out what you know that your list does not say. A question like 'what happened on the first Monday that made you write that one down?' is worth more than any idea the model could add, because the answer is material only you have.
Answer the questions in your own words, quickly, in the chat. The answers are the real brainstorm. The bookkeeper's answer to 'what do clients ask most in March?' turned into a count: fourteen calls in one week about the same deadline.
Send this Here are ten ideas I wrote for a monthly email to my clients, and a line about who they are. Do not add ideas, do not rank these, and do not write any prose. Ask me six questions, one each for the six ideas you find most specific, that would draw out what I know that the idea does not say yet: a moment, a number, a conversation, a thing I keep noticing. Wait for my answers before you do anything else.
- 03AI proposes10 minutes
Cluster the list and name the tension underneath
Now ask the assistant to group everything, ideas and answers together, by the question each one is really about. This is a job models do well: finding the pattern across your own words. The bookkeeper's ten, plus her answers, collapsed into two clusters. One was 'clients find out about deadlines too late.' The other was 'clients do not know what I do between their visits.'
Read the clusters as proposals. If one is wrong, say so and why. The misreading often tells you something. Then notice which cluster you keep coming back to. That is usually the piece.
Send this Using only my ideas and my answers above, group them into the fewest clusters that fit, two to four. For each cluster, write the underlying question in one line and list which of my items belong to it. Do not add ideas or examples of your own. If an item fits nowhere, say so rather than forcing it. Then tell me in one sentence which cluster has the most concrete material behind it, and why.
- 04AI proposes10 minutes
Stretch, with constraints
Only now do you ask for ideas, and only inside the cluster you chose, and only ones built from details you gave. Each stretched angle must name the detail it grows from, so you can tell at a glance which ones are grounded and which are the model reaching for generic bookkeeping content. Ask it to mark anything that would need facts you have not provided with [VERIFY].
Expect a third of these to be good. That is a high rate, and it is high because the model is working from your answers rather than from the average of the internet.
Send this Inside the 'deadlines too late' cluster only, propose six angles for a single 500-word email. Each angle must grow from one specific detail I gave you; name that detail beside it. Write each as a one-sentence claim the email would make, not a title. If an angle would require a fact, statistic, or example I have not given you, mark it [VERIFY: what it needs] instead of inventing it. No introduction, no summary after.
- 05You10 minutes
Pick by what you have material for
Take your top three angles out of the chat. For each, make a plain inventory: what do I have, right now, that this piece would use? A number from my own records. A story I can tell without naming anyone. A document I can quote. A thing I did and what happened afterward. Write it down. Vague confidence does not count.
The bookkeeper's inventory for 'the fourteen calls' had a count from her phone log, two anonymized questions, and what she changed in her own calendar afterward. Her inventory for 'why your receipts folder is a lie' had an opinion and nothing else. The first one is the email. The second one is a note for later, once she has collected examples.
- 06You5 minutes
Write the claim and the source note
Finish by writing two things, by hand, outside the chat. First, the claim in one sentence: 'Most deadline panic in March comes from a date you could have put in your calendar in January, and here is the date.' Second, a four-line source note: the reader, what the piece must do, the facts it may use, and what it must not imply (for the bookkeeper: that she is giving tax advice).
The source note is what you carry into the outline workflow. The chat you just had is not. It has your rejected ideas in it, and they will try to come back.
Where it goes wrong
The assistant's ideas are fluent and yours look embarrassing beside them.
You opened the chat too early, or you let it add ideas in step 2. Close it, finish your ten, and when you return, forbid ideas until step 4. Fluency is the model's default, but it is not the same as having something to say.
Every idea the assistant offers sounds like a listicle title.
You asked for titles. Ask for claims, one sentence each, tied to a named detail from your material. A claim can be judged. A title can only be liked.
You picked the boldest angle and have nothing to put under it.
Material inventory, step 5, before you choose. Bold with no material becomes generic in the draft, because the model fills the gap with the average.
The clusters are tidy but wrong: the model grouped by keyword, not by meaning.
Tell it which grouping is wrong and why, in one sentence, and ask it to regroup using only your answers from step 2. If it is wrong twice, cluster by hand. The list is short enough.
You have a chat full of good runner-up ideas and no record of them.
Copy the runners-up into a file now, with the material note beside each. Next month's brainstorm starts there instead of at zero.
Before you call it done
- Ten ideas written by hand before the chat was opened.
- The assistant was asked for questions first, and you answered them from experience.
- The clusters were read as proposals and corrected where wrong.
- Stretched angles were tied to named details, with [VERIFY] on anything needing facts you lack.
- Each finalist has a written material inventory, not a feeling.
- The chosen angle is the one with the most material, and you can say what that material is.
- The claim and the four-line source note exist outside the chat, and the runners-up are filed.
Questions
Why write my own ideas first if the AI will give better ones?
It will give smoother ones, which is different. Your ten carry details, grudges, and half-noticed patterns that no model has access to. Once you have seen a fluent list, your rough ideas look worse than they are, and the material underneath them never surfaces. Ten minutes of your own first protects that.
What if I genuinely cannot think of ten?
Write the five you have, then write five bad ones on purpose: the obvious one, the one your competitor would write, the one you would be embarrassed by. Bad ideas on paper still carry your associations, and the questions in step 2 often turn one of them into the piece.
Can I use this for fiction?
Yes, with one change. In step 5, 'material' means scenes, images, people you have observed, and places you know, rather than numbers and documents. The inventory question is the same: what do I actually have for this story?
How many angles should I leave with?
One to write now, and a short file of runners-up with a material note beside each. More than that and you have a list, not a decision.