The Source-Bank Method: Let AI Expand Your Blog Without Steering You Off Course

Every blogger who has tried to use AI to "just expand this draft a bit" knows the feeling. You paste in a tight, honest 400-word outline, ask for more depth, and what comes back is a 1,200-word piece that is fluent, confident, and only sort of yours. It added a statistic you have never seen, invented a case study about a company you did not mention, and somehow made you sound like a LinkedIn influencer who writes "In today's fast-paced world" for a living.

The problem is not that the AI is bad at writing. The problem is that you gave it a blank space and asked it to fill the gaps. When a model does not have enough of your material, it will quietly reach for its material — generic knowledge, synthetic examples, and phrasing patterns pulled from a million other articles. The result is readable but ungrounded, and it drifts.

I have spent the last year writing bilingual articles with AI help, and I kept running into exactly this. The popular fix is the "outline anchor" method — keep a strict outline and tell the AI to stick to it. That helps, but it still leaves the AI room to reach beyond you. The method that finally worked for me is different in one small but important way: instead of anchoring the AI to a structure, I anchor it to a source. I call it the source-bank method.

Why the outline anchor is not quite enough

The outline anchor works like this: you write a detailed outline, paste it in, and instruct the AI to expand each section without adding new sections or changing the point of view. It reduces wandering, because the AI has a clear shape to follow.

But here is the gap. An outline tells the AI where to write. It does not tell the AI what it is allowed to write with. Under each heading, there is still empty room, and the model fills that room with whatever it has on hand. So you get a section that is in the right place, on the right topic, and still full of examples you never lived and facts you cannot verify.

The source-bank method closes that gap. You do not just tell the AI what shape the article should take. You hand it the only raw material it is permitted to use, and you tell it, in plain words, that anything outside that material is off-limits.

What a source bank actually is

A source bank is a small collection of your raw material for a given article. It is not a finished draft. It is the unpolished stuff underneath one: your notes, your real numbers, your own stories, quotes from people you actually talked to, things you observed, and links or documents you can vouch for.

For a single article, a source bank can be as short as a page of bullet points. For a recurring topic — say, a newsletter about running a small business in Malaysia — it can grow into a living document you keep adding to.

The key property of a source bank is that everything in it is true for you. You could stand behind every line under oath. That is what makes it a bank and not just "some context." When the AI draws from it, the output stays grounded in your real experience. When it tries to invent something, the instruction you gave makes that a violation, and good models respect the boundary far more reliably when the boundary is concrete.

How to build one in ten minutes

You do not need a system. Start with whatever you already have and dump it in.

First, gather the raw material for the topic. If you are writing about your week in business, that might be your calendar, a few WhatsApp messages, a receipt, and a half-written journal note. If you are writing a how-to, it is your actual process — the steps you really followed, in the order you really followed them.

Second, put it into a plain text file or a Notion page. Do not polish it. Typos, fragments, and rough phrasing are fine; the AI is going to rewrite it anyway. What matters is that the content is real.

Third, tag the bank with anything that needs to stay private or stay out of the article. A source bank often contains things you would never publish — a client's name, a revenue figure, a moment you do not want to share. Mark those clearly so you can strip them before pasting, or tell the AI explicitly that they must not appear.

The finished bank for a typical article looks like a page of mixed notes: three real stories, one number you actually know, a quote you remember, and a list of steps. That is plenty. You are not writing the article here. You are just making sure the AI has something true to work from.

The prompt that ties it together

Here is the instruction I actually use. It has been the single biggest improvement to my AI-assisted writing.

You are helping me expand a blog article.

Here is my SOURCE BANK — raw notes and facts about this topic.
Everything in it is true and comes from my own experience:

[ paste your source bank here ]

Here is my OUTLINE for the article:

[ paste your outline here ]

Rules:
1. You may only use facts, stories, numbers, and examples that appear
   in the source bank. Do not invent any statistic, study, quote,
   case study, person, company, or event that is not in the bank.
2. If the article needs something the bank does not contain, do not
   make it up. Instead, end that section with a short note to me in
   brackets, like [NEED: a real number for monthly hosting cost].
3. Keep my first-person voice and my point of view. Do not add new
   opinions I did not state.
4. Expand each outline point into a full section. Keep the same
   structure and do not add new top-level sections.

Two rules do the heavy lifting. Rule 1 turns the boundary into something the AI can check against, because the bank is sitting right there in the prompt. Rule 2 is the safety valve: when the bank is thin, the AI asks for material instead of inventing it. That bracket note — [NEED: ...] — has saved me from publishing made-up numbers more times than I can count, because it makes a gap visible instead of silently filled.

The workflow, section by section

I do not ask the AI to rewrite the whole article in one shot. I expand one section at a time, and I keep the bank and the outline in view.

Start with the outline you already trust. If the outline is weak, the whole article will be weak no matter how good the bank is, so spend a few minutes getting the structure right first.

Then take the first section and ask the AI to expand only that section, with the full source bank and the full set of rules in the prompt. Read the result. Check every factual claim against the bank. If the AI added something that is not in the bank, either the rule was not clear enough, or the model reached for a generic fact — cut it or move it to a [NEED: ...] note.

Move to the next section and repeat. Because you are working section by section, a mistake stays small. You notice a drift in paragraph two before it poisons the rest of the piece, instead of discovering at the end that the whole second half has wandered.

At the end, do one pass where you read the article aloud, or at least read it as a reader would. This is the moment to catch the last thing the source bank cannot guarantee: voice. The bank keeps the facts true. Only you can confirm the sentences still sound like you.

When the bank is too thin

The most honest thing this method will do to you is reveal when you do not actually have enough material. You sit down to write, build your bank, and realise it is only three bullet points. That is useful information.

You have two honest choices. One is to go get more material: check the real number, find the actual receipt, talk to the person, run the test. The other is to write a shorter article that your thin bank can genuinely support. Both are better than the third choice, which is to let the AI invent the missing pieces. A short, true article beats a long, half-true one every single time, and your readers can tell the difference even if they cannot name it.

This is also why I prefer the bracket-note rule over a hard "never ask." The AI asking [NEED: a real number] is a prompt to me to go do my job, not a sign the workflow failed. Filling that gap with something true is where the value of the whole exercise actually lives.

What changes when you write this way

The first thing you notice is that the article stops sounding like "AI content." It sounds like you, because every example in it came from your life. The second thing is that editing gets faster, because you are no longer fact-checking the AI against the internet — you are checking it against a page of notes you already trust.

The third thing is subtler. Over a few weeks of building source banks, you end up with a growing library of your own real material — stories, numbers, and lessons you can reuse across posts. That library becomes worth more than any prompt template, because it is genuinely, irreplaceably yours. AI cannot conjure it, and nobody can copy it.

None of this removes the need for you to be the writer. The source-bank method just changes what the AI is good for. It is no longer generating content into a vacuum and hoping it stays on track. It is expanding your material, on your outline, and flagging the gaps it cannot fill. You keep the wheel. The AI stops trying to grab it.

So the next time you sit down to "just expand a draft," do not start by pasting the draft. Start by writing down what is actually true. Give the AI a bank to draw from, tell it the bank is the only thing it may touch, and watch how much less steering you have to do.