I run several projects with an AI assistant: a website, a charity cricket event, a set of books and a few small tools. For the first few months I did what most people do. I opened a chat, typed what I wanted, and hoped for the best.
Some days the results were brilliant. Other days the AI forgot what we had agreed the week before, invented a detail, or went off in a direction I never asked for. I would correct it, it would apologise, and twenty minutes later we were somewhere worse.
The fix was not a cleverer prompt. It was treating the AI like a new member of staff who is very fast, very capable and has no memory at all. You would not hand that person a job without a folder, a rulebook and notes from last time. This post shows how I set that up, so you can copy it.
You don't need the technical detail, but three facts change how you use it.
1. It predicts words from what is in front of it. An AI assistant writes its answer by working out the most likely next words, based on everything in the conversation plus what it learned in training. It is very good at this. It is not looking things up in a perfect database unless you give it a tool or a document to look at.
2. It only "knows" what is on its desk. This is why ChatGPT forgets your conversation, loses context or ignores instructions you gave last week, and Claude does the same. Think of each chat as a desk. Whatever is on the desk (your messages, files you share, notes you paste in) is what it works from. When you start a new chat, the desk is cleared. Unless your setup brings your notes back, last week's decisions are gone.
3. Mistakes build on themselves. If one answer is slightly wrong and you carry on from it, the next answer treats that mistake as true. Then the next one does too. This is how a conversation "drifts" and why long, messy chats so often end badly. People call the confident wrong answers "hallucinations".
AI hallucination means an AI giving an answer that sounds right but is wrong or invented. It isn't lying on purpose. When it doesn't have the facts on its desk, it predicts what a likely answer would look like, and sometimes that prediction is wrong. The more an AI has to fill gaps by guessing, the more it makes things up. The cure is to give it the facts, the documents and permission to say "I don't know".
So the goal is simple: put the right things on the desk every time, and leave as few gaps as possible.
You can do all of this with ordinary files in a folder. You don't need special software.
Make one main folder, and one folder inside it for each project. Mine looks roughly like this:
Projects/
00 HUB/ <- the map and the rules (more on this below)
Website/
Charity Event/
Books/
Property/
Inside each project folder, keep:
The point is that when you say "let's work on the website", there is exactly one place to look.
Projects vs chats. ChatGPT and Claude both have a Projects feature. A normal chat starts from nothing. A project keeps its own files and instructions, so every chat inside it starts with that background. Use one project per piece of work, the same way you use one folder per project on your computer, instead of hundreds of loose chats in the sidebar.
Built-in memory features help, but they decide for themselves what to keep and they fill up. The most reliable way to make AI remember everything that matters is to keep your own memory files and have it read them at the start of every session. The hub is a small folder that every session reads first. Mine has three files.
START HERE. A one-page map. Each project, its folder, and which file to open first. When I ask for something, the AI matches it to a project from this map instead of guessing.
HOW WE WORK. The general rules for every project. For example:
INBOX. A place for ideas that come up mid-task for a different project, so they are not lost and don't derail what I'm doing.
You don't have to build any of this by hand: one prompt can set the whole thing up with you, and create the files if your AI can reach your computer. The exact prompt we use for this is in the SEO Intel Skills and Prompt Pack.
Keep it short and current. The top section is the most important:
Any job you do more than twice deserves a written method. Mine is a single file with a short section per job. Examples of what goes in it:
The first time I did each of these, it took a long back and forth. Once it was written down, the AI followed the same steps every time, and so did I.
A useful habit: when a job goes wrong and you work out why, add one line to the method so it can't happen the same way again.
This is the part most people skip, and it is the most valuable over time.
Every time you make a meaningful decision, write one line:
21 Nov 2026 - Decided to stop offering paid services on the website and
focus on free guides plus low-cost tools. Reason: visitors want help,
not a price. Check results: 27 Nov.
Then, when you check the results, add what happened. Over weeks you build a history of what worked and what didn't, a bit like a weather record. You stop repeating the same mistakes, and when the AI suggests something, you can both look back and see whether something similar worked last time.
I keep two files:
Once a prompt reliably gives you good results, stop retyping it. Save it, and reuse it word for word. Most AI tools now let you store reusable instructions. Some call them skills, some call them projects or custom instructions. The idea is the same: the method is locked in, so you get the same careful process every time instead of a slightly different one depending on how you phrased it that day.
How to turn a job into a skill, and keep it improving, is in What are Claude skills?.
Signs a prompt should become fixed instructions:
Fixed instructions reduce drift and mistakes. They don't make an AI perfect, so keep checking anything that matters.
Start:
During:
End:
Ready-made prompts for starting a session, ending one and the weekly memory tidy are in the SEO Intel Skills and Prompt Pack.
Most "best AI prompts" lists are tricks. What actually works is giving the AI the same six things a good manager would give a person. A simple pattern for anything bigger than a quick question:
The six-line request template we use, with an example for each line, is in the SEO Intel Skills and Prompt Pack.
The last line matters more than people think. "Build it but don't publish" has saved me from more mistakes than anything else.
Short answer: no, you don't need to pay, and yes, it works in all three of the big assistants. The method is just folders, notes and saved instructions, so it isn't tied to one tool.
On a free plan. At the time of writing (November 2026), all three let free users keep work in one place:
What paying adds. Mostly room. Paid plans give you more usage before you hit a limit, more files per project and, in Claude, better handling of large project folders. If you work on the method every day you'll probably notice the limits on a free plan, but you can start without paying and decide later.
Across different assistants. The folders, memory file, decision log and "How we do it" rulebook work anywhere, because they are your own files. Saved instructions are where the tools differ:
Tip for free plans: keep each project's files few and tidy. One short, current memory file beats ten long ones, and it leaves room under the free limits.
Plans and limits change often, so check the current details on each provider's help pages before you rely on them.
None of this is clever. It is just good filing. But it turns an assistant that forgets everything into one that picks up exactly where you left off, and that is where consistent, accurate results come from.
This guide comes from a real problem on one of my own stores. Here's how to skip it.
Run the free store audit. No account, no card. Built for Shopify; Wix stores get a quick check too.
Run the free auditSkills Pack (Shopify, £14.99): 8 Claude skills that check your store against Merchant Center rules and fix what they find, including suspension recovery.
Skills Pack £14.99I cleared my own suspension without paying anyone. The full story, every mistake and every fix, is free to read.
Read how I did it