← Back to recent questions

The friction of explaining yourself disappears, and that is a bigger deal than it sounds. Once an AI knows your work, your industry, the recurring projects you are juggling, you stop opening every chat by laying out the same background. That sounds like a small saving. In practice it changes how often you reach for the tool at all.

The shift I noticed in my own work is straightforward. I had a Custom GPT set up for writing tasks with my voice rules pre-loaded (Australian English, no em dashes, plain prose, the whole list). For the first six months I would still open plain ChatGPT out of habit, paste those rules in, then ask the question. Now I open the Custom GPT first because the second I do not, the output is wrong. The compounding return on context is real, and you do not really feel it until you go back to using a stranger.

Why this question matters

Most people experience AI as a stranger you keep meeting again. You ask a question, you get a generic answer, you close the tab. Memory and context features turn that stranger into a colleague who already knows what you are working on. The shift is from "useful sometimes" to "useful daily".

The three tiers

The simplest. Turn on memory in ChatGPT or Claude. Both call the feature "Memory". The assistant quietly notes things you tell it and brings them back when relevant. That you write in Australian English, that you have teenagers, that the recurring topic is your work. You do not have to do anything special. It learns by use.

The middle tier. Custom GPTs in ChatGPT Plus, and Claude Projects. These are saved versions of the assistant pre-loaded with the documents, instructions, and context for one specific job. I have a writing assistant pre-loaded with my voice rules, and a working assistant for the Bid Fragility Framework that holds the methodology and the database in context. Set-up is a one-off ten minutes. The saving is every time after that. For anything you do more than three times, this is the difference between fifteen seconds and fifteen minutes.

You are my writing assistant. You help me draft, edit, and tighten
prose for [my newsletter / my work emails / my blog / my book —
fill in].

Voice rules I want you to apply unless I say otherwise:
- Australian English. Analyse, colour, organisation, recognised.
- No em dashes. Use commas, full stops, or restructure.
- Plain prose. No marketing language. No "amazing", "seamless",
  "supercharge", "unleash", "game-changing". No exclamation marks
  unless quoting.
- Short sentences over long ones. Vary rhythm.
- Honest about limitations. Caveat confident claims when warranted.

About me, briefly: [your role, your audience, your subject area,
anything that should colour how you write to me — fill in].

When I paste text and ask for an edit, return the edited text plus
a short list of what you changed and why. When I ask you to draft,
ask one clarifying question first if you are unsure of audience or
length.

Never invent facts about me, my work, my family, or events I did
not describe. If you need a detail I haven't given you, ask.

The deeper tier. Agents that hold context across sessions and act on your behalf. We are not really there yet for most non-technical users in 2026, but it is where things are heading.

Day to day, what you notice is shorter prompts, more relevant answers, and less "let me explain the background again". You start treating the assistant less like a search engine and more like a colleague who has read the brief.

What I would not do

Do not put genuinely sensitive material into memory. Financial details, health information, anything you would not want surfaced later. Memory features can leak across chats in ways the providers occasionally fix and occasionally do not. Treat anything you put into memory as written down, not whispered. Review what is stored every few months. Both ChatGPT and Claude let you see and delete individual memories from settings.

I learned that one the hard way. Driving up the freeway with my younger daughter in the back seat, ChatGPT in voice mode on the dashboard for me, and the chatbot used her name out loud mid-answer. It had stitched her name into its memory from an earlier session of mine. That afternoon we both sat down and audited everything. Memory is useful. It is also a long-running list of facts you may not have realised it was keeping, and it needs a regular sweep.

The honest catch about lock-in

Switching providers later is harder once you have built up real memory and projects in one of them. None of it ports across vendors. Pick the assistant you are most likely to stay with for a year, then commit. The compounding return on context is real, but only if you stick with one. That is also the strongest argument for thinking carefully about which one you build the muscle on, because by year three the cost of switching is no longer just retyping a few prompts.

As of May 2026. Tool features and pricing change quickly; if you are reading this much later, check the current state before relying on the specifics.