I've spent the better part of a decade in HubSpot's orbit, so when Dharmesh Shah writes something, I read it. His recent simple.ai piece on future-proofing your AI workflows made it about two paragraphs before I sat up. The HubSpot co-founder was describing my mornings: markdown files that travel between models, context that lives with you instead of with the tool, a little "About Me" file quietly outworking the model choice itself.
I run my life on those files. A personal constitution. A writing codex. A running state-of-the-union my AI reads before it says a word to me. So this isn't a rebuttal. When one of the most credible builders in software says the future belongs to portable context, that's the sound of an idea going mainstream, and I'm glad to hear it.
But he leaves the hard part as an exercise for the reader. And the hard part is the exact spot where most people quit.
What actually goes in a portable AI context file?
The durable truths. Who you are and what you value. How you sound on the page — the rhythm, the vocabulary, the words you'd never be caught using. What you're working toward right now. The stories that explain you faster than any bio. The people who keep showing up in your questions. How you like to work, and how you don't.
Write those down and any model can pick you up mid-thought, on day one, forever. Skip them and your file is a task list with a nice header.
At RUMO we call each durable piece a context anchor: a plain document you own and paste into any tool. Dharmesh's whole architecture is context anchors without the name. Which, honestly, is the strongest validation a coined term can get.
The context gap is the whole game
Here's the line from his piece worth keeping: "The context gap is dramatically more impactful than any difference in ability between the premier models themselves."
Consider what that means. The industry spends billions chasing benchmark points, and the thing that actually decides whether your AI is useful is whether it knows you. A naked frontier model loses to last year's model with your context loaded. I've watched it happen in my own work. I switch models mid-project all the time, because the anchors travel and the tools don't get a vote.
Dharmesh lists the practical wins: outage insurance, picking the right model for each task, never re-introducing yourself, and being ready for whatever launches next Tuesday. All true. All real. The thing is, every one of those wins assumes the file is already full.
So why does everyone stall at the blank file?
Because Dharmesh hands you a template, and a template is not a self.
His starter file has three headers: what I'm working on, how I want you to talk to me, things to avoid. Good bones. And then you sit down to fill it in, and the cursor blinks, and you discover the uncomfortable truth about self-documentation: you are the worst-placed person to see your own context. Your voice is water and you're the fish. Your "things to avoid" list doesn't exist yet because it's a scar log — you find out what belongs there one bad output at a time.
| Dharmesh's starter file | What it captures | The durable layer underneath |
|---|---|---|
| "What I'm working on right now" | Current projects | Your situation, which changes quarterly, and sits on top of an identity that doesn't |
| "How I want you to talk to me" | Format and tone preferences | Your actual voice: rhythm, go-to phrases, never-words |
| "Things to avoid" | Corrections, added after mistakes | Your values and boundaries, written down before the mistakes |
| (not in the template) | — | The stories that explain you, and the people who matter |
The gap between the left column and the right one is where the future-proofing actually lives. And closing it isn't a formatting problem. It's an excavation problem — the work of surfacing what you know so deeply you've stopped seeing it. That excavation has a name: context mining. The short version is that you don't fill the file by staring at it. You fill it by being asked the right questions, the way a good biographer or a sharp new colleague would ask them.
Where I'd push back, gently
Dharmesh offers Custom Instructions as the "80/20 compromise" for people who won't make files. I get the instinct: some protection beats none. But the compromise quietly re-creates the problem he's warning about. Custom Instructions live inside one tool, locked to one settings pane, rebuilt by hand everywhere else. That's not portable context. That's a smaller cage. I've compared every way to give AI context, and the pattern holds: anything that lives inside the app belongs to the app. Use Custom Instructions as the place you paste your context's highlights, not as the place your context lives.
The file is the easy part
So yes. Dharmesh is right, and it matters that he's the one saying it. Portable beats platform. Markdown beats memory features. The file you own beats the profile they hold.
But the detail from his piece I keep returning to has nothing to do with tooling. It's the promise that when your main tool goes down, you're working somewhere else "in under a minute." That minute is only possible if the file is full. An empty template ports nothing.
The good news: filling it is a guided half hour, not a writing retreat. Start with the deepest anchor, the one every other layer stands on — who you are and what you value. The free constitution builder interviews you and drafts it with you, and you walk out with the first real file in your set.
Then the next time a newsletter lands describing the future of AI workflows, you'll get to do the best thing a reader can do. Sit up, nod, and realize you're already living there.
Frequently Asked Questions
- What is portable AI context?
- Portable AI context is durable information about you — identity, voice, current situation, preferences — stored in files you own, usually markdown, instead of inside one AI tool's memory. Because you control the files, you can hand the same context to ChatGPT, Claude, Gemini, or whatever launches next, and switch tools without losing what your AI knows about you.
- What should I put in my AI context file?
- The durable layers: who you are and what you value, how you sound on the page, what you're working toward right now, the stories that explain you, the people who matter, and how you like to work. Skip chat logs and one-off corrections. If a fact would still be true in three months, it belongs in the file.
- Why does Dharmesh Shah recommend markdown files for AI context?
- Because markdown is plain text with light formatting — every AI tool can read it, no app owns it, and it survives any platform change. Shah's argument is that the context gap between a tool that knows you and one that doesn't matters more than the ability gap between frontier models, so the context should live in a format you control.
- Are ChatGPT Custom Instructions enough to future-proof my AI setup?
- They're a start, not a foundation. Custom Instructions live inside one tool, cap out at a few thousand characters, and have to be rebuilt by hand in every other app. They work as a summary of your context — a place you paste the highlights — but the durable version is a document you own that travels everywhere.
- How do I figure out what to write in my context file?
- Don't stare at the blank page — get interviewed. The patterns that make you legible to an AI are mostly invisible to you, so answering pointed questions works far better than freewriting. That guided excavation is called context mining, and it's how a blank template becomes a file with your actual self in it.
