I play poker occasionally. Real poker, offline. But more often online, just to switch off. Last night another small tournament, like so many others. After every session I have Claude analyse my hands: where did I play well, where did I not. The analysis is always solid, precise, honest, detailed.
But as plain text it does not stick. I read it, nod, and at the next tournament I fall into the exact same traps as before.
So this time I added one more step. No Markdown dump. HTML.
The same analysis content Claude usually delivers as a text block, this time as structured HTML with visual cards, colour coding, and clear navigation. The difference between actually internalising an analysis and just reading it.
And that was the moment Output Engineering stopped being just a working principle.
Karpathy posted exactly this on X last week: HTML outputs instead of Markdown dumps. Welcome to the club.
View post on X →He also coined "Context Engineering," and that was long overdue. But between "how do I give the AI the right context" and "actually understanding what comes out at the end" there is a layer that nobody has named yet.
I call it Output Engineering.
Here is how it works for us:
I regularly do deep research with four different AIs. ChatGPT, Claude, Perplexity, Gemini. Same briefing, same context. I call it Triple Research, even though it is technically Quadruple. The name came first.
What is Triple Research?
The same strategic question is sent simultaneously to four different AI systems, with identical briefing and context. All four perspectives are merged and compared.
The name "Triple" comes from the early days of the method. Today it runs on four systems, but the name stuck.
What you get from that are three documents with three different functions.
The raw document is not there to be read. It is working material for the AI. I upload it and ask targeted questions: which platform gets recommended most? Where do the four AIs contradict each other? What is consistent across all of them?
Claude builds a summary from that which I and my team actually read. Key recommendations, the most important differences, decision basis. But there can be a world between reading something and truly understanding it.
That is why we recently added a third step. It is a game changer: the HTML report.
Last week I did this for a question about a potential AIKIA Labs community. Should I build one? What format? Which platform still makes sense in 2026? Raw document queried, summary read. Then I gave Claude the third instruction: the same thing, as HTML.
You can see the report here: [insert link]
The difference from a text summary is immediately visible. The structure is clear because navigation and visual hierarchy make the document readable in seconds. The colour coding (green for recommendations, red for what drops out) lets you scan without reading a single sentence. Charts, comparison tables, and action plans can be embedded directly.
You understand all the connections much faster and remember them better, because our brains simply store visually processed information differently from plain text.
"The summary gives me facts. The HTML report gives me clarity."
He is right that Context Engineering is the real job. Writing prompts is yesterday. And he is right that HTML is better than a Markdown dump. But that already goes into Output Engineering.
Context Engineering answers one question: what do I give the AI so it can work well?
Output Engineering answers the other: what should come out, for whom, in what format?
I teach both in my AI First courses, because good research dies exactly here. The prompts were good. The context was there. And in the end, out comes a Markdown block that nobody opens again.
The raw document for the AI. The summary for reading. The HTML report for real understanding and good decisions.
Works for community research. Works for poker hands. Works for anything you actually want to internalise.
One research session. Fifteen minutes.
Honestly: how many reports have you produced in the last few months that nobody opened again afterwards?