My prompts have gotten shorter. A lot shorter. Mid-2025 they averaged 12-15 words. Now they're 6-8. "Check logs." That's it.
What changed: search and memory. Not smarter models.
This piece is about why information architecture matters more than model capability for serious agentic work.
Went down a rabbit hole on AI consciousness ("Little AGI") research this week.
The viral claims? Overblown.
The actual papers? Way more interesting. Models that know when they're being tested. 20% accuracy on detecting their own internal states. Misalignment that emerges from narrow training.
Get Claude.ai Cowork. Paste the sheet into a new that. Ask it what you want. It will likely handle it for you.
23.02.2026 06:30 β π 0 π 0 π¬ 0 π 0
There are two cognitive modes when working with text: generating and editing.
AI is great at generating. Traditional editors are great at editing. I spent months trying to force Claude.ai to do both.
Then I switched to Claude Codeβfiles on my filesystem, editable in Obsidian.
Here's the workflow.
Wow.
17.02.2026 03:59 β π 35735 π 16565 π¬ 1291 π 2461
GitHub: 100+ million new repositories in 2024. 25% YoY growth.
My Head of Product friend: "We're seeing 2-3x output from our top engineers. PMs and designers can't keep up."
2026 won't be remembered for AI. It'll be remembered for the software explosion AI enabled.
New post on what's coming.
Stack Overflow still gets read.
It stopped getting written.
Question volume collapsed 95% from peak. The archive is a museum nowβvisited, but not lived in.
Where did developers go? Everywhere else.
open.substack.com/pub/hyperdev...
Or lazy, which is wise mode than oblivious though not as bad as sociopathic. Perhaps the combination of the two is the third?
12.02.2026 06:20 β π 0 π 0 π¬ 0 π 0*sociopath to be fair.
12.02.2026 06:13 β π 0 π 0 π¬ 0 π 0You move on. If you donβt look for an opportunity to connect to a hand reached out in a mutual acquaintance-making, you are either oblivious or a psychopath. Neither bode well.
12.02.2026 06:13 β π 1 π 0 π¬ 2 π 0I think weβre learning that itβs not him, per se. in a normal world heβd be just pathetic as he was in NYC. Itβs his enablers. Never forget them.
12.02.2026 06:10 β π 4 π 1 π¬ 0 π 0Yes
12.02.2026 06:09 β π 1 π 0 π¬ 0 π 0
The IDE isn't the villain. AI code generation layered on top of it is.
41% more bugs. 30% more warnings. An "illusion of competence."
The research is inβand it points to one mistake: mixing generation and editing in the same interface.
open.substack.com/pub/hyperdev...
Anthropic just released Agent Teams in Claude Code - multiple AI agents working in parallel on the same codebase, coordinating autonomously.
A researcher stress-tested it by having 16 agents build a C compiler from scratch. Cost: $20K. Result: 100,000 lines that compile Linux.
"Just ask GPT" is productivity theater.
Three rules that actually work:
AI writes its own instructions
Strategic context beats document hoarding
Different model checks the first
Nine months of testing. New manifesto on HyperDev.
Says the cretin who knows less about βwhat America was aboutβ than any person I know.
02.02.2026 05:26 β π 0 π 0 π¬ 0 π 1open.substack.com/pub/hyperdev...
30.01.2026 13:40 β π 0 π 0 π¬ 0 π 0
I built claude-mpm to orchestrate AI agents for coding. Then I used it to onboard as CTO.
One week. 5 hours of my time. 267 analysis documents.
And answers to questions most new executives don't get to ask for six months.
The "new guy card" is comfortable. I wanted to walk in like a veteran.
Only 3.1% of developers highly trust AI-generated code. Yet 51% use AI tools daily.
That gap doesn't close through belief. It closes through process.
Full analysis with the research, frameworks, and practical verification stack: π
Anthropic just shipped Coworkβ"Claude Code for the rest of your work." I've been testing it while prepping for a new CTO role.
The tool itself: solid step forward. Handles long-running tasks without crashing. Saves output to your filesystem. Works.
Watching the planes land at Maho beach in St. Martin.
14.01.2026 19:31 β π 0 π 0 π¬ 0 π 0