03
Blog
01
System One
Most AI features are an enum, and we kept paying chat models to write essays about which case applies.
6 min read ·
02
Hiring machines
One model writes the resume and another reads it. The only line that still counts is one a stranger can check and would care about.
5 min read ·
03
Show your work
A check that takes its answer key from the thing it checks will always agree. A sum and a verbatim quote can say no.
6 min read ·
04
Retries
The scariest state in browser automation is 'maybe'. Retry code and agents both pretend it does not exist.
6 min read ·
05
Silent fallbacks
One return value meant four different failures, and every caller heard the same thing: carry on. Give your fallback a name, a type and an alarm.
5 min read ·
06
The bug was in Oregon
The portal was in India and my browser was in Oregon. Nothing in the code said so.
4 min read ·
07
Self-healing
Three flaky clicks in testing were one bug. The vision model I paid to heal my agent was looking at a blank screen.
5 min read ·
08
Rules first
For one client, an invoice took 40 minutes of manual tariff and CSV work. A tool that refuses to guess takes about 3 seconds and zero tokens.
5 min read ·
09
Who signs the diff
Agents write the code now. Somebody still has to sign for it, and the industry is hiring fewer of the juniors who grow into that somebody.
5 min read ·
10
Corporate ADHD
Management walks in with answers instead of questions, and the work starts over.
5 min read ·
11
Agents and scripts
In October Gemini chose every click my browser agent made. By January the SOP left the model 1 step in 31.
5 min read ·
12
SOPs as specs
Kiro and Spec Kit pitched specs to programmers in 2025. The customs team I automate for follows one every day and calls it the SOP.
4 min read ·
13
Merged, then reverted
An open-source project merged my docs rewrite in August and reverted it in October. The maintainer was right.
4 min read ·
14
Builder.ai
The viral story said the AI was 700 engineers in India. The record says something duller and worse.
3 min read ·
15
AI-first memos
Nobody can prove a chatbot can't do the job. Shopify now wants that proof before a team can hire.
5 min read ·
17
Glue code
The 'AI software engineer' pitches skip the part where two systems disagree. That part is the job.
4 min read ·
18
Compilers
Google says AI generates more than a quarter of its new code. Compilers generate all of its machine code, and nobody asks to review that.
5 min read ·
19
Fifty people
A percentage hides the rows. At 50,000 student records, 99.9% accurate means about fifty students with a wrong record.
4 min read ·
21
Why C++
The machine, the latency budget or the SDK picks C++ for you. Even the White House report that urges memory safe languages explains why it stays.
5 min read ·