How It Actually Works
A daily video series. Each episode answers one question about AI agents or growth with something I built or measured. I script, film and post every episode. The edit runs through pitch-video, the agent skill I built: lip sync, pause trims, captions and an eval on every cut.
-
01
An AI agent made this launch video as code
AnswersCan an AI agent make a launch video without a video editor?
My motion-lab skill wrote every frame as HTML, then measured the track’s loudness to land the music drop on the right frame.
1,020 framesloudness every 10 msdrop within 1 frame -
02
The eval that checks my AI video editor
AnswersWhat does an eval catch, and what does it miss?
The eval inside my editing agent re-transcribes the cut and checks every edit. It still missed a lip-sync gap I caught by eye.
115 s → 91 s27 cuts, 0 inside a word40 ms gap found -
03
How one markdown file teaches any AI agent a new skill
AnswersHow does a markdown file teach an agent a new skill?
I opened my own SKILL.md: the agent reads one description line until a request matches, and scripts do anything that has to be exact.
133 lines of English62 skills installed21 scripts, 2,895 lines -
04
What actually goes into an AI agent’s context
AnswersWhat is context, and what fills it up?
I measured my own setup, then changed my editing skill so each step writes its output to a file instead of the context window.
64 skills = 24,927 charslimit hit 4 timeslogs moved to disk -
05
What is a growth engineer, and how do you become one
AnswersWhat does a growth engineer do all day?
Product engineer vs growth engineer vs growth hacker, the idea-build-test loop, and why the real work is distribution.
3 roles comparedidea → build → A/B testPostHog: +250% activations -
06
How do I find ideas worth building
AnswersWhere do ideas worth building come from?
My filter, shown on three things I shipped: a problem I felt, that others complain about too, that fits a 2-minute demo.
Chithi: 22 languagescofoundr: 48 h buildTastemaker: 500+ stars -
07
Anthropic taught Claude to dream
AnswersHow does Claude fix its own memory between sessions?
A news breakdown from the source: Claude replays its logs offline and rewrites its notes, plus how to set it up in Claude Code.
413 → 10 memory files1 → 4 badgessame model -
08
Revolut cut its AI coding bill by 90%
AnswersHow does running a model yourself make it 10× cheaper?
Tokens vs hardware: Revolut runs an open-weights model on its own GPUs and keeps Claude for the hard tasks.
2,000 engineers56 B200 GPUs~90% cheaper -
09
AWS let AI agents write all of the code
AnswersHow do you trust code no human wrote?
AWS wrote the tests first against real AWS, then let agents build until they passed. Engineers built the checks.
200+ servicestests before codeevery line by agents -
10
Apple put its engineers on an AI budget
AnswersWhy would Apple ration the AI it told engineers to use more?
Most engineers never hit the old cap, so the cut is about habits: a cheap model for easy work, the big one for hard problems. I do the same in cofoundr.
$300 → $200 a day~$6,000 a month per engineerR&D up 32% to $45.8B -
11
Proton’s CEO tried to force AI on his staff
AnswersWhat goes wrong when you measure how much people use AI?
Proton graded staff on AI usage, skipped a pentest and cancelled company-wide re-interviews a day after the leak. Measure what ships and what breaks instead.
pentest called “loser thinking”cancelled in 1 day20+ hackathons, built with AI -
12
Goldman Sachs took Claude Code away from its traders
AnswersWhy would a bank take AI away from its own people?
Traders vibe-coded tools that touched client data before anyone reviewed them. Goldman kept AI but moved Claude Code to people who can read the code.
~12,000 engineers kept itin-house AI still firm-wideevery tool needs an owner
Need someone who builds it and sells it?
I’m open to growth, social and DevRel roles, plus hard technical builds. I reply fast.