The systems behind the channel itself. Each build follows one framework: the thinking constraint that was blocking progress, the KPI that measures the improvement, and the AI system that solved it. Client & enterprise work lives on the Cases page.
This is literally the pitch: Constraints removed, KPIs moved.
Building each episode page took 2–4 hours of focused writing + assembly. I couldn't publish without being present and uninterrupted.
Jacob-time per episode build. AI reads the brief and builds the page unsupervised overnight.
Nightly engine: claude -p runs on a Task Scheduler cron. It reads a brief I drop at 20:30, asks clarifying questions via Telegram, then builds the full episode HTML and leaves a morning handoff note.
Before I could use a video’s claims in content I had to sit through the whole thing — a gating step on every piece of research.
Extract + full fact-check. Found income claims were 4× inflated vs. real market rates.
The /watch skill: yt-dlp downloads, ffmpeg extracts 80 frames, Whisper transcribes 565 segments, then three parallel research agents cross-verify the claims against external sources.
The site was built for human eyes. Agents scraped it badly and missed key content, blocking any workflow that needed site context.
npx github: — any AI reads the entire site from a single call, in full context.
llms.txt manifest, JSON site index, an npm CLI package, and an MCP server — so any Claude instance can query the full site content without touching the HTML.
Long sessions degraded silently as context filled. I only noticed when output quality dropped — too late to act cleanly.
Silent context deaths. Two live token bars in the terminal show exactly when to /compact before quality drops.
statusline.js — a Node script that polls Claude Code’s internal session state and renders a real-time cockpit bar: model, effort level, context fuel, and session-token fuel.
Colleagues hit repetitive bottlenecks they didn’t know AI could remove. No one on the team could diagnose the problem, let alone prescribe a fix.
Colleagues now bring problems to me. Went from nobody knowing who to ask → the team’s default first call for anything AI-shaped.
No single artifact — the pattern. Applied constraint-first thinking inside one complex product company: audited real blockers, shipped targeted fixes, measured before/after. Results visible to the team compounded into trust.
“I’m not going to find this job on LinkedIn — because it doesn’t exist yet.”
Two real Skåne questions with no honest answer available: which electric-grid areas are heading for a build-stop before you commit to a connection, and whether a “15 minutes to transit” claim on a property listing means anything real, stop by stop.
165 grid companies mapped, 74 of 141 already build-stopped today. On the transit side, two numbers instead of one — the typical trip time and the one you actually have to plan around.
Built with Bengt: one engine, two data layers, one API key with per-layer access. Nätplansradarn reads every Swedish grid company’s own development plan and turns it into a live capacity map. TransitTrue simulates hundreds of possible departures per stop from real Skånetrafiken GTFS data instead of guessing from straight-line distance. Piloted on Lomma→Malmö C, ready to scale to the rest of Skåne.
Every AI answer improves with context — your files, your calendar, the people and plans around you — but almost nobody keeps that anywhere an AI can read it. Building it by hand is exactly the kind of thing people mean to start and never do.
One command hands you everything stored about you as plain markdown — the same files your own assistant reads — ready to drop straight into a Claude or Codex project.
Pocket Operator: a private, invite-only AI assistant with Cortex, its memory system, built in. Send /brain and it asks five questions once — who you are, what you're on, how you like answers, your world, your rules — then writes it to plain text files it re-reads every reply, forever, on any device. Say “export my data” any time and everything ships as files you own, ready to drop into a Claude or Codex project. It's the on-ramp: start collecting your own context somewhere simple, so any AI you use afterward — here or elsewhere — already knows more about you than a blank prompt.
Constraint and KPI get defined first, build second. New builds land here as the episodes ship.
I scope one constraint at a time, build the AI system to remove it, and measure the before/after.
Work with us →