Blog
Notes on what I build — how it works and what broke along the way.
-
Your AI built you a startup. It also built you a time bomb.
Vibe coding gets you to launch in an afternoon. The 30% it skips — security, access control, the parts that don't show in a demo — is exactly what breaks around Day 60. The data, the documented incidents, and a 5-minute gut check for any AI-built app.
AI agents vibe coding build in public Read →
-
I replaced the server with a prompt: a self-running AI news digest for $0
A developer-grade daily AI digest — crawl ~28 sources, let Claude edit, publish a static site — with no server to operate. The cron daemon is a scheduled routine, the database is a git branch, the host is a free CDN, and the editorial runs on a subscription rail at $0 metered API. The full teardown.
AI agents Claude Code build in public Read →
-
Your codebase could document itself — here's who's building that, and how
A wave of tools promises docs that keep themselves current — the repo as its own wiki. Here's the map: the main players (DeepWiki, Google Code Wiki, Driver, DeepDocs, Swimm, Mintlify) and the four distinct approaches they take, what each trades off, and the questions to ask before you trust any of them. No ranking — the space moves too fast for that.
AI agents documentation developer tools Read →
-
I spent a weekend teaching a computer to draw my 2-year-old
A weekend project: turn phone clips of my daughter into a hand-painted storybook cartoon. It came out adorable, a few times absurd (it once handed me a different child), and once a little bit magic — costs and bloopers included.
generative AI weekend project build in public Read →
-
AI speed without the tech debt: the playbook, ranked
AI doesn't create technical debt — ungated AI does. The Carnegie Mellon data says the difference between a codebase that compounds and one that rots isn't the model, it's the harness around it. Here's that harness, ranked by impact, with the one thing to do first.
AI agents technical debt build in public Read →
-
The AI coding honeymoon lasts about one month
A Carnegie Mellon study tracked 807 projects after they adopted Cursor: +281% code in month one, back to baseline by month three — and complexity and warnings that never left. The speed is a loan. Here's what it costs, and how to keep the velocity without the debt.
AI agents technical debt build in public Read →
-
A 0.6B model went from 10% to 92% on a job you're paying a frontier API to do
A developer fine-tuned a model small enough to self-host until it beat prompt-only frontier models at one narrow task — and a 3B model just out-reasoned giants 200x its size. The pattern underneath: the part of your AI feature that's narrow and verifiable is small enough to own. What that means for what you build.
small models fine-tuning AI agents Read →
-
My phone edits a family music video every night while I sleep
Google launched Gemini Omni to edit your videos. I already had the same thing — free, private, on my own phone — except I didn't write it. An agent built the whole app: Kotlin, tests, the nasty bugs and all. A build-in-public teardown.
AI agents Claude Code build in public Read →
-
Don't trust the AI leaderboard they just pasted into the pitch
Meta tested 27 private Llama-4 variants on Chatbot Arena to find one that would podium; the model you can download finished 32nd. Between that, length-biased voting, and benchmark questions leaking into training data, the public scoreboard has quietly stopped measuring capability. Here's how to read it — and what to measure instead.
AI agents benchmarks build in public Read →
-
I ran a 2,520-call model eval on my Claude subscription — no API key, real per-call cost
The full build behind right-sizing an extraction feature: a label-first golden set, a deterministic grader with a tolerant free-text matcher, four prompt scaffolds × three models, graded to the token. Plus the trick that kept it off the metered bill — running every call through `claude -p` on the Claude Code subscription, stripped to a clean ~320-token app call with caching and thinking forced off. Real flags, the turns×cost table, ~$21 of API spend avoided, and the three bugs that nearly faked the result.
AI agents evals Claude Code build in public Read →
-
My two-year-old has her own record label — her first hit is disco polo about carrots
My daughter has a personal record label: ~50 compilations, an animated series, and an Android app — all generated from one text file at ~$0.10 a song. Every song is code. And it's the exact same pipeline I point at a client's boring, repetitive content work.
AI agents build in public content as code Read →
-
Superpowers vs Matt Pocock's skills: two opposite bets on who runs your agent
Two of the best-known Claude Code skill libraries share one SKILL.md primitive but make opposite bets: Superpowers auto-injects a process every turn; Pocock's skills fire only when you ask. A first-hand architecture comparison, with a skill overlap map.
AI agents Claude Code build in public Read →
-
Intelligence is rising, the leaderboard is lying, and the harness still decides
Three things are true at once. Models are getting more capable on a clean exponential — METR's task horizon doubles every ~4 months now. The public leaderboards are increasingly gamed and contaminated. And the harness around a model moves its score as much as a whole model generation. Here's how to read all three when you're deciding what to build on.
AI agents benchmarks build in public Read →
-
The harness spends your money, not the model
Same model, same spec — but swap the agent around it and the bill moves 4-5x and the score moves 20+ points. The model is a commodity everyone shares; the harness is the variable you actually choose. So I'm running the experiment: one spec, one frozen test suite, four harnesses, measured to the token. Here's the rig and the prior art.
AI agents Claude Code build in public Read →
-
Open-weight AI is about three months behind the frontier
Epoch AI clocks the gap between open-weight models and the closed frontier at roughly three months — and it has briefly hit zero. GLM-5.2 lands a point behind Opus-tier coding at a fifth of the price. But the frontier moved up a tier, and that tier is the one a government just switched off. What that means for what you build on.
AI agents open weights build in public Read →
-
A better prompt beat a bigger model in my app — at a sixth of the cost
I built a small rig to right-size the LLM inside a real feature: read a messy sales email, return a structured record. Same eval, three models, four prompts, behind a deterministic grader. On the easy half the cheapest setup passed 100% for $0.0008 a call; on the hard half Haiku with a good prompt matched Opus with the same prompt — overlapping error bars — at one-sixth the cost. The lesson: don't pick the best model, pin a bar and send each task to the cheapest model-and-prompt that clears it.
AI agents evals build in public Read →
-
The model was never the hard part
For two years the bottleneck on AI agents was the model. It isn't anymore. An agent can now spin up a throwaway cloud account, deploy a live app, and self-destruct it in 60 minutes — no signup, no key, no human. That's one of four pieces of plumbing the industry quietly shipped in 2026: accounts, sandboxes, durable state, and auth. Here's the state of each — and the one nobody's solved.
AI agents agent infrastructure build in public Read →
-
The Suno accent trick: coaxing a regional ornament out of AI — and what it can't fake
A footnote to the Podlasie song experiment: if the local voice 'doesn't drawl,' can you prompt that drawl into Suno without training a new voice? Two cheap tricks and one honest trade-off.
AI Suno build in public Read →
-
A song for my wife from Podlasie — straight from an 1892 songbook
I wanted to give my wife a real gift: a Podlasie spring carol from 1892, sung in her own dialect, rebuilt with AI. Hours in 19th-century archives, reading a 130-year-old engraving note by note, reconstructing the dialect — and a twist at the very end.
AI heritage build in public Read →
-
A Fable 5 agent called both World Cup openers — to the exact score
Before the US pulled Fable 5, it built me an agent that called the first two World Cup matches perfectly. The teardown — split judgment from formatting, optimise EV, keep a human on the trigger — plus an honest look at how it's gone since.
AI agents LLM build in public Read →
-
I built a $0 lead-research agent instead of paying for a leads database
An AI agent that scouts ~100 companies a week from free public sources, scores fit, drafts the opener, and never sends. The funnel, four decisions, and why a human stays at the wheel.
AI agents lead generation build in public Read →