Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
We all fell for it…
TL;DR
Theo agrees the real risk isn’t just tech debt, it’s “cognitive debt” — reacting to Lars Fay’s essay, he argues AI can actually crush old-school tech debt like Twitch-scale migrations across 8,000 broken TypeScript files, but it also tempts developers to stop understanding the systems they’re shipping.
Agentic coding becomes a slot machine when developers stop learning the underlying pieces — Theo says the dangerous behavior is repeatedly rerolling prompts instead of reading docs, debugging, or learning fundamentals, especially for juniors and insecure engineers who never built strong mental models in the first place.
The best engineers still use AI heavily, but as a collaborator, not a substitute — he walks through fixing a production Ping outage in Miami by combining his system knowledge with 3 targeted AI queries, using the model to identify a likely Prisma referential-integrity issue and write the cleanup SQL he didn’t want to risk by hand.
Cost is going up on the invoice, but intelligence per dollar is dropping fast — Theo pushes back on the “AI just keeps getting more expensive” line with Artificial Analysis numbers: GPT-5.5 Medium matched GPT-4x High-level performance at under $1,200 versus $2,800, and GPT-5.5 Low delivered roughly Sonnet-level capability for about $500 versus Sonnet 4.6’s $4,200.
AI is widening the gap between great and weak developers — his blunt version: strong engineers gain leverage, while shaky ones become “atrocious,” citing examples like AI-dependent juniors who can’t debug and even long-tenured engineers who still don’t know basics like SSH.
Theo’s core operating rule is to separate high-stakes code from one-off code — code that runs constantly for users should be better understood and higher quality because of AI, while throwaway scripts, migrations, calculators, and CSV wrangling should explode in volume because AI finally makes them worth writing.
The Breakdown
Theo opens with the contradiction: AI coding is amazing and clearly dangerous
Theo starts from the tension a lot of developers feel but don’t say out loud: AI coding tools are undeniably useful, and also kind of wrecking our relationship to code. He says he now opens his editor and sees prompt files, CSVs, and agent runs instead of code, while also noticing some skills atrophying as he keeps “pulling the slot machine lever” hoping the next output finally works.
Lars Fay’s “cognitive debt” framing clicks immediately
Reading Lars Fay’s “Agentic coding is a trap,” Theo says the key insight is not technical debt but cognitive debt. He pushes back on the lazy “AI creates tech debt” take by pointing out AI is phenomenal at brute-force cleanup work humans used to avoid, like giant lint migrations or the Twitch-era GraphQL change that broke around 8,000 TypeScript files.
Speed without understanding is the actual trap
Theo vibes hard with the article’s claim that AI workflows create distance between the developer and the code being committed. He says the “expert orchestrator” fantasy is mostly fake because most people are not writing meticulous plans, and the real problem is that AI rewards avoiding the discomfort of learning the building blocks underneath the system.
His big disagreement: costs aren’t rising the way people say they are
Theo detours into model economics and gets very specific: yes, teams may have gone from spending $100k on Cursor to $500k in a few months, but that’s because they’re using far more AI, not because intelligence is getting structurally pricier. Using Artificial Analysis data, he argues cost per capability is actually falling fast, with GPT-5.5 Medium beating older model tiers at less than half the benchmark cost and GPT-5.5 Low hitting roughly Sonnet-level performance at about one-eighth the price.
The skateboarding analogy explains why people are getting addicted
One of the strongest sections is Theo’s analogy to learning to ollie: people quit before mastering hard things because feeling stupid hurts. AI lowers that pain so effectively that many developers now choose the quick dopamine hit of another generation instead of the slower, more durable path of actually learning the language, library, or runtime.
Why Theo thinks he adapted better than many developers
He argues his own career prepared him unusually well for this shift because he was already a high-context, high-throughput builder juggling many systems and teams. He used to write less than 5% of the code while solving more than half the outages, and says AI amplifies that style — but only because he already understood the layers well enough to infer failure modes.
The Ping outage story: what good AI-assisted debugging looks like
Theo gives a concrete example from a recent outage on Ping, a codebase he hadn’t touched in four years. While manually investigating, he fired off parallel AI requests, used one model to narrow likely Prisma referential-integrity failures, then had another write the exact SQL to clean up bad rows — a back-and-forth he says only worked because he already knew the system deeply.
His closing thesis: use AI to learn more, not think less
By the end, Theo fully endorses Lars Fay’s warning that developers who outsource all thinking are making themselves obsolete. His final framing is crisp: the code that matters and runs constantly should be better because of AI, the throwaway code that runs once should be far more abundant because of AI, and if AI isn’t making you smarter, you’re using it wrong.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
