Four things worth
your next three minutes.
LLMs as programmable systems. Why being first matters more than being best. The AI coding switch. And why SaaS has a vibes problem.
LLMs are becoming programmable via documentation — and that changes the tooling game
Leading models — Gemini, ChatGPT, Claude — and AI-native editors like Cursor are adopting SKILL files: structured documentation that gives the model persistent, task-specific context. The idea is simple: if you can document how to do something well, an LLM can now execute it reliably at scale.
The implication isn't just better autocompletion — it's that institutional knowledge becomes executable. A senior engineer's mental model of "how we build things here" can be encoded, versioned, and run. Teams that figure this out first will build faster and onboard more cheaply. The ones that don't will wonder why their competitors' AI output looks so much more consistent.Being first to apply isn't a nice-to-have — it's a structural advantage
Job applications follow the same attention distribution as search rankings. Being first isn't just better — it's categorically different. Recruiters and hiring managers have finite attention, and early applications get a disproportionate share of it before decision fatigue sets in.
On switching to Claude for professional coding — what actually changed
After running parallel workflows across ChatGPT, Gemini, and Claude, the practical differences on complex, multi-step coding tasks are real. The behavior that stands out: Claude will pause mid-approach, flag that a path is suboptimal, and course-correct before generating a broken solution — rather than confidently producing one and waiting for you to catch it.
This matters more than benchmark scores. In production workflows, the cost of a confident wrong answer is much higher than a slower right one. If you're doing non-trivial Python, data pipelines, or anything with architectural decisions baked in — it's worth running your own test. The quality delta is measurable, not vibes.SaaS vendors built moats out of complexity — and vibe-coders are filling them in with Streamlit
Here's the pattern: a legacy reporting tool (BusinessObjects, Tableau, pick your poison) accretes complexity until the SQL underneath it is genuinely unmaintainable. The vendor calls this "stickiness." It's actually just lock-in dressed as functionality.
A team with Streamlit, Python, and an LLM just rebuilt five years of BO reports in an afternoon. Not all of them. The gnarly ones. The 30-minute replications are the proof of concept — the rest follow once the first one lands. If your SaaS moat is "this is too complicated to leave," you've already lost to the next generation of tooling. The vibe-coders aren't coming. They're already here.