Blog

Vibe coding means building software with AI without reviewing the code. Where the term came from, how it differs from agentic engineering, and when it's safe.

A dated, line-item monthly bill for vibe coding an app at three tiers: agent plan, Vercel, Supabase, weather API and domain, plus how to cap surprise bills.

Vibe code a real app with Claude Code in 11 steps: spec, AGENTS.md, plan mode, git, Supabase with tested RLS, Vercel, and a pre-launch checklist.

Next.js + Supabase works on localhost but breaks on Vercel? The real errors, their causes, the fixes, and the agent "fixes" you should refuse.

What Supabase RLS does, why an exposed publishable key isn't the leak, copy-paste owner-only policies, and a second-user test that proves they work.

Yes, if you check what the agent built. The public incidents, an OWASP-mapped pre-launch checklist, and an audit prompt to paste into your agent.

Three ways to make an AI answer from your own notes and courses and cite the exact source, compared on citations, where your material lives and which AI you can use.

An agent harness is the code around a model that turns it into an agent. Here is what it does, and the evidence that it can matter as much as the model.

The top coding-agent harnesses compared on the decisions that matter: tools, edit format, permissions, sessions, extensibility and which models they run.

1M-token windows changed the math, not the answer. When long context is enough, when retrieval wins, and what modern RAG looks like in 2026.

Five business concepts that decide whether an AI-built app makes money: market sizing, contribution margin, break-even, CAC/LTV, and pricing, with a worked example.

Who ran each Jev, Laya and Von benchmark, what the accuracy, ECE and latency numbers mean, and when a frontier LLM with structured outputs is enough.

Four Claude Code workflows for analytics teams: warehouse SQL, recurring reports, dashboards and cleanup scripts, plus a rollout checklist for data leads.

How to build a tracker or calculator with Claude artifacts, no code: a four-part brief, a test loop, sandbox limits, and safe sharing and updates.

Claude artifacts changed in September 2026. What an artifact is now, what's new, and a four-question test for picking a doc, deck, design, file, or page.

A focused way to start learning AI agents without getting lost in frameworks: learn the core loop, work through real examples, and use a 20-lesson course to build durable understanding.

We built a benchmark to pick a model for tool routing. The first version separated zero of fifteen model pairs. Rebuilding it around disagreement produced a result we didn't expect — the accuracy ranking and the safety ranking come out almost inverted.

Two Microsoft courses, 87 lessons, one six-category rubric. Both scored an A — but the interesting part is where they lost points, and what that says about scoring generated courses at all.