An iterative spec flow built on a small Makefile and a handful of markdown files — issue, todo, lessons, specs — cuts through the opaque magic of workflow tools. The trick is treating specs as sketches that bound the agent's search space, not exact behavior to confine it. Let AI bringing in world's knowledge!
The week of February 23, 2026: Dorsey's layoffs, Cloudflare's one-engineer Next.js rewrite, Karpathy's tweet. Picking up Conductor, OpenCode, and OpenSpec over a weekend and shipping 15K lines on Parley — agentic coding is no longer the toy it was ten months ago.
LLMs are fundamentally stochastic, but that's not new — the whole ML field always was, and it worked. The interesting question is how to ground stochastic computation in deterministic domains. Two paths: build a deterministic shell around the model (coding), or apply it directly where the domain itself tolerates noise (recruiting, marketing, design).
After more time pair-coding with Claude Code, the awe wears off. It's an overly eager junior with raw smarts and no taste — fast, useful for onboarding, but prone to ridiculous code that you'd reject from a new grad.
Five days, ten thousand lines of Lua, no prior experience with the language. Claude Code is not auto-complete — it's a giant pattern-matching machine that needs grounding through tests and architecture. The 25x figure is cheeky, but the shift is real.
