Tag
I built a 25-second loop of zakaria.lu with HyperFrames and Claude Code: seven HTML scenes on paused GSAP timelines, and a frame-by-frame gate that refused to pass a 1 px wobble.
When requirements keep changing, the expensive part is not the code. It is re-explaining everything each session. A method that keeps the spec in the repo and treats every pivot as a reviewed delta.
Two Claude Code mods, Flightdeck and a prompt-cache meter, tested with an 8-agent workflow on a Play/Scala repo. A mod doesn't change the model; it changes what you see while it works, and so what you decide.
I built a small commit-message plugin and ran claude plugin eval on it. The no-plugin baseline showed what the plugin actually adds, and the one red case turned out to be the judge's mistake.
Anthropic's Claude Code team review agent-written code in layers. I mapped their stack onto CV Next: an independent Claude review on every PR, a verdict tied to a commit, and a /ship skill that waits for it. Plus what I left out, and what broke on the first run.
I ran a real feature through Claude Code's opusplan setting, then had Opus deep-review what Sonnet built. The two worst findings were lines in the plan I had approved.
I ran Anthropic's code-modernization plugin for Claude Code headless on a 2006-style billing batch I had booby-trapped. It caught every trap, found four problems I hadn't planted, and still only called its own proof "partly proven".
On a Claude Code subscription the main conversation caches for an hour, but subagents get five minutes. Here is when that costs you a full prefix rewrite, and how to raise it.
I grew a resilience audit from a single agent into a scatter-gather graph with typed edges, an adversarial verification gate and injected failures. Recall barely moved; trust did.
Aircraft maintenance manuals solved the problem of readers who can't ask what you meant. I took ASD-STE100's rules and applied them to tool descriptions, error messages and prompts for AI agents.