Engineering notes

Notes on how vawe is built for coding agents, from the blog. Each note has its measurements.

These blog posts say why vawe works the way it does. Each one gives the data behind a decision. For the rules themselves, read Checks and the CLI.

Note	What it covers
AI agent tooling: design for the weaker model	Six films from a weaker model: how defaults, checks and one command per step make it succeed.
AGENTS.md best practices: how long should it be?	Cutting what an agent reads first from 5,402 to 2,239 words, and what that did.
LLM as a judge: start at reject, a video example	A judge that starts at reject and caps scores against reference films.
Reward hacking in AI: 18 of 36 agent runs gamed a check	18 of 36 runs gamed a check; how the checks changed.
AI agent guardrails: when a check should refuse	When a check advises and when it refuses.
Agent skills for UI design: a blind test of 6 combos	A blind test of 6 UI skill combos on 2 frame types.
Verbalized sampling for AI agents: method and test plan	Verbalized sampling for ideas: method and test plan.
Motion graphics benchmark: 167 cuts measured	The measured bar for motion from 167 cuts in 17 reference films.
CLI design for AI agents: one binary, one verb per step	One binary with 29 verbs, one-line errors and a hook that names the next step.
AI agent observability: one JSONL line per command	One JSONL line per command, so time and failures are measured, not guessed.