Nudge AI to generate exceptional code all of the time.
The quality of code an AI model is trained on roughly follows a normal distribution. The model therefore sees advanced, expert-level technique — but it rarely produces it unprompted, because typical (median) code dominates the training signal. Left to its own devices, the model regresses toward the mean and emits competent-but-ordinary code.
The bet of this project: the expert capability is already in the model; we can reliably pull it to the surface. With the right prompts, skills, rules, and conventions, we can nudge — or push — an AI to generate exceptional code consistently, instead of only when a request happens to land in the upper tail of the distribution.
A growing, version-controlled collection of the artifacts we develop while testing that hypothesis:
- Prompts — reusable instructions that bias generation toward expert technique.
- Skills — packaged, invocable procedures that encode a high-quality way of doing a task.
- Rules — path- or context-scoped constraints that enforce quality automatically.
- History — experiments, before/after comparisons, and notes on what actually moved the needle (and what didn't).
Everything here is treated as infrastructure-as-code for AI behavior: a single source of truth that is reviewable, reproducible, and version-controlled.
- Now: Claude Code. The first artifacts target Claude Code's prompt, skill, and rule mechanisms, because that is the tool in daily use.
- Later: other AI coding tools and models. The underlying ideas (surface latent expert capability; resist regression to the mean) are not Claude-specific, and the repo is structured to grow toward broader support.
Early and exploratory. Expect the structure, vocabulary, and conventions to evolve as experiments teach us what works. Findings — positive and negative — are part of the deliverable.
Contributions, experiments, and counter-evidence are all welcome. See CONTRIBUTING.md for guidelines and CODE_OF_CONDUCT.md for community expectations. Contributors are recognized in CONTRIBUTORS.md.