BenjaminStelzer

Benjamin Stelzer

I build practical Agent Skills and developer tools for Codex, Claude Code, and AI-assisted software work.

The projects share one concern: an AI agent can produce a great deal of motion without necessarily producing the result that was asked for. My tools keep the result, the boundaries, and the evidence visible while the machinery does its machinery.

Why this work?

AI agents are useful precisely because they can move quickly across code, research, planning, and writing. The catch is not that they do too little. It is that they can do a convincing amount of the wrong thing: another abstraction, another plan, another polished paragraph, and suddenly the process has eaten the task.

I build small, explicit contracts around those failure points. Not to make an agent timid, and not to turn every rename into a governance summit. The aim is to let it move quickly without losing ownership, meaning, authority, or the difference between evidence and optimism.

How I work

• The result comes first. Process earns its place by making the requested outcome safer, clearer, or easier to verify.
• Facts survive presentation. Better prose, cleaner UI, and tighter summaries do not get permission to improve the underlying truth into something else.
• Boundaries stay visible. A read-only review does not become an edit, a suggestion does not become authority, and a passing check proves only what it exercised.
• Tools stay composable. Each Skill owns one concern and can work alone. Combining everything by default would merely produce a very organized kind of confusion.

Scoville family

• Scoville Brainstorm explores materially different mechanisms before selection.
• Scoville Research turns web, GitHub, and scholarly evidence into a decision-ready, claim-traceable result.
• [Scoville Code](https://github.com/