The Lab

Reproducible, honest, small-N-but-documented experiments. Rubrics written before runs, limitations always stated, raw data published, reports immutable once results land.

One-shot prompting vs plan-execute-verify on bug-fix taskspre-registered

A plan-execute-verify loop will complete more of a fixed bug-fix task set than one-shot prompting with the same model and tools, at higher cost per task, with a lower false-success rate.

protocol 2026-07-07
Pilot: one-shot vs plan-execute-verify on seeded-bug tasksresults published (pilot)

A plan-execute-verify loop will complete more of a seeded-bug task set than one-shot fixing with the same model, at higher cost, with a lower false-success rate.

protocol 2026-07-07
Superpowers vs baseline Claude Code on a fixed task setpre-registered

A Claude Code session with the superpowers framework installed will show a lower false-success rate and higher completion on a mixed task set than a stock session, at measurably higher token cost per task.

protocol 2026-07-07
Auto-trigger vs slash-command skills: does auto-triggering actually fire?pre-registered

Auto-triggered skills fire on under 80% of tasks they should handle when prompts are naturally phrased, and the miss rate is driven by description wording rather than task difficulty.

protocol 2026-07-07
Memory vs no-memory on a five-session simulated projectpre-registered

An agent with an indexed-fact-file memory system will repeat fewer mistakes and reconstruct context faster across a five-session project than an identical agent without memory, with the gap widening by session.

protocol 2026-07-07
Skill-trigger reliability across 25 paraphrased requestspre-registered

For a single well-regarded skill, fire rate degrades predictably as request phrasing moves from the description's vocabulary toward natural task language, and a description rewrite following the trigger-test protocol recovers most of the loss.

protocol 2026-07-07