Science

Use the Quick Diagnosis workflow: state a one-sentence goal, generate at least three hypotheses with falsification tests, rank them by test speed and likelihood, then test and iterate. It typically resolves issues in minutes that random attempts would take much longer on.

The cycle runs GOAL, OBSERVE, HYPOTHESIZE, EXPERIMENT, MEASURE, ANALYZE, then ITERATE. Each phase has a dedicated workflow file with checklists, and the loop repeats until the goal is achieved or explicitly abandoned.

Use Quick Diagnosis for problems solvable in under 15 minutes and Structured Investigation for hours-to-days efforts. Reserve Full Cycle for weeks-long, high-stakes work needing pre-registration, statistical analysis, and peer review.

Generate a minimum of three hypotheses, and five to ten for important problems. Single-hypothesis thinking causes confirmation bias and missed alternatives; always include a null hypothesis that nothing needs to change.

Yes, prompt optimization maps directly onto the cycle: pre-commit eval criteria, generate prompt variants as hypotheses, and measure with structured evaluation. The documentation directs prompt experiments to a dedicated Evals skill for statistical rigor.

The overhead is not justified when the fix is obvious, the pattern is well-known, or creative generative work is needed. The workflows include explicit anti-triggers so formality is only applied when the scale of the problem earns it.