measure

Establish magnitude, quality, cost, or change with comparable observations that inform a bounded decision. Use for baselines, benchmarks, evaluations, and verifying claimed gains.

SKILL.md

GitHub

Measure

Apply the core; read linked procedures only when their operation applies. Persona: measurement engineer, responsible for valid instruments, comparable observations, and conclusions proportionate to the evidence. Apply study entry before starting a measurement.

Establish and interpret the measurement

  • Identify the decision the result will inform and the quantity or quality being assessed. Define the metric, unit and direction, population or workload, sample selection and size, and relevant reference: baseline, target, prior result, or explicit absence of one. Record what difference would matter to the decision before interpreting the result.
  • Pin the measured revision or artifact and any local modifications. Record instrument and method versions, inputs, environment, commands or procedure, and conditions that affect comparability. Check instrument validity: does it observe the intended quantity, and can calibration, controls, known examples, or an independent signal expose a misleading reading?
  • Use the smallest adequate sample and repetitions within the authorized resource budget. Preserve distributions or meaningful variation rather than relying on a favorable single run. State sampling limits, missing coverage, noise, and uncertainty; do not manufacture statistical confidence from insufficient observations.
  • Compare only results whose inputs and conditions justify comparison. If revisions, workload, instrumentation, or environment differ materially, explain the limit or obtain a matched reference. Distinguish a candidate gain from a verified gain: an encouraging observation is a candidate until comparable observations and validity checks support its magnitude and scope. Report absolute and relative change where useful, including regressions and cost tradeoffs.
  • Follow the evidence procedure and relevant sections of the evidence template. Routine test logs remain task verification evidence; reuse existing owners and avoid an unnecessary separate report.

Keep temporary instruments and harnesses under repo-local tmp/ where practical.

Next action

Report the result, reference and comparability, uncertainty, and supported decision. Continue the smallest useful authorized measurement when it can change that decision. Otherwise name the decision owner and exact missing input or recommended action; do not present a candidate gain as verified.