Experiment Narrative
Purpose
Write a transparent result narrative from experiment logs, tables and plots.
This skill is deliberately narrow. It produces a reviewable artifact rather than pretending to complete work that requires missing evidence.
Required input
Start by collecting: hypothesis, setup, metrics, results, artifacts, limitations.
If a required input is missing, ask for it or mark the resulting assumption explicitly. Never fill a factual gap with a plausible detail.
Workflow
- state the hypothesis and comparison first.
- describe only setup details needed to interpret results.
- report effect size and uncertainty before interpretation.
- connect each conclusion to an artifact.
- include failed runs and boundary conditions.
Output contract
Return: A result section that separates observation, interpretation and limitation.
Keep three layers separate:
- Source facts: what was supplied or observed.
- Reasoning: how the facts were organized or interpreted.
- Decisions: what the user should approve, change or verify.
Quality gate
Before delivering, check:
- Baseline and evaluation split are clear.
- Metrics are defined before results.
- The conclusion does not exceed the experiment.
If a check fails, report the gap instead of silently lowering the standard.
Example shape
A ranking experiment narrative reports offline gains, confidence intervals and the missing online validation.
Boundaries
- Do not invent names, measurements, citations, permissions or outcomes.
- Keep private or sensitive material limited to what the task requires.
- Prefer a small, testable artifact over a broad list of generic advice.