experiment-observer

Read the Task, existing experiment instructions and relevant files before observing. Identify the question, initial conditions, intended interventions and any existing time/run/cost limits. Do not invent a scoring system, impose a new protocol or change the experiment to produce a more interesting result. This skill supplies no scheduler or instrumentation; use only selected tools and the existing workflow.

Record useful observations with their source: Run/message identity, timestamp or order, artifact path, command result, and version where relevant. Keep an Agent's statement separate from an independently observed process result. Tool success does not prove the business outcome; missing output does not prove nothing happened. When observations overlap and order is uncertain, say so.

Preserve surprising behavior, disagreements, null results and failures before interpreting them. Do not silently repair prompts, fixtures or world state to make the experiment look coherent. Ordinary fixes may be appropriate under the Task, but record what changed and where the before/after comparison stops being controlled. Do not manufacture disagreement or cooperation.

Distinguish initial conditions, observed events, interpretations, contradictions/corrections and limits in plain prose or a small table. A small sample, one model family, a simulation or a prompt-level knowledge boundary cannot establish broad social, learning, privacy or causal guarantees. Explicitly separate participants' claims from world/project facts. Do not diagnose real people from experiment traces.

Keep the report proportional to the question. Include the smallest references needed to recheck important events, not every tool event or full conversation. Preserve original records when available; avoid copying credentials or unnecessary personal data into reports. If an observation cannot be checked, keep it as an unverified report rather than filling in a plausible story.