Start with a question narrow enough to answer
Replay review can become an endless collection of clips unless the analyst knows what they are looking for. Begin with a decision-relevant question: does this opponent rotate early after losing control of a particular area, or do they usually wait for a second cue? The question is not a verdict. It is a filter for gathering examples that can support or contradict a provisional account.
Define the unit of analysis before tagging. It might be a round, possession, encounter, draft phase, objective sequence, or a short interval around a choice. Use a consistent start and end point. If one analyst records the first movement and another records the final commitment, their labels may appear to disagree while describing different moments.
Keep the event and the explanation apart
An observation should describe something visible or otherwise directly represented in the source: a player moved toward an area, a resource was used, a teammate remained in position, or a rotation began after a particular event. An inference goes further: the team may have been responding to a sound cue, a planned timing, or a mistaken read. Inference can be useful, but it needs a separate field and a different confidence level.
This separation protects the dataset from turning an analyst’s first impression into apparent fact. “They panic when pressured” is not an observation. A more careful note might say that, in four reviewed situations where pressure arrived before a resource threshold, the team changed position within a given interval. Whether that response was panic, protocol, or adaptation remains an open question.
Build context into the record
Actions rarely mean the same thing in every state. Capture relevant context such as map or arena, side, score, time remaining, player availability, resource state, patch or ruleset, tournament stage, and the events immediately preceding the action. The goal is not to record every conceivable detail. It is to preserve the variables that might change the interpretation or make two examples incomparable.
Where information is incomplete, mark it as missing instead of silently filling the gap. Hidden information is especially important: a team may react to a cue unavailable to the observer, or know something from communication that does not appear in the replay. An apparently surprising choice can be rational from the player’s perspective.
Sample with care, not convenience
Convenient clips are often memorable clips. A dramatic comeback, a famous opponent, or an unusually clear sequence can dominate review even when it is not representative. Decide which matches and situations belong in the sample, record the selection rule, and avoid using only the examples that support the preferred story. A balanced review includes ordinary rounds, failed attempts, and situations where the suspected behavior did not occur.
Several examples from one match may not amount to several independent pieces of evidence. They can share the same plan, opponent, map, and coaching context. Keep match identifiers and group related observations so later analysis can account for clustering. When the sample is small, say so plainly; a tidy chart does not repair a narrow evidence base.
Make annotations reviewable
Useful annotation is economical and reproducible. A record can include a stable clip identifier, timestamp or round, source location, event label, context fields, analyst name or role, and a short note on confidence. Preserve enough information for another reviewer to locate the clip without rebuilding the trail from memory. If permissions or platform rules limit sharing, store only what is appropriate and point authorized colleagues toward the permitted source.
Develop a shared label guide with concrete examples and edge cases. Ask two reviewers to code a small overlap of the footage, then discuss where their readings differ. Disagreement is informative: it may expose vague labels, ambiguous events, or a genuine interpretive boundary. A model trained on inconsistent categories can produce precise-looking summaries of an unstable idea.
From source to scouting note
Once the material is assembled, summarize the observed pattern with its denominator and time span. “Six of ten comparable situations in the reviewed set” tells a reader more than “usually.” Include counterexamples and define what “comparable” means. If the analysis uses automated event detection or classification, retain a route back to source clips so people can inspect mistakes and verify labels.
Finally, write down what evidence would change the read. Perhaps a larger sample, a different map, a roster substitution, or several new examples in the opposite direction would weaken the hypothesis. This small discipline turns scouting into an updateable practice rather than a fixed reputation. It also lets the team prepare without pretending that yesterday’s footage dictates tomorrow’s choice.
“Evidence is not stronger because the software made the label. It is stronger when another analyst can trace the label back to the play.”Fieldnote on annotation
For the next step, see how evidence becomes a tendency model, or review the wider preparation workflow.