Brief 04 / Analyst workflow
Building a useful review
The deliverable is not a dashboard. It is a short, inspectable explanation of a recurring situation and a question the team can test.
Method brief / Human-led interpretation
The evidence trail
A productive review starts with a question small enough to answer and important enough to revisit. “Why did we lose?” invites a story built around the final result. “What information was confirmed before these entries?” creates an observable task. AI can assist with search, grouping and summarisation, but the review remains a human process of checking evidence, hearing the player's account and choosing what to practise.
Scope the review before collecting clips
Write down the game, version, team context and decision type. Define inclusion rules before selecting examples. If the question concerns coordinated entries, include unsuccessful and successful attempts meeting the same conditions. Avoid choosing only memorable failures. A small, honestly described sample can be useful for discussion; it should not be presented as a comprehensive assessment of the team's behaviour.
Decide what the review will not cover. Mechanical execution, communication, strategy selection and adaptation are related, but trying to solve all of them at once makes recommendations vague. The first session may simply identify whether a cue was available. Further review can investigate whether it was communicated and how the response was organised.
Preserve the original evidence
Retain the replay identifier or authorised source, timestamps, player perspective, patch information and any extraction notes. If a parser or transcription tool is used, record its version and known limitations. An event label derived from a tool is not the original event. Keep a path back to the footage so another analyst can inspect the claim without trusting your intermediate files.
Use only material you are entitled to access and process. Publicly watchable broadcasts are not automatically licensed for unrestricted downloading, redistribution or model training. Private scrims and voice communication require additional care. Publisher rules, tournament agreements and team confidentiality obligations remain relevant even when a tool runs locally.
| Brief field | Purpose |
|---|---|
| Question | States the decision being investigated |
| Evidence window | Identifies clips and selection boundaries |
| Observation | Describes what can be checked directly |
| Interpretation | States a proposed explanation and alternatives |
| Next test | Defines the cue and behaviour to revisit |
Give the model a bounded job
Ask for tasks that can be audited: retrieve similar sequences, suggest candidate event labels, or summarise notes with timestamp references. Avoid asking a language model to infer unrecorded intentions or produce authoritative player rankings from sparse clips. A fluent response can conceal missing evidence. Require the output to distinguish direct observation from inference and to acknowledge information absent from the input.
Review the model's mistakes systematically. Check false matches and missed examples, not merely the most convincing retrieval. Where a classifier is used, examine whether it performs differently across maps, roles or patches. If a model was tuned using the same material used to demonstrate its effectiveness, the apparent accuracy may not transfer to new replays.
Run the discussion without hindsight
Show the sequence up to the decision point before revealing the outcome. Ask the player what they were tracking and which alternatives they considered. Their account is valuable context, though memory also has limits. Compare it with visible evidence without turning the exercise into a test of whether they agree with the model.
Offer more than one plausible explanation where the replay cannot resolve the issue. A late rotation may reflect a communication delay, a deliberate coverage responsibility or an incorrect estimate of opponent timing. Identify what evidence would distinguish these explanations. This keeps the discussion specific and reduces the temptation to replace uncertainty with a judgement about attitude.
Close with one practical test
Choose a behaviour the team can describe in its own language. State the trigger, responsibility and cancellation condition. Agree when to revisit it and what would count as evidence that the original interpretation was incomplete. A practice drill can explore the trade-off, but success in the drill does not by itself establish a competitive advantage.
Store the conclusion beside the original evidence and retain the uncertainty note. If the next patch changes the relevant resources or routes, reopen the question. A useful review system is allowed to revise itself. Its value lies in the clarity of the evidence trail and the quality of the conversation, not in the number of metrics it displays.
