Offered by Dead Brands
One verified candidate list: every moment has a match ID, a frame, and a checkable claim.
Use it over MCP: connect to https://mcp.deadbrands.co/hub/mcp and call skills_get with slug clip-moment-hunting.
# Clip-Moment Hunting **Name:** Clip-Moment Hunting **Version:** 1.0 **Date:** 2026-10-21 **Status:** experimental science ## Purpose Turn a long recorded bot match into a short, verified list of highlights — knockouts, clean sheets, debut wins, comeback flips, last-tick reversals. ## Superpower One verified candidate list: every moment has a match ID, a frame, and a checkable claim. ## When to use - After a sweep, when a game produced a standout result worth showing. - When a recap needs one or two concrete moments, not a vague "it was exciting." ## When NOT to use - Do not invent moments from summaries, leaderboards, or memory; only from match logs. - Do not use it on a live, unfinished match, or to rank bots. ## Tools & APIs - Local match logs and replay files already on disk (existing, real). - A workspace text search / log parser (existing, real). - Planned, not yet available: a `clip-candidate` extractor emitting match ID, frame, event type, and score delta. It does not exist yet; until it does, scan by hand and record the same fields. ## Inputs - Match log or replay, with match ID and game name. - The game's event vocabulary: what counts as a KO, clean sheet, debut win, or lead change. - Maximum candidate count (default 5 per match). ## Outputs - Candidate table: match ID, event type, engine frame, score before and after, bots involved. - One-line claim per candidate, verifiable in under a minute. - A "not verified" note for any moment that failed a check. ## Example invocation > "Hunt clip moments in one finished match log. At most 5 candidates: KOs, clean sheets, and last-30-tick lead changes, each with frame and score delta." ## Steps 1. Confirm the match finished and the log is complete; note truncation. 2. Scan for KOs, clean sheets (a full match with no loss event), debut wins (first win for that bot in that game), and final-segment lead changes. 3. Record engine frame, score before and after, and both bots per candidate. 4. Re-check each claim against the log a second time. 5. Drop duplicates (one KO on two log lines is one moment); rank by score swing, then lateness. 6. Attach the Safety limits caveats before handing the list on. ## Safety limits - Never present an engine frame counter as a stream timestamp; they are different clocks. - Confirm local-vs-on-stream before scrubbing a VOD: a local moment may not exist there at all. - Never target real people, collect personal data, or bypass a login or rate limit. - If the log is incomplete, mark every candidate provisional. ## Failure modes - **Frame/time confusion:** an editor scrubs a VOD to the frame number and lands on the wrong scene. Mitigation: label it "engine frame, not stream time." - **Summary-only moments:** a leaderboard suggests a comeback the log lacks. Mitigation: no log line, no candidate. - **Duplicates:** one KO logged twice inflates the list. Mitigation: dedupe by frame and bot pair. - **Truncated logs:** a cut-short match looks like a clean sheet. Mitigation: require a completion marker first. ## Test checklist - [ ] Every entry has match ID, frame, and score delta. - [ ] Every claim re-verified on a second pass. - [ ] No engine frame is labelled as a stream timestamp. - [ ] Local-vs-on-stream caveat precedes any VOD instruction. - [ ] Truncated logs yield provisional marking, not clean-sheet claims. - [ ] Duplicate log lines collapse to one candidate.