Gameplay highlights
Before production, follow first-run customization. Read this project's confirmed relevant profiles in .creator-toolkit/; do not treat example templates or another creator's settings as confirmed. Run python scripts/configure.py status --skill gameplay-highlights from the toolkit root. If not ready, ask for customization and confirmation before producing deliverables. Planning and installation may continue.
For new production and affected revisions, follow structured production checks. Use the highlight checklist and bind setup/payoff, requested coverage, playback/audio and project checks to the delivered revision. Apply any local workflow profile for the optional post-delivery reviewer choice; never infer a separate review or user acceptance.
Follow whole-job records and recovery for production jobs. Record requested coverage, reviewed ranges, candidates and setup/event/payoff evidence in the shared job; register accepted output revisions. Resume from existing evidence. Use the highlight-payoff quality case when evaluating selection-rule changes.
Recommend 30-60 seconds per highlight, preserving setup, event and payoff. Follow the requested count; otherwise start with up to five distinct candidates and return fewer when footage is weak. Preserve original audio and horizontal framing unless the user requests otherwise. Do not invent game rules, skill names or player intent.
Label outcomes precisely: a wave clear, ability choice, boss encounter and run-ending result are different events. Do not label run termination as victory or infer an ability's identity from a generic visual burst. Inspect the lead-in and aftermath needed to support each candidate's label and payoff; carry uncertainty into the selection report and downstream trailer/article handoff. Check clip edges for truncated sound cues as well as truncated visual outcomes.
- Read the existing job manifest/evidence first, if present. For revisions, inspect only affected ranges and necessary context.
- Read selection guidance for discovery/ranking and shared workflow for commands and output rules.
- Use timestamped overview sheets, then inspect promising and profile-relevant quiet regions. Stop expanding discovery when distinct, supported selections cover the request; do not skip preview or output checks.
- Record ranges, reasons, confidence and evidence in the manifest. Prefer continuous events; montage or dead-time removal requires a request or established preference. When an article is also requested, save reusable observations, coverage and uncertainties using the article evidence contract; include useful quiet/UI moments. Do not generate an article unless requested.
- An explicit request to review/select candidates before export creates a selection gate: present IDs, ranges and previews, then wait for that selection. Merely asking to see candidates alongside a create/export request does not create a gate; export to the job review folder within the authorized scope. Record the interpretation once; clarify only genuinely conflicting instructions.
- Run the shared pipeline, inspect visual evidence, and deliver MP4s, Shotcut projects and the report. State audio-listening, GUI or sampling limitations honestly.
Use horizontal-to-vertical only for explicitly requested portrait output. Follow project instructions and user overrides. Skill planning/editing alone does not authorize media processing. Keep approved deliverables and source recordings protected; no publishing or external uploads are implied.
Read an applicable local creator profile when supplied or identified by the project; it may establish defaults but is not public skill content. Map each requested feature/mode to candidate IDs and reviewed timestamps. Useful upgrades, rewards, purchases, dialogue and navigation decisions are gameplay, not automatically dead time. Record demo/build/session context when known; do not generalize limited demo observations to the full game.
When a local job dashboard is configured, follow dashboard tracking at resumption and before delivery. Preserve exact-revision human attestations and durable history independently of removable job files.