yikart/AiToEarn / skills/codex-plugin-aitoearn/skills/aitoearn/references/byted-mediakit-media-understand
byted-mediakit-media-understand
VerifiedUnderstand videos, audio and images with MediaKit MCP tools. Use for media summaries, visual event searches, dialogue analysis, sound descriptions, and timestamped evidence for editing. Use sampled-image understanding, ASR and audio understanding; submit video to an understanding model only when the user explicitly requests direct video understanding. Refine only the intervals needed to answer the question.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 8.6 KB |
| evals/evals.json | config | 5.8 KB |
| references/mcp.md | markdown | 3.2 KB |
| references/sampling.md | markdown | 3.1 KB |
references/sampling.md
View on GitHub# Sampling and time mapping
## Start with coverage
For an ordinary visual overview, approximately 24–48 frames across the requested interval is a starting budget, not an accuracy guarantee. Let duration and the question determine the interval. Short action clips need closer samples; long, mostly static footage can start sparsely. A request for all shots or every occurrence requires systematic coverage and refinement, not a capped overview presented as exhaustive.
With uniform interval I over duration D, reserve about `ceil(D / I) + 1` tile positions, accounting for endpoint behavior. With R rows and C columns, allocate at least `ceil(estimated_frames / (R * C))` sheets. In sprite mode, `snapshot_limit` counts sheets, not tiles. Check the actual returned coverage; the estimate alone does not prove that the ending was sampled. Obtain the last frame separately if needed.
Prefer readable sheets of about 4x4 tiles. Split larger batches rather than squeezing text and faces into one huge image. The service suggests sprite edges no larger than 16384 pixels, but being below that limit does not guarantee readability by the understanding model. Use independent images at source-appropriate resolution for small text or detailed poses.
For long sources, first map the whole requested range, then inspect relevant sections. For complete analysis, work through all sections in manageable batches. Reuse evidence instead of extracting the same frames for multiple questions.
## Keep time honest
- Prefer returned frame times or labels. Only calculate times after confirming the start, interval and layout order; do not assume blank padding tiles represent extra frames.
- Scene-change samples are nonuniform. Never calculate their times from tile indices and a fixed interval.
- For locally decoded media, retain actual frame presentation timestamps, especially for variable frame rate footage. Requested seek times are not automatically exact decoded-frame times.
- For a clip beginning at source time S, map clip-relative time t back to S+t; preserve any additional documented timing transformation or audio offset.
- If reliable mapping is unavailable, cite image/tile positions and say that source timing is unknown.
## Refine and stop
Around a relevant event, obtain before/during/after samples and narrow the interval only as needed. Inspect individual frames when exact visible state changes matter. For sound, analyze the corresponding audio interval; model timestamps and ASR segment boundaries remain estimates unless independently localized.
When estimating synchronization, carry the uncertainty of both event times into the comparison. If that uncertainty is larger than the discrepancy the user asks about, the evidence cannot establish synchronization or desynchronization. Do not substitute a video-model summary for this missing precision.
Stop once sufficient evidence supports the requested conclusion. Report remaining gaps rather than repeatedly increasing resolution or launching whole-video understanding. No fixed frame budget guarantees completeness, and no speed improvement is claimed without measuring a comparable task.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness. | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependencies | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |
| executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is available | Stars: 323k | Updated | Claude CodeCodex | Verified | Install |