media-analysis

Media Analysis

Start every video with a local contact sheet. The CLI combines roughly one frame per second, scene-change candidates and the last decoded frame, with real source timestamps. Inspect the returned sheets before choosing any cloud inference; preparing images alone is not a completed analysis.

Choose the next step after inspection

• Visual questions stay local. Read structure, scene order, composition, visible copy and product presentation from inspected images. Open full-size frames for small text and product details. Refine short shots, brief subtitles and action order with denser local frames; read local frames. A one-second survey can miss an entire short shot. Scene candidates and ending frames improve coverage but do not prove every cut or action was captured.
• Use cloud analysis only for a remaining important speech gap. If spoken content needed to answer the question cannot be established from readable subtitles, use the original retained local video with --cloud. Read cloud execution and recovery. A talking person, an MP4 extension or a desire for a more complete-looking report is not enough.
• Stop when the evidence answers the question. Existing readable reports and frames can answer follow-ups. Do not purchase cloud analysis to confirm a supported visual answer. For an unspecified "analyze this video", deliver the visual/subtitle findings with their scope; resolve a material speech gap when the topic cannot otherwise be understood.

Subtitles are visible text, not verified original speech. Frames cannot establish music, voice tone or unheard dialogue, and cannot exclude an off-screen narrator. Label uninspected audio as unassessed. A failed local preparation remains failed; do not bypass it by buying cloud analysis.

Images remain direct visual inspection: view accessible images, or use postplus media prepare --source <url> for ordered image posts (see image evidence). Keep multiple sources and their evidence, reports and failures separate.

Start with local visual evidence

postplus media analyze video-analysis \
  --video ./reference.mp4 \
  --json \
  --output ./contact-sheet.json

Follow the CLI's structured result and reported next action; do not infer recovery from free-text messages. Use existing PostPlus credits for requested work within existing authorization; estimate-only requests do not execute. Do not recharge, buy subscriptions, publish or change external advertising spend without applicable authorization. An initial reservation is not a final price or a hard cost cap. If a required hard final-cost limit cannot be enforced, do not execute; explain that limit. Correct definite pre-submission field or format errors without changing the requested intent, mode or output constraints. After submission or an unknown outcome, resume the same operation through its returned checkpoint or action; never resubmit uncertain work, repeat exhausted recovery, or switch providers to bypass failure. Keep raw results, operation identifiers and recovery commands in execution records. In ordinary chat, give the result, meaningful limitations, known cost and any necessary user decision; do not narrate uploads, polling, internal identifiers or JSON. When asked how media was made, provide the actual model, submitted prompt in the user language and references actually submitted with their roles.

The default command returns contact-sheet and frame paths for Agent inspection, without uploading for inference. A local video needs no account connection; remote source acquisition can still cost credits. Reuse the returned local file for refinement and any later cloud analysis to avoid collecting it again.

video-analysis is the CLI capability key, not another skill. For the explicit cloud step, PostPlus supplies one fixed analysis prompt for the full original timeline. Do not supply, replace or append custom questions, formats or marketing instructions. There is no short/deep cloud branch. Use command help for local frame controls; use postplus media schema --json only when repairing an unknown cloud input.

Deliver

For local work, answer the user's question with real inspected images, measured timestamps and a brief note on sampling/refinement and audio coverage. Separate visible facts and captured wording from interpretation. A visual timeline survey is not a complete audiovisual record. Do not impose the cloud report template on local observations. Embed supporting images and keep their saved attachments readable after temporary source cleanup; preserve user files.

For cloud analysis, deliver the returned model Markdown unchanged, linking the complete report when long. Do not substitute a summary, marketing verdict, cleaned-up transcript or rewritten shot list. See report scope. Do not schedule an extra Agent completeness review, strategy synthesis, or paid rerun as part of this flow. If separately asked for interpretation or strategy, use the saved original report and deliver that work separately without changing the analysis prompt or replacing the original record.

Keep sources, reports and failures separate. Preserve user files and delivered evidence. Screenshots cannot establish full motion or audio coverage. A failed analysis remains failed; partial screenshots or subtitle observations must be labeled with their limits.

Wait and recover

For source acquisition and cloud tasks, follow structured progress and next actions. After a disconnected wait, use the returned poll/checkpoint action for the same task, account and environment. Do not restart an uncertain paid analysis. Let PostPlus perform bounded recovery; do not resubmit uncertain work, repeat exhausted recovery, or switch providers to bypass failure.

For a clear local parameter rejection before submission, use command help for one correction. For readiness failures, use postplus doctor --skill media-analysis. Use existing PostPlus credits within the requested scope and existing spending authorization. Estimates do not guarantee a final maximum; do not execute an explicit hard final-cost limit that the server cannot enforce. Do not purchase a redo for wording or perceived content quality.

On terminal failure, explain the affected analysis, reason, known cost status and necessary user action. Save the operation id and technical recovery action in execution records. An unknown result or missing usage does not mean no charge. Do not claim unseen or unheard material was analyzed.