media-acquisition

Media Acquisition

One shot, one source, decided by a rule rather than by which vendor sounds impressive. Apply it per shot, never per project — a single video routinely mixes all four tiers.

Requires

scripts/prekey.py (keying a generated clip to alpha) ships with this skill — stdlib-only, wants ffmpeg on PATH.

Everything that actually fetches needs pip install 'video-studio-engine @ https://github.com/scrollmark/social-skills/archive/refs/heads/master.tar.gz', invoked as video-studio <name> …. The base install covers stock_archive, stock_pexels, stock_pixabay, stock_shutterstock, gen_image, gen_minimax, gen_replicate and verify_clips (the return check). Add [sourcing] for source_clips (user URLs and licence-filtered search, needs yt-dlp) and [generate] for gen_veo; the bracket takes both at once, [sourcing,generate].

Skip that install and this skill is not usable as a workflow. It is a selection rule wrapped around ten programs; the ladder still reads as a briefing that tells a human which library to open and why, but nothing searches, nothing generates, nothing checks what arrived, and the licence filtering that makes the tiers safe is inside those programs, not in this prose.

The Ladder

1. Does the subject exist in the real world?
   NO  -> GENERATE (fictional product, invented scene, stylised abstraction)
2. Is it public, historical, scientific, civic or natural?
   YES -> ARCHIVE  (free, no key: space, weather, wildlife, infrastructure, record)
3. Is it everyday or commercial — interiors, people at work, food, cities?
   YES -> STOCK    (free with a key, modern catalogues)
4. Otherwise GENERATE. Then: does the MOTION carry meaning?
   YES -> generate video    (a person acting, a process, a camera move)
   NO  -> generate a STILL + camera drift  (~10x cheaper, and steadier)

Ordered by how likely the result is to be wrong, not by quality. Every failure this pipeline has actually shipped came from generated footage — invented pseudo-text on a prop, an unplanned cut mid-clip, a presenter whose face changed between scenes. Real footage has none of those modes, costs nothing, and is already licensed. Generation is the fallback.

Archives go first among the free tiers because they need no credential at all, which makes them the only sources guaranteed to work on a fresh machine. Their honest limit: deep on science, space, history and nature; thin to absent on staged modern subjects. Check what a query actually returns before promising a scene from them.

Tie-break on licence breadth, not catalogue size. Prefer a source whose licence covers the whole catalogue over one that varies per item. An agent choosing hundreds of assets cannot do per-item rights reasoning reliably and one bad pick is a real liability — that is why Pexels outranks the mixed-licence libraries, and why the archive and effects scripts filter by licence rather than trusting search rank.

Shot Type Beats Subject

The ladder keys on the subject, and that is not sufficient. A plastic toy brick is real, everyday and mass-produced, so step 3 says stock — and stock lost all four brick scenes to paid generation: a studio-lit isolated brick came back as a child with wooden cubes, a macro of the underside came back as a family on a rug.

Stock is strong on subjects in context and a desert for objects in isolation. A studio hero shot, a product macro or a mechanism close-up is a generation job even when the object is utterly ordinary.

Generation's known weaknesses — faces, legible text, identity across shots — simply do not apply to a rigid unbranded object, which is why it wins there. Judge generation per shot, not per video.

Vocabulary is a sourcing decision. "block" biases toward wooden cubes and Jenga; "plastic building brick" returns masonry. Search the term people tag with, then re-read the result against the shot you asked for.

Check What Came Back

Search results are untrustworthy in five specific ways and none of them announce themselves. video-studio verify_clips --project <dir> catches all five — two queries returning the byte-identical clip, a monochrome clip in a colour piece, a clip shorter than its scene (it loops and the jump lands mid-shot), a silent score, narration text with no audio. Nonzero means do not build. Orientation mismatch is a warning, not a failure; a 4:3 archival clip in a vertical piece is often deliberate.

Frame-check real footage for unwanted text the same way you would a generated frame. Pseudo-text reads as texture; the stock clip with "MAKE MONEY" burned into it reads as a statement.

Never regenerate an existing clip. Cache by scene and layer id under the project directory. Re-fetching to "make sure it's fresh" costs money and can come back a few frames different, which shifts every cut after it.

Provider-specific failure modes — duration limits that are silently rejected, errors delivered inside HTTP 200, endpoints that rot, billing gates that look like rate limits — are in references/provider-landmines.md. Read it before quoting a cost or planning a batch. Costs and rights per provider live in the studio-setup skill's SERVICES.md; check it before anything is published.

Anti-patterns

• Defaulting to generation because it always returns something. It returns something wrong more often than the free tiers do, and it bills for it.
• Generating video for static b-roll. A 6s clip is ~$0.36; a still is ~$0.03. If the motion carries no meaning, move the camera in the composer instead.
• Asking a generator for readable text. It invents letterforms. Real text is composer typography, every time.
• Offering a source with no key. It wastes the user's turn and makes the cost estimate wrong. Check availability before you offer.
• Treating a user-supplied URL as rights-cleared. They own that call, but say so out loud when the URL is obviously third-party content.