Animal Docu Shorts — the AI animal documentary production standard

The production standard of a running AI animal documentary channel, as thirteen Claude Code skills.

Everyone who tries this hits the same wall. The footage looks fine and nobody watches. The animal turns into a slightly different animal between cuts. A leg appears out of nowhere at the moment of a leap. Search for answers and you get "try a better prompt".

This pack is what a working operation does instead.

—

THE MEASUREMENT THAT PAYS FOR THIS PACK

Same channel, same class of subject, changing nothing but the opening cut's picture:

A myth-bust statement over a mood shot — the animal standing, or scenery — held 19.6% of viewers past three seconds, and the feed throttled it to 7.1% impressions. An action peak in the first 1.5 seconds, with an action verb in the first sentence, held 67.3%.

Three times the difference, and a wider margin than any other element of the script. The lesson is not "write a better hook sentence" — the sentence was fine. It was the frame behind it. That single finding restructures how you choose subjects, because if the action peak sits in the middle of the story you have to pull it to the front.

—

COST CONTROL IS A ROUTING TABLE, NOT A PREFERENCE

Every cut declares a motion tier that maps 1:1 to an engine. At one service's measured prices for 5 seconds at 720p, that is 7.5 credits against 22.5 — a factor of three. Over a twelve-cut episode, routed is 165 credits and everything-on-the-expensive-model is 270.

Both directions are errors. Smearing a dynamic cut on the cheap model is obvious. Rendering a static cut on the expensive one is the quiet one, and it is how a credit budget disappears.

—

THE ANATOMY LOCK

Write no extra limbs and you get more limbs — a negative activates the concept, the same way "no dust" produces dust. Every count goes in positively: exactly four limbs: two front, two rear.

And the lock only holds while the part is visible. A foreleg held its amputation stump until 3.3 seconds, was hidden by an explosion of dirt, and from 4 seconds came back as a complete foot with four toes. No prompt prevents that. The fix is cut design: never put unusual anatomy and violent occlusion in the same cut.

The pack also records that a 12-frame inspection strip passed a clip whose defect a human then found by eye — so cuts with unusual anatomy get six frames per second, cropped onto that part alone.

—

WHAT IS IN IT

Thirteen skills covering topic discovery → deep research → script → script review → TTS → stills → i2v clips → subtitles → assembly and render → final review → upload, plus a channel constitution.

Judgement happens in exactly two places: before you spend a credit, and before you publish. Every other skill confirms its own stage finished and grades nothing — because with several places passing judgement the standards drift apart, and then nobody knows which one is right.

A complete worked script is included (hippo, 12 cuts) that passes the pack's own length tool, along with that tool.

—

LANGUAGE — READ THIS BEFORE YOU BUY

This pack came out of a Korean-language channel, and three values carry that origin. All three are parameters rather than assumptions, and the README collects them in one place.

The one that matters: the length tool's characters-per-second default is 6.4, measured on a Korean narrator. The tool counts Latin letters too, and English packs far more of them into a second — leave the default alone on an English script and it comes out roughly twice too long. The pack tells you to calibrate against your own voice in five minutes, gives a derived starting figure of about 12.5, and the tool itself warns you when the default is still in place on a non-Korean script.

Everything else — the routing, the anatomy lock, the hook measurement, the review gates, the subtitle pixel budget which already handles both CJK and Latin glyph widths — is language neutral.

—

HONEST DISCLOSURE

This is a production standard and a methodology, not executable code. Claude Code carries it out, or you fold it into your own scripts; it is not a one-click program.

No outcome or revenue is promised. Channel growth depends on subject matter, consistency and the market.

Statements about AI models — moderation strictness, parameter constraints — were measured during operation and can change with model versions. Each skill includes the method for re-measuring rather than asking you to trust the number. The pack explicitly records that the ranking of which model is more lenient has inverted between versions.

Generation credits for images, i2v and TTS are your own cost. The routing skill exists to reduce that cost, not to remove it.

It takes you as far as making one episode properly. We installed nothing but this pack and produced one end to end — eleven of the thirteen skills, about 200 credits (TTS 7, stills 66, clips 130), passing the format, loudness and review gates. Producing several a day at the same quality is a different problem, with automatic gates and a production ledger, and it is a separate product.

Shorts only (9:16, 40–55 s). Mid-form and long-form are a different process, sold separately.

—

REQUIREMENTS

Claude Code or a Claude agent environment · image generation · i2v video generation (the routing assumes two model families) · a TTS with per-sentence synthesis, file download and re-synthesis · ffmpeg · licensed BGM. The YouTube Data API is optional — the upload skill includes a manual checklist.

—

LICENCE

Producing for your own or your company's channels is covered by the base licence. Delivering to third-party clients as agency work requires a separate agency licence. Reselling, redistributing or reproducing the skill text is not permitted. Full text in LICENSE.md.

Support: questions on using the skills for 30 days from purchase, five exchanges. Environment debugging and per-project consulting are not included. Updates are for life — re-download from the purchase page.