ai-shortfilm-prompts: a Claude Code Skill that writes film-grade video prompts

The problem: prompt advice that is mostly adjectives

Text-to-video models respond to concrete physical description and ignore praise. The README makes this the central argument. Its naive example reads "Epic cinematic shot of a beautiful female mech warrior... Highly detailed, 4K, photorealistic, movie-quality, dramatic lighting," and the repository's claim is that words like epic, stunning and movie-quality give the model nothing to anchor to, so the output lands as generic game CG.

The target reader is someone who already generates video and has hit that ceiling: the clips look like stock renders, the camera drifts without intent, and the same idea produces a different result every attempt. The project's answer is to replace taste adjectives with camera bodies, lens specifications, physical reactions and an explicit ending. It is aimed at people who write prompts by hand and want a repeatable structure, not at people who want a hosted service that generates the clip for them.

How the 5-stage structure and the 21 genre templates fit together

The mechanism is a Skill: a directory of instructions and reference files that Claude Code loads when the task matches. The repository layout shows .claude-plugin/ for plugin metadata and skills/shortfilm-prompt/ for the Skill itself, with examples inside it. The README points to skills/shortfilm-prompt/examples/02-skill-output-sample.md as a full sample output with a ten-item self-check, which suggests the Skill is expected to produce a structured document rather than a single line.

The structure has five stages. The README's excerpt shows four of the labels in use: core theme, atmosphere, camera, and a timed beat section, plus an ending. The core theme line packs genre, location, aesthetic, subject and treatment into one pipe-separated string. Atmosphere names the camera body and lens. Camera describes the movement. The timed section gives second ranges with physical events. The ending states what does not happen, which is the part most hand-written prompts omit.

Around that structure sits a library. The v0.4.0 release notes describe a 17-genre library, while the README and the browser builder both say 21 genres. The difference is not explained in the release notes, so treat the count as version-dependent and check the templates directory in the revision you install. Alongside the templates, v0.3.0 added a cheat sheet, preset negative prompts, a failure-case library and a machine-readable index, and cases.md is described as a failure-to-fix gallery. That pairing matters: a template library without failure cases teaches you to copy, not to diagnose.

Installing the Skill and getting a usable prompt out

The README advertises plugin installation through Claude Code, which is what the .claude-plugin/ directory is for. The repository does not print the exact plugin command in the install section available here, so use the install section of the README rather than guessing a subcommand.

If you would rather not install anything, the README offers a browser builder at prompts.aiolaola.com/en/build.html that runs entirely in your browser. The stated flow is: pick one of 21 genres, fill a few blanks, and receive the full multi-shot prompt plus a negative prompt. The README also notes there is a one-click copy path on the site for any template or case.

The lowest-commitment test is the single-line prompt the README publishes. It is meant to be pasted directly into Sora, Seedance, Kling or Veo with the placeholder replaced:

Anamorphic widescreen cinematic. Simulated IMAX film camera +
Panavision C-series lens (35mm focal, f/4 aperture). Handheld
shot — extremely subtle, breath-like camera float throughout.
{{your scene description}}.
No score. Production audio only.

What you should see is a clip with a specific optical character rather than a generic cinematic wash, and production audio instead of a generated score. The repository's explanation for why it works is that real camera bodies plus a described float anchor the model to film aesthetics. That is the project's claim, not a measured result.

For the full Skill path, the README points at methodology.md for the breakdown and examples/02-skill-output-sample.md for a complete output with the ten-item self-check. The check is the part worth adopting even if you never install the plugin: it gives you a way to grade a prompt before you spend generation credits.

Where the method breaks down

The repository is a prompt library, not a renderer. Nothing here generates video, queues jobs, or reports cost. The README links to a separate project by the same author, OpenShorts, for turning a story into a three-shot mini-drama with per-shot cost quoting, which is an implicit admission that this repository stops at the prompt.

The Skill is also coupled to Claude Code. The plugin path is the documented install, and the alternative is a browser page. If your team writes prompts in a different editor or inside a generation platform's own UI, you are copying text between tools. The tags list other environments, but the README's install section is about Claude Code.

There is a documentation gap around versioning. The README says 21 genres, the v0.4.0 notes say 17, and the release notes do not reconcile them. A prompt library lives or dies on which templates you actually have, so that discrepancy is worth resolving before you build a workflow on a specific genre.

Finally, the project inherits a single film's look. The method is distilled from Zombie Scavenger, and the README's own example carries that lineage in its core theme line: "gritty hard sci-fi mech | rainy dock | battle-damage aesthetic." The aesthetics it encodes are handheld, low-saturation, film-grain, battle-damaged. If you are producing clean corporate footage, product animation or children's content, the templates will fight you, and the negative-prompt presets tuned for that look may remove exactly what you want.

The alternative: writing prompts from a model vendor's own guide

The obvious alternative is not another repository. It is the prompt documentation published by whichever model you use, plus your own notes. Sora, Kling, Veo and Seedance each ship guidance written by the people who trained the model, and that guidance tracks model updates faster than a third-party library can.

The difference in approach is scope and specificity. A vendor guide tells you how that model weights camera language, duration and audio. This repository gives you a fixed five-stage skeleton, a genre library and a failure gallery that are model-agnostic, and the README's single-line prompt is explicitly written to be pasted into any of the four. The trade-off is that a model-agnostic structure cannot tell you that one model ignores your lens specification while another over-weights it.

A second alternative is the browser builder at prompts.aiolaola.com, which is the same project without Claude Code. That is not really a competing tool; it is the escape hatch for people who do not want a plugin. The genuine choice is between a curated, opinionated structure and the vendor's own documentation. If you generate across several models, the structure is worth more. If you have standardized on one, the vendor guide is closer to the source.

Licence, maintenance and the cost of upgrading

The repository is MIT licensed, with LICENSE and NOTICE files at the top level plus a NOTICE.zh variant. MIT is permissive: you can reuse the templates and Skill files in commercial work, provided you keep the copyright and permission notice. The presence of a NOTICE file suggests there is attribution material the project wants carried along, so read it before you redistribute the templates inside your own product. This is a description of what the files say, not legal advice; if you are embedding the library in something you sell, have counsel read the NOTICE.

The last push was on 2026-09-15, and the repository is not archived. The release cadence visible in the notes runs from v0.3.0 and v0.3.1 on 2026-06-05, through v0.4.0 on 2026-06-24, to the September push. The v0.3.0 notes describe a broad batch of additions (cheat sheet, negative-prompt presets, failure-case library, machine-readable index, dead-link CI), while v0.3.1 is a one-line plugin skill path fix. That pattern is worth knowing when you upgrade: most releases add material, and occasionally one fixes the install path, which is the kind of bug that makes a plugin silently stop loading.

Upgrade cost is low if you consume the templates as text. It is higher if you have forked the Skill directory, because the plugin path has already changed once and the genre count has shifted between the README and the release notes. Pin the revision you use and diff the templates directory on upgrade rather than pulling the default branch.

Editorial conclusion

Adopt it if you already write prompts for Sora, Kling, Veo or Seedance and want a fixed structure instead of adjectives, and if you work inside Claude Code or are willing to use the browser builder. Skip it if you need the model to run the generation, if you want a hosted API, or if your workflow is a GUI with no command line. Before committing, read methodology.md and cases.md, run the Skill once on an idea you have already shot, and compare the output against the ten-item self-check in skills/shortfilm-prompt/examples/02-skill-output-sample.md.

Frequently asked questions

Official sources

1. jnMetaCode/ai-shortfilm-prompts on GitHub
2. License: MIT
3. Project website
4. README
5. Releases