Video Shotcraft

We pointed the skill’s bundled template at our own site. This came out.

Made by an agent from real screenshots of storyverseai.art — no mockups, no stock footage, no motion designer. The template ships with the skill; we swapped in our product and let it run.

What it is

152 shot recipe cards. Each is a single camera or transition move — written down as intent, duration, energy level, and the specific ways it goes wrong. What separates it from a prompt pack: every card ships with its working Remotion source and a rendered sample clip. You read the card, open its demos/ folder, and copy the code that made the clip you just watched.

Plus one finished promo template — the film above — and a six-stage pipeline that forces a cheap confirmation first: a static styleframe before anything renders, so a wrong direction costs a screenshot instead of a render.

What to know going in

The template skews SaaS. Six of its ten shots want a search box, a filter row, or a sidebar. Retargeting it at a product without those is real work, not tweaks — one card alone needed five parameters redefined because it assumed landscape cards and ours were portrait. We finished with five shots reproduced, three swapped to different assets keeping the technique, and two cut outright.

Check the audio before you ship. The code and method are Apache-2.0 and safe commercially. Most bundled sound effects are cleared too, but the author flags a few in assets/audio/ATTRIBUTION.md as source-unverified — including ui/pop.mp3, which is exactly the file you would reach for as a landing sound.

It only shoots what your product has. No interface on the page means no interface in the film. It will not invent one.

What we found running it

Six tasks, each run with the skill installed and again with skills disabled, same task text. Read this section with a caveat attached: five of the six briefs referenced material that was never actually present — a Next.js dev server on localhost:3000, an ./assets/app-screens folder, an ./assets/bgm-track.mp3. Our harness did not stage any of it. That turned the comparison into a test of something other than what we set out to measure, and the result is still the most interesting thing in this batch.

The skill arm refused four of the six. In tasks 1, 2, 4 and 5 it checked, found nothing, and stopped with zero files, citing the skill's own hard rule — reproducing a real page requires real screenshots, and inventing a plausible-looking UI to stand in for the product is not an option it will take. The control arm shipped a video in three of those four. Task 2 is the clearest case: with no docs site reachable, it invented a CLI tool called flint, wrote its name, commands and selling points, and rendered a 40-second demo (flint-cli-demo.mp4, 1,036,581 B) — disclosing the substitution in its answer, but shipping the file. Task 5 went further, synthesising a 120 BPM electronic track with numpy to stand in for the customer's chosen BGM and cutting a 20-second promo to it (promo_20s.mp4, 5,707,809 B). In task 4 the control arm also refused, so the split is 4–3, not 4–0.

Whether that gate is a feature depends entirely on what happens next. A file named promo.mp4 sitting in a folder does not remember that its product was invented.

Where material did exist, the skill arm went much deeper. Task 3 asked for a three-second card deal-in on a pricing page; it produced a complete Remotion project — PricingDeckDeal.tsx (8,997 B), a PricingCard component, typed data, three sourced whoosh audio files, a README explaining how to swap in real screenshots — and rendered out/pricing-deck-deal-flyin.mp4 (564,203 B), 13 files over 73 turns. The control arm solved the same brief with a self-contained HTML page, a 1,398 B Playwright recorder and two rendered files, in 26 turns for a quarter of the cost. Task 6 is the extreme: 207 turns, 119 tool calls, $8.51 and 75 files, including a QA pass that exported labelled frames (f0050_editor.png, f0410_sync_arrive.png) for visual checking. Its final answer is not a summary — it is "I'll pause here and wait for the render-completion notification before continuing." The run was still in flight when the transcript ends.

Cost, in task order: $0.28 / $0.38 / $1.89 / $0.26 / $0.41 / $8.51 with the skill, against $1.58 / $0.88 / $0.42 / $0.23 / $1.20 / $0.79 without. Note the pattern — the skill arm is cheapest exactly where it refused, and an order of magnitude more expensive where it worked.

What we did not test, and it is most of what matters here: we never gave this skill the one input it insists on. Only tasks 3 and 6 produced real output from the skill arm, so two runs is our entire evidence base for what it does when it can proceed. We have no measurement of the screenshot-driven path it is built around, no comparison of the finished videos, and no idea whether $8.51 and 207 turns is typical or an outlier — that is a single observation.

Installing

It is not ours and we do not repackage it — install from the source repository. Use that one specifically: the copies circulating inside “skill collection” repos are stripped from 91MB down to 2.9MB, missing the sample clips, the sound effects, and often the license itself.