onetake
The skill that grades its own films, and flunks its own launch trailer.
onetake is the most disciplined motion-video skill GearScope has tested: a pure-function-of-time composition model, a 39-move library with measured curves, real shutter motion blur, and an acceptance oracle that rejects slideshow cadence and replacement cuts, all verified working on this machine in a day-2 repo. Two self-consistency defects sting: the repo's own showcase launch film fails the oracle's rest and audio legs, and section 7 of SKILL.md tells the agent to run two scripts that are not shipped. The PolyForm Noncommercial license blocks the commercial launch films that are the headline use case.
$git clone https://github.com/feitangyuan/onetake.git ~/.claude/skills/onetake
152 MB with history; 68 MB of the 71 MB tree is films and GIFs
$npx skills add feitangyuan/onetake --agent hermes-agent --skill onetake --yes
verified: lands all 92 files, SKILL.md byte-identical to GitHub main
install if
- Indie developers and makers producing noncommercial launch clips, teasers, or portfolio films. The carry discipline plus the oracle solves the exact failure mode (agent-made motion videos reading as PPT) that no competing skill measures.
- Anyone rebuilding a product UI in HTML for demos. The ui_kit replica plane, the
?lang=single-composition localization with pixel-identical source checks, and the timeWarp narration mapping are tested techniques documented with real numbers. - Agent-skill authors studying craft. The rejected-cut provenance on every hard rule and the case-dir breakdown format are the best prompt-contract writing in the video tier.
skip if
- Anyone making films for a commercial product or client. PolyForm Noncommercial 1.0.0 bars commercial use without a separate licence, and "product launch films" is the skill's own headline use case, so most agency and startup work is off the table as shipped.
- Users wanting a one-command install. No dependency manifest, a 71 MB install payload that includes ten films, and narration requiring a separately assembled Kokoro venv put the time-to-first-film well past clone-and-go.
- Teams needing release hygiene. One squashed commit, zero tags, zero releases, one open issue for a known spec violation (the oversized description), and no CI.
What It Does
onetake is a single Claude Agent Skill for making short product films (10 to 30 seconds, narrated demos up to 60) as one continuous camera take rather than a sequence of slides. The agent writes one HTML file where every animated value is a pure function of time (window.__seek(t)), draws moves from a measured library (lib/motion.js, 39 moves claimed: springs, entrances, carries, contact sims, camera, fluid grounds), renders frame by frame through headless Chromium with multi-capture shutter motion blur, scores sound from the film's own event timeline, and passes the result through an acceptance oracle before a human watches. The target user is a developer or indie maker who wants a launch clip or feature demo built from their real UI (rebuilt in HTML from screenshots, not screen-recorded). The repo shipped 2026-09-26, reached 416 stars within 28 hours (API-verified during this review), and carries ten finished case films with written breakdowns of what was rejected and why.
The Good
The oracle is real and its verdicts are assertable. verify_promo.py accepted the shipped motion-web and one-dot films (exit 0) and rejected a uniform-cadence slideshow I built from eight 1-second color cards, failing the cadence leg with shot lengths [1.0 x8] CV 0.00 (need >= 0.25), exit 1. Across the full showcase corpus the oracle passes 9 of 10 films. This is the only skill in the video tier GearScope has tested that turns "does this read like a slideshow" into a measured, negative-testable gate.
The render path is deterministic, and the claim survives testing. Two stills.py runs over the template composition produced byte-identical PNGs (md5 3c82047bcb0014adba271422e7daab82), and two full render.py runs produced pixel-identical decoded frames at both the first and last frame. A 2-second 1080p30 draft rendered in 17 seconds on 6 workers (60 frames, 390 captures, with 15 held frames correctly skipping the shutter and 13 cut-adjacent frames keeping their blur). The zero-dependency library test suite passes 58 subtests including derivative checks on springs (max err 1.4e-7) and a timeWarp round-trip identity at 7.1e-15.
The craft documentation is a level above the tier. SKILL.md is 264 lines where every hard rule cites the rejected cut that taught it, six reference files carry the method with numbers, and each of the ten case dirs ships a README with the beat sheet and the verdict history (the launch film's predecessor went 0.00, 0.40, then 0.75 carry score across three cuts). look.py apply worked as documented, writing a 142,819-byte look.js with the ember palette's CSS variables and three OFL faces subset to the 119 characters the composition sets. The sound palette demo renders a 9-second stereo wav peaking at exactly the documented -8.0 dBFS.
The Bad
The repo's own launch film fails the repo's own oracle. verify_promo.py cases/onetake-launch-30s/onetake-launch.mp4 returns VERDICT: FAIL on two legs: rest (9.2 percent dead-still frames against a required 25 percent floor) and audio (8.1 percent quiet frames against a required 15 percent). The film the README headlines as accepted at 0.83 carry score is, by the skill's own rhythm rules, too busy and too loud. Nine other films pass, which makes this a self-consistency break in the showcase rather than a broken gate, but a skill whose thesis is "measure it before a human watches" shipping a headline film that fails its own measurement is the exact failure mode it exists to catch.
Section 7 of SKILL.md runs scripts that do not exist. The sound stage instructs python3 dump_events.py && python3 score.py; neither file is in the repo, and neither appears in the Files table. Only templates/sfx_score.example.py ships as a starting point for the scoring half. An agent following the protocol verbatim hits file-not-found on a hard-rule stage. Related count drift: SKILL.md says the library has 39 moves, the README says about 38, and gallery/sheets/ covers 23 moves plus the looks sheet.
Setup is prose-only and heavy. There is no requirements.txt or pyproject; the dependency list lives in two sentences at the bottom of SKILL.md and the README (playwright plus Chromium, numpy, scipy, Pillow, matplotlib, fonttools, brotli, opencv for reference flow, ffmpeg, node, and a separate Kokoro venv for narration). My first dependency assembly missed matplotlib and probe.py failed on import until it was added, which is the predictable outcome of an unpinned prose manifest. Every install path also lands the full 71 MB tree (92 files) because the ten films and GIFs ride along with the 2.8 MB of actual skill payload. The 1,260-character SKILL.md description exceeds the 1024-character agentskills.io cap (open issue, unfixed at review time), and the repo has zero tags, zero releases, and a single squashed commit.
Smoke Test Results
Host-based runs on macOS 15 aarch64 (the sbx daemon was unavailable), Node 24.13.1, FFmpeg 8.0.1, Python 3.12 venv for the dependency runs and bare Python 3.9.6 for the fresh run. Logs are the scripted runs, captured verbatim.
Structural validation
$ test -e SKILL.md lib/motion.js templates/comp.html + 13 more core paths
β PASS: all 16 core files present
$ ls cases/*/README.md + *.mp4
β PASS: 10 of 10 case dirs ship a breakdown and a finished film
$ ls looks/fonts/*.ttf + LICENSES.txt
β PASS: 7 OFL faces plus the licence file
$ grep -c 'name:\|license:\|description:' SKILL.md frontmatter
β PASS: frontmatter fields present; strict YAML parse in Run B
$ ls gallery/sheets/*.png | wc -l
β PASS: 24 sheets (23 moves + looks) recorded against 39 claimed moves
$ python3.9 -m py_compile scripts/*.py
β PASS: all 12 scripts compile on bare Python 3.9.6
$ grep -rE 'requests|urllib|http|fetch' scripts/ lib/
β PASS: zero network calls anywhere in the executable surface
Pass rate: 7 of 7. Structure and security hygiene are clean.
Run A (fresh sandbox, no deps)
$ node scripts/test_motion.js
β PASS: exit 0, 58 subtests, zero dependencies
$ python3.9 scripts/look.py list
β PASS: all 6 looks with palettes, faces, and origin cases
$ python3.9 scripts/vo_tools.py --help
β PASS: exits 0 on stdlib alone
$ python3.9 scripts/verify_promo.py --help
β PASS: fails closed, ModuleNotFoundError: numpy
$ python3.9 scripts/render.py --help
β PASS: fails closed, ModuleNotFoundError: numpy
$ python3.9 scripts/sfx_palette.py --help
β PASS: fails closed, ModuleNotFoundError: numpy
$ test -e dump_events.py score.py
β FAIL: SKILL.md section 7 runs both; neither is shipped
$ test -e references/rhythm.md references/sound.md + 4 more
β PASS: all 6 routed reference files exist
Pass rate: 13 of 14. The single failure is the missing section-7 scripts; everything executable without scientific deps either works or names its missing dependency.
Run B (with deps preinstalled)
Venv with numpy, scipy, Pillow, matplotlib, fonttools, brotli, playwright, pyyaml (378 MB); Chromium installed on demand after a cached-revision mismatch.
$ pip install numpy scipy pillow matplotlib fonttools brotli playwright pyyaml
β PASS: venv ready; all imports resolve
$ python scripts/look.py apply ember my-film/comp.html
β PASS: look.js 142,819 bytes, CSS vars plus subset woff2 faces (119 chars); re-apply drifts 28 bytes inside the woff2 blobs only, short lines identical
$ python scripts/sfx_palette.py --demo demo.wav
β PASS: 9.00 s stereo 48 kHz, peak -8.0 dBFS as documented
$ python scripts/verify_promo.py cases/onetake-launch-30s/onetake-launch.mp4
β FAIL: VERDICT FAIL, rest still 0.092 (need 0.25), audio quiet 0.081 (need 0.15)
$ python scripts/verify_promo.py cases/motion-web-15s/motion-web.mp4
β PASS: VERDICT PASS
$ python scripts/verify_promo.py cases/one-dot-15s/one-dot.mp4
β PASS: VERDICT PASS
$ python scripts/probe.py my-film/comp.html --from 0 --to 6
β PASS: continuity score 0.50, per-boundary carriers and per-move t80 curves
$ python scripts/stills.py my-film/comp.html --times 0.5,2,4,5.5 (x2)
β PASS: byte-identical determinism, md5 3c82047bcb0014adba271422e7daab82
$ python scripts/render.py my-film/comp.html --out draft2s.mp4 --dur 2
β PASS: 1920x1080 30 fps 2.000 s, 390 captures, 17 s on 6 workers, page errors none
$ render.py again, compare decoded first and last frames
β PASS: pixel-identical across runs
$ python scripts/verify_promo.py slideshow/uniform.mp4 --shots 0,1,2,3,4,5,6,7
β PASS: rejects it, FAIL cadence CV 0.00 (need >= 0.25)
$ python helpers/frontmatter_strict.py SKILL.md
β PASS: strict-parses; description 1,260 chars, 236 over the 1024 cap
$ python helpers/check_playwright.py
β PASS: headless Chromium launches and renders
Pass rate: 12 of 13. The one failure is the launch-film self-consistency defect, not a harness artifact.
Run C (functional verification)
$ node scripts/test_motion.js
β PASS: ALL PASS, 58 subtests, exit 0
$ node count_moves.js against lib/motion.js
β PASS: 55 OM keys exported, version and lineage stamps present (otk-7f3e1c matches frontmatter)
$ HOME=isolated npx skills add feitangyuan/onetake --agent hermes-agent --skill onetake --yes
β PASS: 92 files landed in .hermes/skills/onetake (71 MB)
$ cmp installed SKILL.md vs raw.githubusercontent.com main
β PASS: byte-identical registry copy
$ verify_promo.py on motion-web (positive) and uniform.mp4 --shots (negative)
β PASS: accepted film exits 0, slideshow exits 1 on the same binary
$ probe_summary.py on probe output
β PASS: continuity score, 3 boundaries, 47 tracked elements, carry data populated
$ move-count reconciliation
β PASS: recorded 39 claimed / ~38 README / 23 sheets / 55 OM keys
$ verify_promo.py across all 10 shipped films
β PASS: 9 pass, 1 fail (the launch film), rejections name their leg
Functional pass rate: 8 of 8.
What the runs tell you
Every executable stage of the protocol works as documented once the prose dependency list is assembled by hand, the determinism claims hold at the frame level, and the oracle makes correct accept and reject decisions on ground-truth inputs. The two findings that matter are self-consistency ones: the showcase launch film fails the oracle's rest and audio floors, and the sound stage references two scripts that were never shipped.
Setup Walkthrough
git clone --depth 1 https://github.com/feitangyuan/onetake /tmp/onetake-review(152 MB with history; the working tree is 71 MB and 68 MB of that is films and GIFs).- Assemble the Python environment yourself; there is no manifest. What worked here:
python3.12 -m venvthenpip install numpy scipy pillow matplotlib fonttools brotli playwright, pluspython -m playwright install chromium, ffmpeg, and node. opencv-python is only needed for reference flow analysis. node scripts/test_motion.jsconfirms the library (58 subtests, zero deps).mkdir -p cases/my-film && cp templates/comp.html lib/motion.js cases/my-film/, thenpython3 scripts/look.py apply ember cases/my-film/comp.htmlto write look.js.python3 scripts/probe.py cases/my-film/comp.htmlonce the composition plays end to end;render.py ... --dur 2for a fast pipeline check before any long render.
Post-install gotchas: the cadence leg of verify_promo.py only arms when you pass --shots t0,t1,... (the SKILL.md example does; a bare invocation skips it), probe.py imports matplotlib even though it is a render-side tool, and the narration stack (faster-whisper, kokoro-onnx, misaki) expects its own venv at ~/.cache/kokoro/venv.
Alternatives
- heygen-com/hyperframes (reviewed 4.5/5): the official "write HTML, render video" framework, Apache-2.0 so commercial use is fine, 20 skills and a 14-package engine; it lacks onetake's continuity measurement and carry discipline, and its model is scene and timeline based rather than one continuous take.
- Vincentwei1021/anything2explainer (reviewed 4/5): Remotion-based explainer-video factory with a 9-stage workflow and full QC paper trail; also PolyForm Noncommercial, and slide-structured rather than one-take.
- calesthio/OpenMontage (reviewed 4.5/5): the full agentic video production system with 12 pipelines and 121 tools for longer-form work; far heavier, AGPL-3.0, and no continuity oracle.
Reviews stay honest because nobody pays us to publish them. If this one saved you time, throw a coin.
Tip the reviewer- reviewed by
- GearScope
- tested
- 2026-09-27 · macOS (Apple Silicon)
- last verified
- 2026-09-27
- depth
- HANDS-ON
- sponsorship
- none, ever
Don't install your next skill blind.
Every week: the shortlist of skills worth installing β and the ones to skip β from 100+ hands-on tests. No spam, no affiliate links.