franken-manim-mega-skill
Made by JeffWork on or with franken_manim (fmn), the pure-Rust manim rebuild. Use when authoring Rust scenes, rendering 4K MP4s, using the fmn CLI or fmn-python portal, or fixing engine bugs or speed.
From Jeffrey's Skills.md: professionally crafted skills for any AI agent, installed and kept in sync by the jsm CLI. Subscription only, $20/month.
Works With
Quality and Utility come from canonical skill metadata. Community ratings appear separately as stars when subscriber reviews exist.
Subscribe to download
Public previews are available, but downloads, CLI sync, and premium content require an active subscription.
Sign In to Subscribe- Full skill content with deterministic versioning
- CLI sync + local search index
- Access to Jeffrey’s premium library
Save to your account
Saved skills live in your cloud list. Run jsm sync on each machine to pull them locally.
What's New with this Skill
2Updates, enhancements, and release notes for this skill.
Updated: franken-manim, a 4K chapter on a 36-core Mac went from 15.6 to 73.1 fps
Frames travel to ffmpeg over a local socket, and macOS's default 8 KB buffer throttled them. With the buffers at 1 MB, one chapter rendered at 73.1 fps instead of 15.6. Passing --threads 32 runs one render team instead of a 32-thread team and a 4-thread one, and a render-heavy chapter took 367 s instead of 748. The output bytes were the same both times.
- 1 MB local-socket buffers took one 4K chapter from 15.6 to 73.1 fps, with byte-identical output
- --threads 32 cut a render-heavy chapter from 748 s to 367 s
- Troubleshooting entry for text that breaks into strokes during a morph
Measured on a 36-core M5 Ultra with x264 at CRF 16, preset slow, 3840×2160 at 60 fps, 4 to 6 chapters at a time. A sixteen-minute explainer now renders, masters and verifies in about twenty minutes.
- Socket buffers.
sudo sysctl net.local.stream.sendspace=1048576 net.local.stream.recvspace=1048576, with the same settings in/etc/sysctl.confso they survive a reboot. --threads 32. The default plan on 36 cores splits the threads into teams of 32 and 4 that take frames in turn, so the small team sets the pace on render-heavy chapters.
New skill: franken-manim-mega-skill, an agent writes and renders 4K manim videos in Rust with no LaTeX install
An agent writes scenes for franken_manim, the Rust rebuild of 3Blue1Brown's manim, which typesets math without LaTeX and renders 4K60 MP4s with ffmpeg as its only external program. Long videos become chapters under the engine's 64 GiB raw-frame cap, rendered in parallel, joined without re-encoding and verified for decoding, sync, loudness and true peak. Each capability is classed live, partial, design or broken from the code at a named commit, so the agent never promises a missing feature.
- Sizes name their build: 9.2 MB for the explainer program (macOS build) and 13.0 MB (12.4 MiB) for the fmn CLI, against 10,875.4 MB installed for manim CE plus MacTeX on macOS. In a sandbox that blocks TeX, Homebrew and host fonts, the glyph probe PNG comes out byte-identical.
- Math is typeset with no LaTeX install: 99.994% of a frozen 3Blue1Brown formula corpus typesets without error, though only 49 of 111 checked formulas land within 5% of LaTeX's box. A glyph probe sheet comes first, because \rho draws as ϱ and \theta as ϑ.
- One render captures frames serially, so a video renders as one process per chapter: 8 chapters and 30,431 4K frames in 141 s on a GPU host. Chapters are joined by stream copy, which requires that one encoder made them all.
- The 64 GiB raw-input cap divided by 12,441,600 B per 4K NV12 frame gives about 5,523 frames, or 92 s at 60 fps, so chapters stay under about 85 s. The explainer's largest chapter was 4,415 frames.
- Every capability is classed LIVE, PARTIAL, DESIGN or BROKEN from the code at a named commit, with each later change marked by its commit. A CUDA raster is design only, a Metal raster runs only in a CLI feature build and Studio preview, and GPU encoding is the only GPU path that ships.
- The fmn::render facade takes sound and a quality setting and gives ffmpeg 24 h. The fmn CLI drops audio from video and has no quality setting, and both the CLI and the fmn-python portal stop ffmpeg after 600 s.
- Finals are 3840×2160 at 60 fps with explicit quality: CRF 16 for x264 through RenderOptions::video_quality, or NVENC constant quality 16 with 256 kb/s AAC through a runner wrapper. The default CRF 23 encode, 230 to 460 kb/s at 1080p60, was rejected as blurry.
- verify_video.sh fails a file on decode errors, an A/V delta over 0.2 s or a missing video stream, and EXPECT_* variables add resolution, frame rate, BT.709 tags, loudness and true peak. With a loudness expectation set, a file without audio fails.
- A speedup must leave the output bytes unchanged. ab_bench.sh alternates base and candidate runs, checks every exit code and compares framemd5; the recorded 2.2x (22.01 s to 9.97 s for 600 4K frames) kept every frame's framemd5 identical.
- Public claims go through a checklist written after reviewing an X post: where it ran, which binary, MB or MiB, and no speed comparison with legacy manim, which was never measured head to head. The rules rest on 52 recorded incidents and 16 operator cards.
The skill comes from franken_manim's first end-to-end production, the Hoeffding's D explainer (github.com/Dicklesworthstone/franken_manim_hoeffding_explainer): 8 chapters, 3,000+ lines of native Rust and 8:27 of narrated 4K60 video. The README describes the 1.0 design in the present tense, so the skill trusts the code instead: STATUS-LEDGER.md classes every capability as LIVE, PARTIAL, DESIGN or BROKEN at a named commit, and marks each later change with the commit that made it. Rust authoring through fmn::render is live, the fmn-python portal runs 1,052 of 2,814 corpus scenes (37.4%) to