Nº 135 · AI ·15 min read · November 16, 2026

99 Percent of What? Five Claude Skills and Their Units

99 Percent of What? Five Claude Skills and Their Units
Fig. 01 99 Percent of What? Five Claude Skills and Their Units

About a century ago, the Soviet filmmaker Lev Kuleshov is said to have cut one shot of an actor's blank face against three different images: a bowl of soup, a girl in a coffin, a woman on a divan. As the story goes, the audience saw hunger, grief and desire, and praised the acting. The face was the same piece of film each time. The footage is lost and the accounts differ on the details, so take the three images as the legend tells them.

Five READMEs brought it back this week, because each one leads with a number: 99.6 percent, 25 seconds, under 45 cents, 83 minutes, 70 to 80 percent. Each of those is the blank face. What it means comes from the shot beside it, and that shot usually sits a few screens further down: percent of what, seconds of which step, cents for which part, minutes until which moment.

How I checked

This is a documentary review, not a hands-on test. I read each repository's own files on November 16, 2026: the README, and the results file, skill file, intake template or limits section where the unit is written down, named in each entry. Stars, forks, licence and last push come from GitHub in one snapshot that morning, at 10:11 UTC. Where a claim depends on one file, the link goes to that file at a fixed commit, so it still says what I quote after the repository moves.

I did not install or run any of the five. Every number below is the author's unless I say I counted it. Four are under the MIT licence and one, rea-video, under Apache 2.0.

1. rea-video

A command-line tool with an agent skill that takes a finished video apart shot by shot and says how each shot was made: filmed, generated by AI, or built in code.

Link. github.com/JuneYaooo/rea-video, by JuneYaooo.

What it does. It cuts the video into shots and splits the audio into voice, music and effects. Then it gathers evidence for each shot, such as frame cadence, sensor noise and AI-content tags in the metadata, and makes a call. You get a report to read, and a layered folder of text files in an open format called VIDEO.md for your agent to read. The report also lists captions with their timing, where the cuts fall, the music's tempo, and what you would have to record yourself. A second part plans a remake, grouping shots by what a given video model can take in one generation. It does not render any video.

Who it is for. The creative at an agency who received a reference reel with the note "like this", and has to say before the quote which shots need a camera and which need a model.

What it saves. The job it replaces is stepping through the reference frame by frame and guessing. The command-line guide says a 30-second video takes about a minute without a vision model. That is the author's figure.

Cost to install. No command to type: the README has you tell your agent to install the repository and break down a file. You need ffmpeg and Python 3.10 or newer. On first use it downloads the separation models, which the guide puts at about 67 MB for voice and about 149 MB for sound effects.

Install https://github.com/JuneYaooo/rea-video, then break down this video: /path/to/video.mp4

The catch. The accuracy table is honest and worth reading column by column. The 99.6 percent is the how-it-was-made call, weighted by shot duration, on a held-out set of original files. One column to the right, on ten videos re-encoded by a chat platform, the free offline tools score 88.6 percent, and the README explains that recompression wipes out the noise evidence. A reference a client sends is nearly always a platform download. For that file, plan on the second or third row of the table: a vision model you pay for, or an agent that can look at frames. Either way the reference leaves your machine for that provider.

The percentage also covers the picture only. The sound has its own section in the results file, which is in Chinese. With the tools alone, detecting sound effects passed on 2 of 4 videos and naming the source of the music on 2 of 5. The three misses were narrated videos whose generated music was judged to be stock library, and the file says why: a generated instrumental and a library instrumental cannot be told apart from the signal. Read the report's line on where the music came from as a guess unless a platform tag backs it.

Verified. November 16, 2026, at commit b5cbfee. 9,439 stars and 1,393 forks, the largest repository this week; Apache 2.0, created October 29, last push November 16. The README's table says 15 tuning videos and 16 held out, and the results file lists 15, 9 and 7. The README comes in twelve languages, Portuguese among them, and says the deeper docs are in Chinese for now. Read: the README, eval/RESULTS.md and docs/cli.md.

2. music-skill

A skill in which the agent writes instrumental music note by note, and scripts on your computer perform it on sampled instruments, mix it and master it.

Link. github.com/owoevan-dev/music-skill, by owoevan-dev.

What it does. No audio-generation model is involved. The agent writes a score into a project file, a reviewer script questions it (clashing notes, parallel fifths, a melody that ignores the plan), and a renderer plays each part on recorded instruments. You get an MP3 and a WAV, the MIDI, one stem per part and the project file, in which any single note can be changed and rendered again. Because it is written to numbers, it can hit an exact length, loop cleanly, or peak on a given bar. The master lands on a loudness target, -16 LUFS by default and -14 for streaming.

Who it is for. The editor who needs a 40-second bed that swells on the product reveal, drops under the voice, and can have four bars moved when the cut changes on Friday.

What it saves. It stands in for regenerating a whole track to fix one moment, and for digging through a music library for something that almost fits. The README's examples rendered in 25 to 156 seconds on a 24-thread laptop. I did not time anything.

Cost to install. Two commands in Claude Code, and then the heavy part. The renderer and sound banks are about 3.7 GB of downloads, and with all thirty banks installed the README tells you to keep about 12 GB free. It needs Python 3.10 or newer, FluidSynth and FFmpeg. No GPU and no account.

/plugin marketplace add owoevan-dev/music-skill
/plugin install music-skill@music-skill

The catch. Twenty-five seconds is how long the machine takes to make a file. Nobody has heard that file yet, and the repository says so in three places. In the limits: "The agent cannot listen", and a person has to hear the result. In the skill's own rules, the agent is told never to say or imply that it listened. And under the 160 automated checks, which test the plumbing: whether the music is good "is another matter, and only listening can tell".

So the real unit of time here is your ears, per revision. The limits also tell you where the free instruments run thin: no steel-string acoustic guitar, no choir, no vocals of any kind, and section strings that speak slowly. If the brief is a folk guitar under a voice, this is the wrong tool this month.

Verified. November 16, 2026, at commit 424e7bd. 234 stars and 27 forks; MIT, created November 2, last push November 14. The README presents six example pieces and the examples folder holds six, plus a before-and-after demo. Read: the README, PRIVACY.md, the sources-and-rights reference, and the listening rules in SKILL.md.

3. Papermorph

A skill that turns a textbook PDF, a story or a book into narrated, animated web pages: lessons, paper-puppet picture books, short documentaries and vertical shorts.

Link. github.com/DozenTwelve/Papermorph, by DozenTwelve. The live bookshelf lets you judge the output before installing anything.

What it does. It plans a book in chapters or scenes, writes a script and a storyboard, voices the narration so that every word has a time, and animates the picture to those times. Then the model looks at its own output through contact sheets and blank-frame checks and fixes what it finds. The result is a folder of static pages with no server, database or account behind it. Any chapter can also be exported as a 1080p MP4 with an .srt, or packaged as SCORM for a learning platform.

Who it is for. The producer at an educational content studio who has to show a publisher one real animated chapter before anyone signs for forty.

What it saves. For a pilot chapter, it covers the storyboard, the animatic and the narration sync in one pass. I have no time figure, and the README gives none.

Cost to install. One command. It needs uv, ffmpeg and Playwright's Chromium and Firefox. Narration is free through Edge TTS online or Kokoro on your machine. ElevenLabs voices, clips made from a still, and painted sets are optional, paid, and need an API key.

npx skills add DozenTwelve/Papermorph --skill papermorph --agent claude-code

The catch. Two prices sit four lines apart. The first says the newest picture books cost under $0.45 each in voices. The second says an illustrated physics chapter costs about $10 to $18 of model usage. The cents are for one line of the budget. The dollars are per chapter, and the physics book on the shelf has 9 chapters. By my multiplication of the README's own range, a book that size is $90 to $162 in model usage before any paid voice or painted set. The algebra book beside it has 68 chapters. Price the series by the chapter before you promise it by the book. The skill keeps a cost ledger, so the first chapter will tell you your real rate.

On source material, the FAQ gives a short answer to a long question: you can make a book from any PDF for your own use, and publishing one based on someone else's textbook is yours to clear. The project's own examples use public-domain books or pictures drawn for them. And install it with the flag above. The skill folder is 141 files and about 1.6 MB, while the whole repository, shelf included, came to 7,753 files and roughly 778 MB when I added up the file listing.

Verified. November 16, 2026, at commit 618aeea. 1,734 stars and 221 forks; MIT for the code, with pictures and recordings under site/ keeping their own licences; created October 3, last push November 12. Read: the README and the skill's SKILL.md.

4. astro-conversion-kit

Eight skills that rebuild a live WordPress site as a static Astro site with the same addresses, content and look, and prove it with scripts.

Link. github.com/ethanplusai/astro-conversion-kit, by ethanplusai.

What it does. It reads the public site slowly, without logging in: every page's HTML and head tags, redirects, media, fonts and screenshots. It rebuilds the pages in Astro and runs gates against that saved copy on every commit: routes, redirects, search tags, visible text, links, media. Then it compares screenshots pixel by pixel at three widths. A run ends with a pull request, a staging deploy that tells search engines to stay away, and a launch runbook. It refuses to touch DNS, log in, submit a form on the live site, or handle more than one site per run.

Who it is for. The producer at a small studio whose brochure site runs on WordPress and needs a plugin patched every other week, with nobody on staff whose job that is.

What it saves. It replaces a manual rebuild and the page-by-page comparison that follows it. The author's average over 45 conversions is 83 minutes.

Cost to install. One command, or one sentence to your agent with the site's address. Requirements are strict: Node 22.23.4 or newer on the 22 line, or 24.22.1 or newer; git; Playwright's Chromium; macOS or Linux, with Windows not supported. A Vercel or Cloudflare account is needed for the staging deploy.

npx skills add ethanplusai/astro-conversion-kit

The catch. The 83 minutes run "intake to pull request". Merging, moving DNS, exporting Search Console data and the 30 days of watching afterwards are on your side of that line. What lands in the pull request is also a faithful copy: the default build reproduces the live site, "faults included", and broken links or missing alt text arrive as a list of suggestions. Search, comments, carts and logins are reported as out of scope.

The default I would change is in the one message the agent sends you before it starts. Staging stays public unless you say otherwise: hidden from search engines, and open to anyone who has the link. For a client's site, answer yes to protection in that message, because the skill is built never to ask a second time, and replying "go" accepts every default.

One more thing I looked for and did not find. The capture needs no login, so it works on any public site, and the demos show it: the README links rebuilt copies of sxsw.com, harvard.edu and angrybirds.com, which the examples file describes as not affiliated with those organisations. The intake template has ten questions and none asks whose site it is. That question is yours to ask before you type the address.

Verified. November 16, 2026, at commit 8beb1de. 132 stars and 25 forks, the smallest repository this week; MIT, created October 22, last push October 31. I counted eight folders with a SKILL.md. Read: the README, docs/faq.md, the intake template, and the last section of docs/examples/README.md.

5. skill-diet

Two skills that rewrite your other skills to be shorter, and test that the short version still does the job.

Link. github.com/nicobailon/skill-diet, by Nico Bailon.

What it does. It sorts everything a skill says into levels. The description is loaded in every session, the main file when the skill is used, reference files only when a case needs them. It cuts what the model already knows and what was written for people, then runs agents on the same tasks with the old skill, the new one and none, and keeps the slim version unless it does worse. A second skill, skill-model-gap, runs a skill on several models and looks for why the weaker one fails.

Who it is for. Anyone who has been installing skills from columns like this one since August. Every description in that folder is read in every session, used or not. It also fits the person who maintains a studio's house skills, such as the brief template and the file-naming rules.

What it saves. Context, which is money and attention on every request. The author reports cuts of 70 to 80 percent of the tokens on real skills. I did not measure any.

Cost to install. One command in Claude Code. The test then needs a small adapter script that your own agent writes, which took from under 4 to 16 minutes in the author's trials, and it runs real sessions that you pay for.

/plugin install skill-diet --marketplace nicobailon/skill-diet

The catch. The full sentence is that a diet cut 70 to 80 percent "with no drop in test results", so the test is the unit. It is 10 to 15 one-line requests, each run three times per version. The limitations spell out what follows: the checks cover those tasks, and to be sure your own use is among them you run the diet right after using the skill, or name the case when you ask. If the one thing you rely on is the unusual thing, say it out loud before anything is cut.

Those tasks are sampled from your real session logs, and for skills that teach a tool the diet reads the logs to see which calls are common. If client work went through those sessions, your agent will be reading it again. And the README does not say what happens when the original author updates a skill you have slimmed. I would diet a copy and keep the original where the updater can find it.

Verified. November 16, 2026, at commit b93e147. 259 stars and 10 forks; MIT, created October 24, last push November 6. Two skills promised, two SKILL.md files counted, 18 files in all. Read: the README, the main SKILL.md and references/eval.md.

What I did not verify

I did not install or run any of the five. Every behaviour above comes from a README, a results file, a skill file or a template I read.

I did not run rea-video on a real platform download, so its accuracy figures are the author's own evaluation on the author's own videos. I did not listen to music-skill's six example pieces, which means I have no opinion on how they sound, and that is the only opinion that counts there. I did not open Papermorph's bookshelf chapter by chapter or price a chapter myself. I did not visit the three demo sites of the Astro kit or compare them with the originals. I did not put a skill on a diet.

No piece from my own skill library this week: the ones left either already ran here or are still waiting on a rights review.

The Hacker News search did not answer from here and Reddit refused the request, so I cannot say whether any of the five was discussed. Stars move by the hour, and the numbers above carry the morning they were read.

What I still do not know

The music-skill README answers the commercial question as far as its own tools go: no sound bank asks for credit, and nothing in the chain stands in the way of a release. Then it stops at the edge of what a README can know: whether you hold a copyright in music an agent wrote for you depends on your country's law and on how much of it you shaped.

I do not know where that line falls in Brazil for a cue where I chose the tempo, the instruments and the bar where it peaks, and the agent chose the notes. When a contract asks me to assign the rights to a piece of music, I want to be sure I hold them. Until someone who practises this area tells me otherwise, I would treat such a cue as a temp track.

Edition 13 is due next Monday.

Sources and verification

Method: a documentary review of each repository's own files, read on November 16, 2026, with one GitHub snapshot at 10:11 UTC. None of the five skills was installed or run.

About the author

Read the manifesto→ Write in