references
Skill67 starsChanged 17 days agoMITView on GitHub
The reference engine of the AI Video Factory. Use whenever the creator shares videos or links they like as references ("aquí te dejo estos videos de UGC", "mira este video", "analiza este TikTok/Reel/YouTube", "quiero algo como esto", a pasted video link or an attached .mp4, "reference", "referencia", "breakdown", "desglose", "reverse engineer this video"), or asks what is in the reference bank. Measures every video locally, writes a frame-by-frame breakdown (hook, script, structure, every shot and join, voice, music, effects, on-screen text, look), asks only what it can't tell, publishes the breakdown page and adds it to the bank in the creator's words. The creator never types a command.
12 other projects use a version of this.
Adopted by
Most recently taken up by
- uwonu606/claude-skillstheir own version, first seen Nov 16, 2026
- nerymurillohnd/the-claude-factorytheir own version, first seen Nov 16, 2026
- swarm-routing/skillstheir own version, first seen Nov 15, 2026
- jmche/skills-hubtheir own version, first seen Oct 20, 2026
- arthur-plautz/research-toolkittheir own version, first seen Oct 15, 2026
- darellchua2/civiltekk-opencode-claude-skillstheir own version, first seen Sep 18, 2026
Other versions12 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 12
- alfredxw/denova★ 1ksame name
- openbkn-ai/bkn-sdk★ 27same name
- gohypergiant/agent-skills★ 25same name
- howells/fiction★ 23same name
- jeffsenso/dolibarr-skills★ 9same name
- v-open-platform/ai-app-skills★ 8same name
- darellchua2/civiltekk-opencode-claude-skills★ 7same name
- nerymurillohnd/the-claude-factory★ 1same name
- swarm-routing/skills★ 1same name
- arthur-plautz/research-toolkit★ 0same name
2 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository9 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- castSkill
- cinematographerSkill
- format-longformSkill
- format-seriesSkill
- onboardingSkill
- story-architectSkill
- studioSkill
The file
We split this skill into the rules it gives an agent
1 rule. 0 are common, 0 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 9 edited versions we read keep these.
0
Distinctive
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 18
Read overview.jpg, every shot-NN.jpg and every join-NN.jpg with the Read tool.
“Common” is measured: kept by at least 6 in 10 of the 9 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
References: the creator gives videos, Claude does the rest
The creator talks; Claude runs engine/refs.mjs. Never ask the creator to type a command or remember an id. Everything in this skill is enforced in code (engine/lib/refs.mjs, scripts/check-refs.mjs); the steps below are how to drive it.
1. Take every video in
- A file (attached or a path):
node engine/refs.mjs add "<path>" --note "<short name you choose>" [--lang es]. - A link (TikTok, Instagram, YouTube, X, any page yt-dlp reads): the same command with the link. Only links the creator gave; say what was downloaded (title, platform, length). If a link fails, say why in one line and ask for the file instead.
- Several at once: run
addfor each. The note is a short readable name ("ugc-skincare-mirror"), never the creator's typing.
add measures on its own: transcript word by word (never silence detection), cuts and every shot's length and motion, the hook (first 3 s), pauses, the bed under the voice, the mix, the music's tempo, sound hits outside the voice, on-screen text (local OCR), brightness and color, and a format guess with its confidence. It cuts refs/<id>/sheets/: overview.jpg, shot-NN.jpg (4 fps) and join-NN.jpg (the half second on each side of every cut).
2. Look at everything, then write the breakdown
Read overview.jpg, every shot-NN.jpg and every join-NN.jpg with the Read tool. Use refs.mjs show <id> for the numbers. Then write breakdown.json (contract: Breakdown in engine/lib/refs.mjs):
formatand why;summary;hook: type, seconds, the words, why it works;structure: beats with times;shots: one entry per measured shot (what, framing, camera, action) and for every shot but the lasttransitionOut{type, how it joins the next};voice(who, delivery),music(present, style, role),sfx(times and what),onScreenText(style, examples),look;takeaways(what we learn, at least 3) andneverCopy(characters, lines, logos, signature compositions).
Save it with node engine/refs.mjs breakdown <id> --file <json>. breakdownRules refuses a missing shot or join, beats outside the video, music that contradicts the measurement, measured on-screen text with no style described, or measured sound hits with no effects listed. Fix and save again; never weaken the breakdown to pass.
3. Ask only what you can't tell
When the measured guess is under 70 % or disagrees with what you saw (UGC vs influencer is the usual one), ask one question with buttons (AskUserQuestion), for example "¿La guardo como referencia de UGC o de influencer?". Also ask what they liked if it isn't obvious (hook, rhythm, look, all of it) and put it in the takeaways.
4. Show the page, approve in their words
node engine/refs.mjs page <id>, then publish refs/<id>/page.html with the Artifact tool, sending the sheets as files (root: refs/<id>, files: ["sheets/overview.jpg", "sheets/shot-01.jpg", ...]). Give the link. On their approval: node engine/refs.mjs approve <id> --said "<their exact words>" (and --format-said "<their words>" when the format was their call). Never approve with words they didn't say, never as a test.
5. In production
studio.mjs new lists the approved references of the format. Offer them with buttons; record the choice with node engine/studio.mjs refs <slug> --use id,id --said "..." or --none --said "...". The concept gate waits for it; a launch plan is calibrated to the chosen references; the script gate refuses lines that repeat a reference's phrases (copyOverlap over 20 %). Use the bank's ranges (refs.mjs bank --format <f>) and takeaways when writing; never copy a reference's characters, lines, logos or signature shots.
What stays private
Reference media (refs/**/source.*, refs/**/sheets/) is ignored by git; profiles, breakdowns and pages are kept. The public edition ships the engine, never the bank.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

