agentleFS
lohi-ailohi-ai/babysit/.claude/skills/marketing-video/SKILL.md

marketing-video

Skill0 starsChanged 3 days agoMITView on GitHub

Plan and produce marketing videos for your own product on TikTok, YouTube (Shorts and long-form), Instagram and Facebook Reels, X and LinkedIn, for any product and any kind of video, from research to a script strangers can retell, art, voice, edit and per-platform files; you post them.

6 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

11 rules. 0 are common, 8 are distinctive, and 3 are very specific to this project.

  • 0

    Common

    Most of the 6 edited versions we read keep these.

  • 8

    Distinctive

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Code makes the video: Every format, a story as much as a demo, is built as code: the place, the characters, the lettering and the motion …

    line 31
  2. Very specific:

    Run mvf doctor before the first video.

    line 97
  3. Very specific:

    Signing in to an AI provider for Soot: Ask the user to run ! soot login <provider>.

    line 100
  4. Very specific:

    Naming the judge model on first use: Ask once, offer the providers that mvf judge status lists as signed in, then run mvf judge setup …

    line 102
  5. Distinctive:

    Research before writing: Read what the audience says and what the most-viewed videos on the topic, and for the benefit, actually tell their …

    line 152
  6. Distinctive:

    Claims stand on facts: Every quantity, price, "free", superlative or guarantee in a beat or in the copy names a fact id.

    line 158
  7. Distinctive:

    Only approved art is shown.

    line 177
  8. Distinctive:

    One video at a time: Write 3 candidate scripts, then pick one script and one hook and produce only that one.

    line 184
  9. Distinctive:

    Speak to people who have never heard of it: Use their words and their situation first.

    line 191
  10. Distinctive:

    Draw what the audience already pictures: If the product, character or world has a known look (an adaptation, a mascot, brand art), the cast …

    line 202
  11. Distinctive:

    Never log in to, upload to or post on any platform, and never spend money (ads, boosts, paid APIs) without the user's explicit say.

    line 278

“Common” is measured: kept by at least 6 in 10 of the 6 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

marketing-video

A founder or indie dev with no marketing team gets a platform-ready video, made by an agent. This is a manual stopgap until 2found's marketing product (2market) ships. Follow the preamble for bootstrap and telemetry. Shared refs are filesystem paths beside this skill's directory, so read them by path, not as skill://.

It works for any product: an app, a SaaS, a game, a book, a physical product, a service, a course or a creator. It also works for any kind of video: a demo, a tutorial, a story, an explainer, a testimonial, a comparison, a list, a skit or a long-form essay.

This skill sets the standards and the direction; the craft is yours. It says what a video must achieve and what must never be broken, and passes on lessons from past campaigns. It does not prescribe a structure, a shot list, a look or an edit. Invent those for the product, the audience and the platform, and use whatever tool does it best.

You own every step; mvf assists. Its presets, recipes, effects and rig are shortcuts, and each has a raw form beside it: raw HTML or SVG plates and effects, raw GSAP tweens, hand-authored scenes and captions, extra CSS. Use those, or another tool entirely, whenever the shortcut is not what the video needs. A check FAILs only on a standard, or on something that would break the render, the platform or the timing. Taste is WARN advice that you answer in writing.

Code makes the video. Every format, a story as much as a demo, is built as code: the place, the characters, the lettering and the motion are HTML, SVG, CSS, canvas and GSAP, beside the real product's captures. Image generation only supports a coded scene. It can give a texture, a far layer behind coded parts, a portrait or a prop that the code moves, a thumbnail, or a reference. It never gives the scene itself. A video of generated pictures panned and zoomed under text is a slideshow, and slideshows have been the weakest cuts this skill has made, xianxia above all. A full-frame plate is the rare exception, and you answer it in writing (A12).

There is no default look. Design the picture for this video: the place, the characters, the lettering and the motion are yours, written in code. A new campaign names no style preset ("style": "none" in campaign.json). The presets (stickman, ink, anime), the code-drawn rig and direction.json are still there for a video that wants one, and nothing asks for them (07 Production).

One campaign is one folder, marketing/campaigns/<slug>/ by default. Each stage leaves one artifact on disk, so the work survives a crash and can be audited.

How it runs

The owner touches a campaign three times, and the agent does everything else.

  1. The brief. What to promote, the goal, and where it goes (platforms, language); the viewer benefit only if the owner names one. Ask once, only for what the request leaves out, and write every other assumption into campaign.json.
  2. The script. The owner reads the winning script's page (mvf sheet <sid>) and says yes, no with a reason, or "you decide" (mvf approve). This happens on the campaign's first video; nothing is voiced before it.
  3. The cut. The owner watches the delivered files and posts them.

Contexts

A campaign runs for hours, and a context that carries every download and every scene's code loses the brief. Keep the main context for reading the material, writing and deciding, and give the rest to fresh agents on the strongest model available (Opus on Claude Code), each with a brief on disk and a result on disk:

  • Gathering (stages 1-3.5): downloads, frame sheets, transcripts and source pulls go into 00-material/ and the source lists. You then read the pack yourself and pick the ideas: a script written from a helper's summary has lost the details it needed.
  • Each scene, after the key frame is yours: one agent per scene, given SCENE_GUIDE.md, the key frame, its beat and the storyboard. It returns snapshots; you look at them. Renders run one at a time.
  • Every review: the predict, overlap, coldread and viewer helpers.
  • Production as a whole: once the owner approves the script, stage 7 starts in a fresh context that reads the disk (mvf next).

These roles are part of their stage, so a caller's "no subagents" default does not cover them. With no agent tool, work in one context and re-read campaign.json and the script before each stage.

Between those points the agent loops on mvf next, which names the one next action and says whether the agent or the owner takes it. The agent does the action, runs its check, and asks again. mvf status shows the whole board. A FAIL blocks; a WARN is advice to follow or answer in writing. An AGENT: line is a step mvf cannot do, such as a page that needs a browser or a video it cannot read. Do it yourself before that stage's check.

Backbones

JobEngineFirst use
Edit and renderHyperFrames (HTML to video, npx hyperframes, Node 22+)the first scaffold downloads the pinned version
VoiceOmniVoice (clone or designed voice; Whisper times the words)uv builds its environment on the first take (about 2 GB, once)
Picture parts (supporting code, never the scene)Qwen Image 2.1 Turbo, and RMBG-1.4 cutouts (the image-gen skill)image-gen/scripts/setup downloads the weights once; free and local
JudgeMicro Fish on the user's Soot: a society of persona Soots on a shared feed, and an expert panelthe user signs in to a provider and names the model, then mvf judge setup

Run mvf doctor before the first video. It checks each backbone and prints the fix for anything missing. The agent fixes what it can; NEEDS lines are the user's to do:

  • Signing in to an AI provider for Soot. Ask the user to run ! soot login <provider>. Never ask for, paste or write a key.
  • Naming the judge model on first use. Ask once, offer the providers that mvf judge status lists as signed in, then run mvf judge setup --provider P --model M.

Other engines plug in as extensions, but these four are the defaults the checks expect.

Stages

#Stage -> artifactThe bar it clearsReference
0brief -> campaign.json, material -> 00-material/goal and KPI, the viewer benefit only when the owner names one, brand, platforms, constraints; what people already like or know about the product00
1product -> 01-product/facts.jsonevery claim the video may make, sourced and dated; real screens01
2audience -> 02-audience/segments.jsonwho it is for, by need, from evidence, in their words02
3trends -> 03-trends/trends.jsonwhat the platform rewards this fortnight, observed03
3.5research -> 03-research/research.jsonwhat the audience asks, what the most-viewed videos say, the specific details; one idea picked per candidate script, each built on something from inside the subject and ranked by what the audience already watched03b
4strategy -> 04-strategy/angles.json2-3 angles, each a benefit for a segment, standing on facts, checked fresh against the obvious videos and this owner's past ones04
5script -> 05-scripts/<sid>.json3 candidate scripts, each delivering its benefit early and in plain words, complete enough that a stranger can retell it, written from its ending, with 2-4 key moments, hook variants and platform copy; the kind of video and its channels from the atlas05
6pre-test -> 06-pretest/<round>/decision.jsonMicro Fish (or another honest test) picks one script and one hook, and strangers can retell what it told them; the owner reads the winner before production06
7video -> 07-video/<sid>/one video: real product, the picture drawn in code (approved art only as its parts), measured audio, reviewed by looking, by a cold viewer who never read the plan, and judged by a simulated panel07, craft
8delivery -> 08-delivery/<sid>/<canvas>/<platform>/the platform's specs, QA, captions, copy, AI labels08

Open only the reference for the stage you are in, and only the sections you need: rg -n '^## ' references/<file>.md lists them, and the Bar comes first. Stages 1-4 are research, so a guess is written down as inference, never as a fact. Go back a stage whenever the evidence says so. Run only the stages the user needs, and say where you stopped.

Standards (FAIL in the checks; breaking them costs people)

  • Something from inside worth retelling, and a use. A stranger stays for something from inside the subject: a scene, a twist, a line, a rule of its world, a fact they did not know. They save and send the video for what it does for them: something to apply, decide, understand or do. A video needs both. A story with nothing to keep fails the second half. A guide about the product with nothing from inside it fails the first. Facts about the product's packaging (its length, price, size, where to get it) are neither, so give them one line at most. Each angle picks its benefit; the campaign sets one only when the owner asks. Every script says in plain words what the viewer can use (spine.use). For an answer the use is the answer, in the topic's own words; only a lesson, a decision, a match or a how-to is turned into advice about the viewer's life.
  • Understood in one hearing, and complete. A stranger who hears the video once, at speaking pace, can retell what it told them in a sentence or two, and that retell is the answer the hook promised, not the question again, a slogan or "it depends". Every term the point needs gets its plain meaning the first time it is heard; a term it does not need is cut, never kept unexplained. A conditional answer names the condition and gives each result. A short video that leaves the viewer asking "so what is it?" has given nothing, however well it opens. Length follows what the viewer needs to follow: cut repetition and asides, never a step of the explanation.
  • Research before writing. Read what the audience says and what the most-viewed videos on the topic, and for the benefit, actually tell their viewers, not only how they open. Ideas come from that evidence, not from memory, and are ranked by it: the views of the top videos that carry the same kind of payload, never the writer's own scores. Never remake what an earlier campaign said; mvf lint FAILs a script whose lines repeat one.
  • Claims stand on facts. Every quantity, price, "free", superlative or guarantee in a beat or in the copy names a fact id. Re-read price and limits before every export (re-check any older than 30 days).
  • Real product on screen. UI is a captured screenshot or recording, never a drawn mock-up. A missing screen is a labelled placeholder and a report line.
  • The first seconds earn the rest. Most viewers decide within 3 s. The spoken hook ends near the platform's limit (2.5 s on TikTok and Shorts, 2.0 s on X; more than 1 s over fails), X needs on-screen text because it autoplays muted, and frame 0 shows something. The first real value lands early (about 6 s for a short, 15 s for long-form; more than 2 s late fails), and the payoff keeps the hook's promise.
  • Measured, not hand-typed. Timing, loudness and specs come from the audio and the rendered file. Estimates are labelled.
  • Rights. Show only own, licensed, public-domain or official-promo material (brief, credited, with your own commentary). unknown may inform the script but never appears on screen. Never bake a licensed track into the file; trend sounds are added in the app.
  • No AI likeness of a real person, and never prompt an image model with a real person's name. Generated art is never the product UI, a logo or words. Only approved art is shown. Realistic AI picture or voice gets the platform's AI label.
  • Honest testing. A pre-test records its true method: Micro Fish is a simulated panel, never "users", and agent readers are labelled as such. The winner beats a control written to lose, and the panel's retell, quoted, matches the script's takeaway. The owner reads the first script of a campaign before any production.
  • One video at a time. Write 3 candidate scripts, then pick one script and one hook and produce only that one. The machine does not have the resources for more.

Direction (lessons, not rules)

  • Start from what people already like or know about the product, not from what the product wants to say.
  • Speak to people who have never heard of it. Use their words and their situation first; insider names and jargon come later, for the fans. Making it reach means explaining a word, not avoiding it: a script that dodges the terms but still leans on them ("the lower five", "level three") loses strangers and fans at once.
  • Short is not thin. The checks measure the hook, the timing and the jargon; none of them can tell that a script said almost nothing. A script revised round after round to pass them shrank to a hook, two quotes and "both sides are half right", and the owner could not tell what it was about. Before any pre-test, write the retell a stranger should give, and check that the body actually says it.
  • Draw what the audience already pictures. If the product, character or world has a known look (an adaptation, a mascot, brand art), the cast drawn in code takes its signature traits, and so does any generated part.
  • Whoever made the cut cannot score it. Three cuts their maker scored 7 of 10 had every character under a third of the frame height and more than 60% of the frame empty; the most-viewed videos on the same topics fill theirs. The size on screen is measured (D16, PX6), and the score is the cold viewer's, given beside frames of those videos (07-review).
  • Design one frame before building them all. The same three beats were built from direction.json and by hand: the generated scenes were a small rig on a flat ground, the hand-written one a lit place with figures filling the frame. Finish one scene by hand, compare a frame of it with the most-viewed references (mvf keyframe), and hold every scene to it (07-production).
  • Agent readers judge entertainment, not benefit. Two scripts that every agent panel passed were rejected by the owner because they gave the viewer nothing to use. Ask readers what they would use, save or send, and make them retell what the video told them: a reader who read the script as text can echo its lines without having understood them. They are also polite about boredom, so ask what they would tell a friend (tell_a_friend). If most readers say "nothing", the script bores, however useful it is.
  • Brand in the picture, not in the pitch. Show the product where the value happens. Keep its name out of the hook, speak it once at most, and save selling words for one soft CTA: most viewers never reach the end, and a pitch reads as an ad. When the owner wants the product in the first seconds, let it be seen and heard working there (the real player playing, its own sound in a pause after the hook) while the hook still speaks about the subject.
  • The picture is the whole frame. Safe zones say where text and faces may sit; they do not say where the picture ends. A cut built to the zones as if they were the canvas put every subject in the top half of the phone and scored 6, a template.
  • Every angle names who a viewer would send it to, and why.
  • Each reference gives its stage's bar first, then the lessons. Treat the lessons as evidence to weigh. Where the advice and your judgement differ, follow your judgement and note why.

Tool

mvf is a stdlib Python CLI. Run it as PYTHONDONTWRITEBYTECODE=1 python3 <this-skill-dir>/tools/mvf.py <cmd> from inside a campaign.

  • It needs python3 and ffmpeg; yt-dlp for video references and transcripts; mvf doctor checks the rest.
  • mvf init <slug> --product NAME [--root DIR] writes the folder and its .gitignore; pass --root when the repo keeps campaigns elsewhere. For the next episode of a series, mvf format save <name> --from <sid> keeps the last one's look, captions, ambience, rigs, props and voice, and mvf init <slug> --product NAME --format <name> starts from them. mvf next says what to do, mvf status shows the board, and mvf --help lists every command.
  • Pass --campaign <dir> (before the command) whenever more than one campaign is open or the shell's directory may move, such as parallel agents sharing a shell; otherwise mvf writes into the campaign that contains the current directory.
  • The checks (mvf check <stage>, mvf lint, mvf qa) report two levels: FAIL is a broken standard and blocks the next stage; WARN is advice, so follow it or write down why not (a script's warn_notes). The checks count and match words; they cannot tell whether a script is clear. Never bend a script to quiet a WARN that is wrong for it: answer it.
  • The voice, art, scaffold, direction, sound and export helpers are one way to build the video, not the only one. Any route that meets the standards is fine.
  • Local engines, data sources and recipes are gitignored extensions (contract).
  • Fresh helpers (mvf predict, mvf overlap, mvf coldread, mvf viewer) get a brief with only what their role may see, a budget of 15 tool calls, and a short reply; the details go to a file.

The cold viewer and its fixes, the prediction and overlap tests, the writing tells, the motion numbers, the cast sheet and the generated-design tells adapt ideas from cinecode (MIT); no code was copied. Hit-stop, snapped and stepped poses, after-images and the blow lessons in craft adapt ideas from pixelfx (MIT); no code was copied.

Boundary

Never log in to, upload to or post on any platform, and never spend money (ads, boosts, paid APIs) without the user's explicit say. Delivery ends at local files plus caption and post text; the user posts. Keep secrets out of campaign files, and read them from env vars only. Observe platforms logged out (pages that need a real browser, such as TikTok, through browse), and do not like, follow or comment.

Report

STATUS: DONE | DONE_WITH_CONCERNS | NEEDS_CONTEXT | BLOCKED
CAMPAIGN: <folder> | STAGES: <run / skipped and why>
FILES: <delivered paths with duration, size, loudness, QA verdict>
OPEN: <unverified trends, placeholder screens, untested hooks, what the user must watch and post>

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.