agentleFS
pooriaarabpooriaarab/skills/ad-lofi-ugc/SKILL.md

ad-lofi-ugc

Skill1 starsChanged 9 days agoMITView on GitHub

Make cheap lo-fi UGC-style ad creatives and short video ads that read as real phone photos, not ads. Covers the camera-and-era-first image prompt rule, the exact ffmpeg degrade recipes that turn clean AI stills into 2000s camera-roll photos and video, one consistent synthetic character across clips (character bible + last-frame chaining), muted-first sound design (voiceover, ambient bed, ffmpeg SFX), and the QA gates that catch the real failures (objects that change identity, rotated images, garbled text, policy violations, runaway spend). Use when asked for UGC ads, lo-fi ads, fake-phone-photo creatives, TikTok/Reels-style video ads, cheap ad batches on a small budget, or 'make it look like a real person filmed it'. Pair with ad-creative-generation for clean photoreal stills, ad-image-prompt-library for text-exact HTML formats, and ad-video-styles for polished generated-footage styles.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

4 rules. About 0 are common, 1 is distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    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. Very specific:

    Write the character bible: if a person recurs. One paragraph, pasted verbatim into every prompt.

    line 36
  2. Distinctive:

    Strip PII before anything ships: No real names, account ids, keys, wallet balances or private domains in prompts, logs or this repo.

    line 72
  3. Very specific:

    Never send a fake User-Agent or Referer to get past a block.

    line 88
  4. Very specific:

    When the MCP route is not set up, use other sources: the open-source prompt and hook repos in references/providers.md, and the ad libraries …

    line 91

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “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

ad-lofi-ugc

A repeatable workflow for lo-fi UGC ad creative: stills and short vertical video that look like a real person shot them on an old phone. Built and measured on a real campaign run in November 2026. Total creative spend for a 12-image + 5-clip batch was under $1.00.

The look is the product. A clean, glossy AI image reads as an ad and gets scrolled past. A grainy 2004 point-and-shoot photo with a crooked horizon reads as a camera roll and earns the second look. Everything here serves that.

Pairs with:

  • ad-creative-generation — clean photoreal stills, reference-image rules, vision QA on un-degraded output.
  • ad-creative-templates / ad-image-prompt-library — text, captions, UI and end cards. Never let the image or video model paint text (see references/prompts-images.md).
  • ad-video-styles — a different family: polished generated footage with static overlay type. Use this skill when the goal is "looks amateur on purpose".
  • ad-video-editing — the fast-cut post stage: dead-air removal, pacing, SFX on cuts, captions and render QA for the finished video.
  • ad-experiments — how to size budget and read results once ads ship.

The pipeline, in order

  1. Find the format. Mine a free viral-format library for structures that already work. A person can browse darkugc.com by hand; see the format research section below for the rules.
  2. Write the character bible if a person recurs. One paragraph, pasted verbatim into every prompt. See references/characters.md.
  3. Generate stills with the camera-and-era-first prompt rule. See references/prompts-images.md.
  4. QA the stills: orientation, hands, objects, no text. See references/qa.md.
  5. Degrade each still with the ffmpeg recipe. This is what makes it lo-fi. Same file.
  6. Animate with image-to-video motion prompts kept deliberately small. See references/prompts-motion.md.
  7. Chain shots by feeding each clip's last frame into the next shot. See references/characters.md.
  8. Add sound last, and design muted-first. The video must work with sound off. See references/sound.md.
  9. Run the QA gates: object continuity, mute test, policy lint, spend meter. See references/qa.md.
  10. Log every prompt and every billed call. The prompt file is how the next run repeats this one.

Non-negotiable rules

  • Camera and era first, then subject and moment, then flaws, then negatives. That order is what made the prompts work. Full rule and the exact prompts in references/prompts-images.md.
  • The model never paints text. Captions, headlines, lower-thirds and end cards are HTML renders or ffmpeg overlays added after generation. Blank signs, taped notices and "no readable text" prompts still produced garbage letters twice; the fix is to keep text out of the frame entirely.
  • Keep motion small. Tiny shake, slow push-in, one continuous shot. Every time we asked for more (walking, cuts, zooms) the model invented people, phones and morphs.
  • Muted-first. Captions carry the story; sound is a bonus layer. Export a _muted.mp4 of every video as proof.
  • One writer per spend log. Two workers racing one spend.json clobbered each other's entries. Lock the file, price-check before every call, hard-cap the batch. Details in references/qa.md and references/failures.md.
  • Strip PII before anything ships. No real names, account ids, keys, wallet balances or private domains in prompts, logs or this repo.

Format research: viral-format libraries

darkugc.com is a free public library of viral short-form ad videos with view counts. Its terms ban scraping, bulk download and any automated access to the site or its API, including bots, scripts, headless browsers and "AI agents". Follow them.

  • Human use: a person opens https://darkugc.com/, sorts by Views, and filters by category and format family. Brand pages exist per advertiser. The person notes the story structure by hand.
  • Agent use: only through the route the site names for agents, its MCP server at https://darkugc.com/mcp, with a free account the owner signs in to. https://darkugc.com/llms.txt states the AI policy. Read it first.
  • Never send a fake User-Agent or Referer to get past a block. A block is a no. Do not call /api/ paths directly, and do not save bulk copies of its data. The site's robots.txt disallows /api/.
  • When the MCP route is not set up, use other sources: the open-source prompt and hook repos in references/providers.md, and the ad libraries run by the platforms (for example the Meta Ad Library API).

Copy the structure, never the footage, the face, the words or the music. A library is inspiration, not source material.

Cost shape (measured, 2026-11-12/13)

StepModelPriceVerified
Stillsbytedance/seedream-v5.5 @1.5k on WaveSpeed$0.02/imageyes
Image-to-videowavespeed-ai/wan-2.2/i2v-480p-ultra-fast$0.01/s ($0.05 per 5 s clip)yes
TTS voiceovercheapest decent WaveSpeed TTSper-call, check firstpartially

Full provider table with alternates and unverified marks: references/providers.md. Every price must be re-checked on the day; prices and model ids change.

References

  • references/prompts-images.md — prompt structure rule, exact working prompts, ffmpeg degrade recipes for stills and video.
  • references/prompts-motion.md — image-to-video motion prompts that worked, and what to keep small.
  • references/characters.md — character bible method, anchor contact sheet, last-frame chaining.
  • references/sound.md — voiceover, ambient bed, ffmpeg SFX, muted-first design.
  • references/owner-feedback.md — owner rules R1 to R26 and the cost ratchet. The reviewer checks each rule.
  • references/qa.md — continuity check, mute test, orientation check, policy lint, spend meter.
  • references/providers.md — cheapest working models and prices, dated and marked verified or unverified.
  • references/failures.md — every failure hit and the fix.

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.