Repository navigation
Releases
v0.5.0 — forty templates, an industry layer, and four defects a green build could not see
Eighteen pull requests since v0.4.0. No breaking change — nothing that validated before stops validating, no slot renamed, no environment variable newly required.
What grew
| Forty templates, grouped by what they do | Eighteen became forty: four 3D frames, four data frames, maps, and a catalogue image ordered by the job a frame does rather than the batch it arrived in. |
| An industry layer | INDUSTRIES.template.json — 16 industries, each carrying both halves: what to write and what counts as evidence there, then the genre, frames and palette. The three regulated ones (healthcare, finance, real estate) carry a legal block with links to the source, and a test requires the link. |
| GEO, both of them | The local genre, where the story is where, and geo-audit.mjs, which grades whether a post still means anything once an answer engine quotes one paragraph of it. Kept in one place so nobody does the wrong one. |
| Captions and transitions, made here | --captions burn, and fade · swipe · slide · iris · pixelize with the arithmetic that keeps the video's length exact. |
| A palette read off a live page | theme-from-url.mjs takes background, ink and accent from a real site, under the contrast rule the validator already enforces. |
| A robot on the upstream registry | registry-watch.mjs runs daily, diffs a committed snapshot, opens exactly one PR when upstream moves — it reports, it does not add. |
What this release kept finding
The less flattering half, worth stating plainly because it is the same fault four times. Every one of these shipped green:
- three templates that rendered as a blank box
- four tiles in the catalogue image emptied by the very code written to keep them full
- two frames that turn a missing clip into a black rectangle
- an English README five features and 22 templates behind the Vietnamese one
A passing test suite says the code did what the tests ask. It says nothing about whether anyone looked. Rendering a picture and opening it found what 227 tests did not, four separate times. Most of the guards added this cycle exist to turn "someone should have looked" into "CI looked" — including one that reads every composition and demands a visible stand-in behind anything drawing supplied media, and one that compares the tool names the two READMEs mention, because several older guards concatenated both files and so accepted "documented in at least one language".
Full detail in CHANGELOG.md.
227 tests, up from 121 at v0.4.0. Still no package.json, no node_modules.
v0.4.0 — the gate catches the model, and one piece becomes several
Two threads: a pre-release audit, and what came out of reading facebook-skills.
| The gate catches the model, not just Markdown | oaicite, "As of my last update", an unfilled [Your Name]. Publishing one of those is worse than a stray ## — the first looks careless, the second announces that a machine wrote it and nobody read it back. |
repurpose |
One published item → several more, each from a different angle the source already contains. For one person that's 3 posts a week versus 12. It refuses when the source is thin. |
ads-report without ad spend |
It used to stop when no ads server was connected, leaving everyone who hasn't started buying reach with nothing — despite having the same question. |
| Attribution that stands alone | Two notices named the wrong owner by implication; the root NOTICE covered 14 of 20 folders, missing two vendored Apache-2.0 ones. |
| Repo features as code | Wiki off — recorded in repo-about.json rather than clicked once and forgotten. |
What was deliberately not taken
facebook-skills calls the em dash "the biggest AI tell of 2026". This repo's own reference article uses thirteen of them in ordinary Vietnamese prose and reads well.
Adopting the rule would have failed good writing, and a gate that cries wolf is a gate people switch off. Same for the English vocabulary swaps (leverage → use). Both are one command away in the registry for anyone writing English:
node scripts/install-skills.mjs fb-humanizerAlso
SECURITY.mdnow covers three surfaces the repo had started shipping without documenting: MCP connections that write to live ad accounts, a video backend that bills per second, and a skill installer that writes third-party code into the folder your agent reads and obeys.- Six attribution tests, including one that keeps
LICENSEplain MIT — appending prose is what makes GitHub reportNOASSERTIONand drop the badge.
121 tests, up from 108. Full detail in CHANGELOG.md.
v0.3.0 — plain-text posts, three video backends, 18 templates, 108 tests
Thirteen changes across the output the kit produces, the tooling around it, and the checks that keep both honest.
Most of these were found by running the thing, not by reading it. The kit went from no tests to 108. Six of the fixes were invisible to a passing build — a render that succeeded, a validator that passed, a documented setup that could not work, a meta-skill that had stopped mentioning half the repo.
| Posts are plain text | Meta: / Slug: blocks and Markdown reached readers verbatim — including in the repo's own reference sample. make-post.mjs now refuses to send it. |
| Skills, fetched not vendored | An SEO auditor, three design skills, six marketing skills — on demand, each with its licence and a NOTICE recording the commit. |
| Ads feed back into content | Five Pipeboard MCP servers plus an ads-report skill that produces queue items rather than a dashboard. |
| Campaign visuals via Canva | Hands back a file, not a design link, because everything downstream takes a path. |
| Three video backends | html (free, default, unchanged) · api (Veo/Imagen — bills per second, three gates before spending) · remotion (scaffolds, does not install). |
| 18 templates, 5 genres | Four new frames that draw from their data, and sequences answering "which frames, in what order". |
Fixes worth knowing about
- Two templates rendered landscape inside a 9:16 video — correct CSS, correct viewport, stale
data-*. The render succeeded and the validator passed; only the contact sheet caught it. - A theme never reached
scene.inputs, so a caller's#f59e0bstayed amber on a paper-blue video. The paper-blue sample had been hand-editing that exact hex. .env.examplenamedVBEE_API_KEY; the code readsVBEE_TOKEN. Anyone following the documented setup for the Vietnamese TTS provider could not make it work.- The
.gitignoreCI guard could never have failed — it usedgit ls-files --others, which sees nothing on a clean checkout. - Five templates defaulted to the upstream author's brand, which is what renders when a caller leaves a slot empty.
Full detail in CHANGELOG.md.
v0.2.0 — theme engine + white/ocean-blue sample video
Almost every template in this kit ships a dark canvas. Wanting a light one used to mean
forking fourteen templates and maintaining fourteen more. Now it is one line:
{ "aspect": "9:16", "theme": "paper-blue" }Templates are recoloured into a throwaway copy at render time — nothing under
video-templates/ changes.
Assets
| file | |
|---|---|
claude-fable-5-paper-blue-1080x1920.mp4 |
2 min 28 s · 16 scenes · all 14 templates · white canvas, ocean-blue ink |
claude-fable-5-paper-blue-voice.mp3 |
the narration alone, for CapCut |
Everything needed to regenerate it byte-for-byte is in the repo:
examples/ai-video-social/sample-output-paper-blue/.
What the theme engine handles that find-and-replace does not
- Hex inside JS and
data-composition-variables— that is where a template keeps its
caller-facing colours. Skipping them left a themed comparison card with one orange side. mix-blend-mode: screenflips tomultiplyon a canvas going light. Screen over white
paints white, so an aurora blob or a glitch layer vanishes with a successful exit code.- Accents darkened to clear 3:1 against the canvas. Flipping lightness alone leaves amber
#ffb020at 1.5:1 on white. - Hue squeezed with its ordering preserved, so two accents that differed still differ.
Which way to flip is measured, not guessed: theme-probe.mjs screenshots each composition
with Chrome and measures it. Static CSS parsing is not reliable — frame-bold-poster is
light at 16:9 and dark at 9:16.
Template fixes
Two templates had line-height below 1 — fine for Latin capitals, wrong for a language whose
vowels carry a horn and a tone mark above and a descender below. Both were caught by looking
at a contact sheet; both rendered successfully and exited 0.
frame-bold-poster— tilted headline lines overlapped: "ngày" landed on "rồi tắt"frame-build-minimal— the tilde was clipped off "Ỡ", and the per-character hero split
"NGƯỠNG" into "NGƯỠ / G"
v0.1.0 — video pipeline, 6 voice providers, stock media
First tagged release. The kit now goes from a topic to a published 9:16 video without
leaving the repo.
Sample output
claude-fable-5-sample-1080x1920.mp4 — 2 min 12 s, 1080×1920, real Vbee narration,
stock B-roll from Pexels and a live screenshot of the source page. 15 scenes.
It is attached here rather than committed because it is 16 MB. Everything needed to
regenerate it byte-for-byte is in the repo at
examples/ai-video-social/sample-output/:
script.json, media-lock.json (which pins the exact stock clips) and script.txt.
node scripts/video/render.mjs examples/ai-video-social/sample-output/script.jsonclaude-fable-5-sample-voice.mp3 — the narration track on its own, for CapCut.
There is a matching written article from the same source in
examples/ai-news-social/sample-output/
— one source, two formats, which is the point of the kit.
What's in this release
- Video pipeline ported to dependency-free
.mjs: validate → TTS → SFX → resolve media → render → mux - Six voice providers (
omnivoice,elevenlabs,vbee,fptai,viettel, plus a generichttpadapter). Only the local one needs a server - Stock B-roll from Pexels and Pixabay, web screenshots via headless Chrome, uploads to R2 / Cloudinary
- 14 video templates, plus a fetcher for the 146 in the upstream HyperFrames registry
contact-sheet.mjs— one labelled frame per scene, because four real defects shipped that no validator rule could catch- Multi-platform posting through one Make.com webhook
Known limits, stated plainly
- R2, Cloudinary and Viettel AI adapters are
UNVERIFIED— written from official docs, never run against real credentials. They say so themselves inhost-check.mjs --hostsandtts-check.mjs --providers. The R2 SigV4 signer is verified against AWS's published test vector. - Rendering needs a real machine — FFmpeg, Chrome, and a voice. GitHub Actions cannot do it.
- CI does not install crawl4ai; it only compiles
crawl.py.