convez.ai
SkillAutomaçãoAvançado
egorlapygin-ops

Video Hosting Skills — Publicação com Revisão Humana

egorlapygin-ops/video-hosting-skills

Drafta, edita e publica vídeos com varianção de conteúdo e aprovação humana

Ver no GitHub
InicialPluginScore · de 0 a 100

Par de skills Python para Claude que automatiza a criação de variantes de título, descrição e capa para vídeos, mantendo um editor humano no controle via Google Sheets antes da publicação. Integra rascunho automático, revisão intermediária e publicação final em um fluxo estruturado.

Instalação

$npx convez-ai egorlapygin-ops/video-hosting-skills

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Publicação em massa de clipes com variantes de título e descrição geradas por IA
  • Fluxos de revisão editorial com aprovação intermediária antes de publicar
  • Automação de composição de capas de vídeo com validação de especificações de design

Veredito

Solução robusta e bem pensada para automação de vídeo com governança humana. A arquitetura de three-way variants é inteligente — reduz retrabalho ao oferecer escolhas em vez de regeneração total. Requer Python, integração com Google Drive/Sheets e hospedagem customizada, mas o fluxo é produção-ready para equipes editoriais.

README

A tradução para o português deste README ainda está na fila.

video-hosting skills — human-in-the-loop publishing

Two Claude Code skills that turn a folder of edited videos into published clips on the DataPeople video hosting, with a human editor in the middle. Automation prepares the work; a person decides; automation carries out the decision. The stop between the two is the whole point.

videos + designer templates
        │
        ▼
 /video-hosting-draft  ──►  Google Sheet (3 title + 3 description + 3 cover variants per video)
                                   │
                                   ▼
                            EDITOR reviews in the sheet:
                            picks a variant (1/2/3), edits text, sets status = approved
                                   │
                                   ▼
 /video-hosting-publish ──►  hosting API (publishes ONLY approved rows, writes back id + url)

Why three variants? An LLM rarely nails the title on the first try — variant 1 is safe, variant 2 is over-cooked, variant 3 is often the keeper. Showing the editor all three turns "rewrite from scratch" into "pick and tweak". Cheap for the model, a saved half-hour per clip for the editor.

The two skills

.claude/skills/video-hosting-draft

Stage 1. Never publishes.

  • vh_draft.py inspect --src <Drive folder URL | id | local path> — pulls videos + templates + rules into a work dir and prints a JSON manifest (videos with transcripts, templates with pixel sizes, verbatim rules text). Claude reads it.
  • Claude writes 3 title + 3 description variants per video and translates the designers' rules into a normalized cover spec → variants.json (see examples/).
  • vh_draft.py run --src <...> --variants variants.json — composites 3 covers per video, uploads them to Drive (so they render in the sheet), and creates/fills the editor Sheet with status/pick dropdowns. Saves ~/.config/video-hosting/state.json.
  • vh_covers.py — the cover compositor (pure Pillow) + the SPEC title validator. Renders a finished thumbnail from a template + a cover spec, and rejects titles that break the contract (length, emoji, ALL-CAPS, trailing punctuation, pill/brand duplication) so they get regenerated rather than truncated.

.claude/skills/video-hosting-publish

Stage 2. Acts only on approved rows.

  • vh_publish.py [--sheet-id ID] [--dry-run] [--real-video] — for each approved, not yet published row: resolves the chosen title/description/cover, POSTs to the hosting via multipart (curl, no extra deps), writes back published id + gallery url and flips status to published. Errors land in the row's notes; draft/published rows are skipped.

The editor sheet

One row per video. Human columns first, machine/output columns hidden at the end.

statusvideopicktitle 1..3description 1..3cover 1..3final titlefinal descriptionpublished idpublished urlnotes
  • status — draft → editor sets approved → publish sets published (dropdown).
  • pick — which variant to publish, 1/2/3 (dropdown).
  • final title / final description — optional overrides; if blank, the picked variant is used.
  • cover 1..3 — =IMAGE() previews of the three composited covers.

Install

Copy the two skill folders into your Claude Code skills dir:

cp -R .claude/skills/video-hosting-draft .claude/skills/video-hosting-publish ~/.claude/skills/

Requirements: Python 3.9+, Pillow, curl, and the gws Google Workspace CLI authenticated (gws auth login) for Drive + Sheets. Drive pulls also work on any link-shared public folder without auth.

Store your hosting token once:

mkdir -p ~/.config/video-hosting
printf '%s' 'vh_xxx' > ~/.config/video-hosting/token   # from the "Your API token" box at /video-hosting

The hosting API (mock, docs at /video-hosting/docs)

methodpathnotes
POST/video_hosting/api/videosmultipart: video (required, discarded by the mock), thumbnail (PNG/JPEG, required — shown in the gallery), title (≤200 chars), description (optional)
GET/video_hosting/api/videosyour videos, newest first
DELETE/video_hosting/api/videos/<id>204 on success

Auth: Authorization: Bearer <token>. Errors: 400 (missing field / bad thumbnail type), 401 (bad token), 413 (>300 MB).

Cover spec (normalized)

The designers ship a SPEC.md per template category (e.g. case, educational). Claude translates each into a normalized cover spec; the compositor applies it. Per-category keys:

{
  "title": {
    "box_rel": [0.50, 0.0, 0.46, 1.0],
    "font": "HelveticaNeue.ttc", "font_index": 1,
    "color": "#1A1A3A", "align": "left", "valign": "middle",
    "max_lines": 3, "min_size": 36, "max_size": 96, "autofit": true
  }
}

box_rel is [x, y, w, h] in 0..1 of the canvas; autofit walks max_size → min_size (step 2) keeping the largest size that word-wraps within the zone and ≤ max_lines. Optional scrim, shadow, eyebrow blocks exist for templates that need them. See .claude/skills/video-hosting-draft/vh_covers.py for the full shape and examples/ for a real, two-category variants.json.

Notes

  • The draft/publish boundary is deliberate: the code stops, hands the lever to the editor via the sheet, and waits.
  • The video file is required by the API but discarded by the mock — publish sends a tiny placeholder by default (fast, dodges the 300 MB cap); --real-video downloads and sends the genuine file from Drive.
Ver o original no GitHub

Bom para

Relacionados