instagram-video-workflow

Download Instagram videos + transcribe with faster-whisper + extract content automatically.

SKILL.md

GitHub

Instagram Video Workflow

Complete pipeline: download Instagram video → transcribe audio → extract useful content.

Tools

  • parth-dl: CLI Instagram downloader (installed via pipx) — downloads Reels, posts, carrousels, profile pics
  • ig-download MCP server: Node.js MCP server at /home/keila/ig-download-mcp-server/ with a download tool (url + path params)
  • faster-whisper: Python library for transcription (installed via pipx)
  • transcribe.py: /home/keila/ig-download-mcp-server/transcribe.py — CLI script that extracts audio with ffmpeg + transcribes with faster-whisper

Workflow Steps

1. Download video

Option A — via MCP tool (uses download tool from ig-download MCP server):

  • Tool name: download
  • Params: url (Instagram URL), path (where to save)
  • ⚠️ This MCP server may fail (Instagram download error); fall back to parth-dl if so.

Option B — via parth-dl CLI:

parth-dl "https://www.instagram.com/reel/ABC123/" -o /path/to/output.mp4
  • ⚠️ -o requires a full file path (e.g. /path/to/reel.mp4), NOT just a directory.
  • 🔍 Dica: Mesmo quando o download falha (ex: caminho errado), o parth-dl imprime o título e uploader do Reel no erro. Capture essas infos antes de corrigir o caminho — elas já dão contexto sobre o conteúdo.

2. Check for audio track

Before transcribing, verify the video has an audio stream:

ffprobe -v quiet -show_entries stream=codec_type /path/to/video.mp4 | grep -q audio && echo "✅ Has audio" || echo "❌ No audio (silent video)"

If only video appears (no audio), the Reel is silent — skip transcription entirely and go straight to caption scraping.

3. Transcribe audio (skip if silent)

python3 /home/keila/ig-download-mcp-server/transcribe.py /path/to/video.mp4 --model base --output transcript.json

Models: tiny (fastest) < base < small < medium < large-v3 (most accurate)

If HF_UNUTHENTICATED warning appears, it's harmless — transcription still works.

4. Extract content

Option A — from transcript (if audio was present): use the LLM to analyze the transcript and extract:

  • Key topics / summary
  • Action items
  • Useful information
  • Quotes

Option B — from Instagram post caption (if video is silent, OR as supplementary source): Use browser_navigate to open the Instagram URL and scrape the caption text from the page.

  • 📝 A caption pode estar visível mesmo com o login dialog aberto. O browser_snapshot retorna a árvore de acessibilidade completa — procure por StaticText dentro de dialog ou generic containers. A legenda geralmente aparece como múltiplos StaticText nodes com o texto completo do post.
  • Se o snapshot mostrar um modal de login, role (ou clique "Close" se houver) para revelar mais conteúdo.
  • O título e uploader também aparecem no snapshot como link com StaticText — colete tudo junto.

Config

The MCP server is registered in ~/.hermes/config.yaml as ig-download:

mcp_servers:
  ig-download:
    command: node
    args:
      - /home/keila/ig-download-mcp-server/index.js

Pitfalls

  • Only works with PUBLIC Instagram content
  • Stories and Highlights require authentication (not supported)
  • parth-dl -o requires a full file path (e.g. /path/to/reel.mp4), not just a directory — passing a directory gives [Errno 21] Is a directory
  • Some Reels have no audio track (silent video-only posts). Always check with ffprobe before running transcription
  • If transcription returns empty segments but the video has an audio stream, try a larger whisper model (e.g. small or medium)
  • transcribe.py uses CPU by default — large models may be slow on CPU
  • The ig-download MCP server uses btch-downloader npm package which may break if Instagram API changes; parth-dl is the more reliable fallback
  • The provider (DeepSeek/OpenCode Zen) does NOT support vision — vision_analyze will fail with unknown variant image_url. Use browser_navigate to scrape captions instead
  • To reload MCP servers in Hermes after adding to config: restart session or /reload