Telegram Skill — User API Integration

All operations via Telethon (user API, not bot API). One session file for everything.

Setup (first run)

1. Get API credentials at https://my.telegram.org/apps → create app → copy api_id and api_hash
2. Set env vars in your .env:

   TG_API_ID=your_api_id
   TG_API_HASH=your_api_hash
   

3. Install Telethon: pip install telethon
4. Run setup diagnostic: python3 scripts/tg_setup.py
5. Authenticate (first time): python3 scripts/tg_setup.py --auth → enter phone + code from Telegram

Session file is saved automatically. All subsequent operations reuse it.

Set TG_SESSION env var to point to your session file, or the skill auto-discovers it.

1. Read Messages (any chat)

# Saved Messages
python3 scripts/tg_read.py --saved-messages --limit 10

# Specific chat/person
python3 scripts/tg_read.py --chat username --limit 20

# Public channel
python3 scripts/tg_read.py --channel SEOBAZA --limit 5

# With explicit session file
python3 scripts/tg_read.py --saved-messages --limit 5 --session /path/to/tg_session.session

Output: JSON lines (one per message) with: id, date, text, has_voice, has_audio, has_photo, has_document, mime_type, file_size, duration.

2. Parse Channels

Fetch messages from public channels with text sanitization (anti-injection, emoji cleanup, link extraction, dedup hashing).

# Single channel
python3 scripts/tg_fetch_channels.py --channel SEOBAZA --limit 20

# Multiple channels
python3 scripts/tg_fetch_channels.py --channels SEOBAZA,seo_inside,serpstat --limit 10

# Since date
python3 scripts/tg_fetch_channels.py --channel SEOBAZA --since 2026-03-01

# Save to file
python3 scripts/tg_fetch_channels.py --channel SEOBAZA --output messages.jsonl

# Pretty JSON
python3 scripts/tg_fetch_channels.py --channel SEOBAZA --pretty

Output: JSONL with: channel, channel_id, id, date, text, text_sanitized, links, hash, views, forwards.

Sanitization: strips zero-width chars, excessive emoji, backtick injection, truncates at 2000 chars.

3. Download Voice Notes

Fetch voice/audio/video_note messages and download the audio files.

# From Saved Messages
python3 scripts/tg_fetch_voice.py --saved-messages

# From specific chat
python3 scripts/tg_fetch_voice.py --chat username --limit 5

# Custom output directory
python3 scripts/tg_fetch_voice.py --saved-messages --audio-dir ./my-audio

Output: metadata JSONL (id, date, duration, media_type, file_path, file_size) + downloaded audio files.

Transcription: Use mlx-whisper or any Whisper model on the downloaded audio files. The skill handles Telegram I/O only — transcription is your choice of tool.

4. Setup & Diagnostics

# Full diagnostic (env, session, deps, connectivity)
python3 scripts/tg_setup.py

# Create/refresh session (interactive auth)
python3 scripts/tg_setup.py --auth

Architecture

skills/telegram/
├── SKILL.md              ← This file
├── scripts/
│   ├── tg_read.py        ← Read any chat (104 lines)
│   ├── tg_fetch_channels.py  ← Parse channels → JSONL (165 lines)
│   ├── tg_fetch_voice.py ← Download voice/audio (219 lines)
│   ├── tg_setup.py       ← Auth + diagnostics (198 lines)
│   └── sanitizer.py      ← Text sanitization (84 lines)
└── references/
    ├── setup.md          ← Detailed setup guide
    └── json-schema.md    ← Output format spec

Design principle: Scripts produce JSON. No business logic, no classification, no transcription. Pure Telegram I/O. Pipe the output to whatever you need.

Session Resolution

Scripts find the session file in this order:

1. --session CLI flag
2. TG_SESSION env var
3. .claude/data/telegram/tg_session.session (project-level)
4. ~/.telegram/tg_session.session (global fallback)

Limitations

• User API only — not a bot. You authenticate as yourself.
• Rate limits — FloodWaitError handled with auto-delay between channels.
• One active session — only one Telethon client per session at a time.
• Media download — large files throttled by Telegram.
• No sending — this skill reads only. It does not send messages.