Claude Behavior Guidelines

Draft vs Send: Follow User's Intent

User says	Claude does	Clarify?
"драфт", "draft", "сделай драфт"	`draft`	No
"отправь", "пошли", "send"	`send`	No
"напиши сообщение" (ambiguous)	Ask what user wants	Yes

Key Rules

•Explicit draft → draft: "драфт", "draft" → use draft command immediately
•Explicit send → send: "отправь", "пошли", "send" → use send command immediately
•Ambiguous → clarify: If neither "draft" nor "send" verb present, ask: "Создать драфт или сразу отправить?"

Examples

User: "сделай драфт для lv: привет" Claude: Uses draft --chat "lv" --text "привет" immediately

User: "отправь сообщение Маше: встретимся в 5?" Claude: Uses send --chat "Маша" --text "встретимся в 5?" immediately

User: "напиши сообщение для Маши: встретимся в 5?" Claude: Asks "Создать драфт или сразу отправить?"

Telegram Telethon Skill

Full Telethon API wrapper with daemon mode and Claude Code integration. Supports interactive setup, background message monitoring, and automatic Claude session spawning per chat.

Package Layout

code

telegram-telethon/
├── SKILL.md                          # This file
├── pyproject.toml                    # Installable Python package
├── scripts/
│   ├── tg.py                         # Main CLI (messages, media, drafts, etc.)
│   └── tgd.py                        # Daemon controller
├── src/telegram_telethon/            # Importable package
│   ├── core/                         # auth, config
│   ├── modules/                      # messages, media
│   ├── daemon/                       # runner, handlers, claude_bridge
│   └── utils/                        # formatting
└── tests/                            # pytest unit + integration tests

Scripts import from src/telegram_telethon. Install the package in editable mode so tg.py/tgd.py can resolve imports:

bash

cd telegram-telethon
pip install -e .
# or with dev tools (pytest, coverage):
pip install -e ".[dev]"

Relationship to `telegram` Skill

The separate telegram skill (single-script telegram_fetch.py backed by telegram_dl) overlaps on list/recent/search/send/edit/download/thread but differs:

With the publish/markdown/schedule ports now complete, telegram-telethon is a superset of telegram on everything except the external telegram_dl auth dependency. Use this skill for:

•publish — draft→channel workflow (frontmatter, media albums, post-publish move + index update, post-flight lint)
•--markdown on send / publish — markdown→Telegram HTML conversion
•--schedule on send / publish — ISO / relative / natural-language scheduled delivery
•Daemon mode + Claude Code spawning, voice transcription (Telegram/Groq/Whisper), delete / forward / mark-read, local draft / drafts / draft-send, lint-channel, non-interactive auth setup.

Prerequisites

Interactive Setup (Terminal)

Run setup wizard on first use:

bash

python3 scripts/tg.py setup

This guides through:

•Getting API credentials from https://my.telegram.org/auth
•Phone number verification
•2FA (if enabled)
•Optional daemon trigger configuration

Non-Interactive Setup (Claude Code)

For use from Claude Code or scripts without TTY:

bash

# Step 1: Provide credentials and trigger code send
python3 scripts/tg.py setup --api-id 12345678 --api-hash abc123... --phone +1234567890

# Step 2: User receives code on phone, then complete auth
python3 scripts/tg.py setup --api-id 12345678 --api-hash abc123... --phone +1234567890 --code 12345

# If 2FA enabled, add password
python3 scripts/tg.py setup --api-id 12345678 --api-hash abc123... --phone +1234567890 --code 12345 --password mypassword

The script auto-detects TTY and switches between interactive/non-interactive modes.

Quick Start

bash

# Check connection status
python3 scripts/tg.py status

# List chats
python3 scripts/tg.py list

# Get recent messages from a chat
python3 scripts/tg.py recent "John Doe" --limit 20

# Search messages
python3 scripts/tg.py search "meeting notes"

# Configure daemon triggers interactively
python3 scripts/tg.py daemon-config

# Start daemon (foreground with logs)
python3 scripts/tgd.py start --foreground

# Start daemon (background)
python3 scripts/tgd.py start

# View daemon logs
python3 scripts/tgd.py logs

CLI Commands

Message Operations

bash

# List all chats
python3 scripts/tg.py list [--limit 30] [--search "term"]

# Fetch recent messages
python3 scripts/tg.py recent [CHAT] [--limit 50] [--days 7] [--format markdown|json] [--output file.md]

# Search messages by content
python3 scripts/tg.py search QUERY [--chat "Chat Name"] [--limit 50] [--format markdown|json]

# Fetch unread messages
python3 scripts/tg.py unread [--chat "Chat Name"] [--format markdown|json]

# Fetch forum thread
python3 scripts/tg.py thread CHAT_ID THREAD_ID [--limit 100]

# Send message
python3 scripts/tg.py send --chat "Chat Name" --text "Message text" [--reply-to MSG_ID] [--file path] [--topic TOPIC_ID] [--markdown] [--schedule "+1h" | "tomorrow 10:00" | "2026-04-10T09:30"]

# Edit message
python3 scripts/tg.py edit --chat "Chat Name" --message-id MESSAGE_ID --text "New text"

# Delete messages
python3 scripts/tg.py delete --chat "Chat Name" --message-ids 123 456 789 [--no-revoke]

# Forward messages
python3 scripts/tg.py forward --from "Source" --to "Dest" --message-ids 123 456

# Mark messages as read
python3 scripts/tg.py mark-read --chat "Chat Name" [--max-id MSG_ID]

Draft Operations

bash

# Save/update a draft message
python3 scripts/tg.py draft --chat "Chat Name" --text "Draft text" [--reply-to MSG_ID] [--no-preview]

# Clear a draft (save empty text)
python3 scripts/tg.py draft --chat "Chat Name" --text ""

# Clear all drafts
python3 scripts/tg.py draft --clear-all

# List all drafts
python3 scripts/tg.py drafts [--limit 50]

# Send a draft as a message (clears the draft)
python3 scripts/tg.py draft-send --chat "Chat Name"

Note: Use "me" as the chat name to target Saved Messages (your own chat). The literal name "Saved Messages" doesn't work as it's localized differently per user.

Media Operations

bash

# Download media from chat
python3 scripts/tg.py download "Chat Name" [--limit 5] [--output-dir ~/Downloads] [--message-id ID] [--type voice|video|photo]

# Transcribe a single voice message (MESSAGE_ID required)
python3 scripts/tg.py transcribe "Chat Name" MESSAGE_ID [--method telegram|groq|whisper]

# Batch-transcribe recent voice messages (omit MESSAGE_ID, use --batch)
python3 scripts/tg.py transcribe "Chat Name" --batch [--limit 10] [--method telegram|groq|whisper]

Publish a Draft to a Channel

End-to-end publish workflow: parse a draft markdown file, resolve the destination channel from folder structure or frontmatter channel:, upload media (single file or album), post-process to move the draft to published/ and insert an entry in the channel index.

bash

# Dry-run preview
python3 scripts/tg.py publish --draft "Channels/klodkot/drafts/20260416-post.md" --dry-run

# Publish now
python3 scripts/tg.py publish --draft "20260416-post"  # slug works too

# Publish scheduled for later
python3 scripts/tg.py publish --draft "..." --schedule "tomorrow 10:00"

The result JSON includes published, channel, message_id, media_count, moved_to, and — crucially — lint_warnings when the final body contains leaked markdown/HTML that Telegram wouldn't render. Post-publish bookkeeping failures (e.g. index write error) surface as warnings but don't roll back the send.

Markdown Formatting on Send

Pass --markdown to convert a markdown-flavored message into Telegram HTML before sending:

bash

python3 scripts/tg.py send --chat "@mychannel" --markdown \
  --text $'## Release\n\n**v2** ships _today_. See [docs](https://example.com).\n\n* fast\n* stable'

Rules (applied in order): ## Header → bold line; * item / - item at line start → → item; **bold** → <b>; _italic_ → <i>; [text](url) → <a href>. Pre-existing HTML passes through unchanged, so the flag is safe to add to content that was already authored as HTML.

Pair with lint-channel below to catch cases where --markdown was forgotten.

Scheduled Delivery

Pass --schedule with one of three formats (naive times default to Europe/Berlin):

bash

# Relative: send in one hour
python3 scripts/tg.py send --chat "@mychannel" --text "..." --schedule "+1h"

# Natural: send tomorrow morning
python3 scripts/tg.py send --chat "@mychannel" --text "..." --schedule "tomorrow 09:30"

# Absolute: send at a specific time
python3 scripts/tg.py send --chat "@mychannel" --text "..." --schedule "2026-04-20T15:00"

The response includes "scheduled_for": "<iso datetime>" when the message is queued for later. Telegram displays scheduled messages in the chat's scheduled-messages view.

Lint Published Messages

Scan a channel (or a single message) for unrendered markdown/HTML that leaked into the raw message text — i.e. the sender forgot --markdown or the HTML conversion failed, so readers see literal **bold**, <b>…</b>, [text](url), or ## Header in the post.

bash

# Scan last 50 messages in @mychannel
python3 scripts/tg.py lint-channel --chat "@mychannel"

# Scan last 200 messages
python3 scripts/tg.py lint-channel --chat "@mychannel" --limit 200

# Lint a single message by ID
python3 scripts/tg.py lint-channel --chat "@mychannel" --message-id 1234

# Machine-readable output for pipelines / CI
python3 scripts/tg.py lint-channel --chat "@mychannel" --json

The detector lives in modules/lint.py as a pure function (detect_unrendered_markup(text, entities)), so it can also be called directly on drafts or wired into a post-flight check after publishing. Content inside MessageEntityCode/MessageEntityPre spans is ignored (inline code / code blocks are expected to contain raw characters).

Obsidian Integration

--to-daily and --to-person are flags on the read commands (recent, search, unread), not standalone subcommands:

bash

# Append recent messages to today's daily note (Daily/YYYYMMDD.md in the active vault)
python3 scripts/tg.py recent "Chat Name" --to-daily

# Append search results to today's daily note
python3 scripts/tg.py search "query" --to-daily

# Append recent messages to a person's note
python3 scripts/tg.py recent "Chat Name" --to-person "Person Name"

The target vault path is resolved by the formatting helpers in utils/formatting.py; there are currently no --vault or --section overrides on the CLI.

Voice Transcription

The skill supports three transcription methods with automatic fallback:

•Telegram API (default) - Uses Telegram Premium's server-side transcription
•Groq - Uses Groq's Whisper API (requires GROQ_API_KEY environment variable)
•Whisper - Uses local OpenAI Whisper model (requires pip install openai-whisper)

bash

# Use Telegram's transcription (Premium feature)
python3 scripts/tg.py transcribe "Chat" 123

# Force Groq transcription
python3 scripts/tg.py transcribe "Chat" 123 --method groq

# Force local Whisper
python3 scripts/tg.py transcribe "Chat" 123 --method whisper

Daemon Mode

The daemon monitors Telegram for messages matching configured triggers and can:

•Reply with static text
•Spawn Claude Code sessions to handle requests
•Resume existing Claude sessions per-chat
•Queue requests to prevent rate limiting

Trigger Configuration

Triggers are stored in ~/.config/telegram-telethon/daemon.yaml:

yaml

triggers:
  # Respond to /claude command in DMs
  - chat: "@myusername"
    pattern: "^/claude (.+)$"
    action: claude
    reply_mode: inline

  # Respond to @Bot mentions in a group
  - chat: "AI Assistants"
    pattern: "@Bot (.+)$"
    action: claude
    reply_mode: new

  # Simple ping-pong in any chat
  - chat: "*"
    pattern: "^/ping$"
    action: reply
    reply_text: "pong"

claude:
  allowed_tools:
    - Read
    - Edit
    - Bash
    - WebFetch
  max_turns: 10
  timeout: 300

queue:
  max_concurrent: 1
  timeout: 600

Trigger Fields

Field	Description
`chat`	Chat name, `@username`, or `*` for all chats
`pattern`	Regex pattern (capture group 1 becomes Claude prompt)
`action`	`claude`, `reply`, or `ignore`
`reply_mode`	`inline` (reply to message) or `new` (separate message)
`reply_text`	Static text for `reply` action

Claude Integration

When action is claude:

•Text captured by regex group 1 is sent to Claude Code via claude -p "..." --output-format json
•Claude sessions persist per-chat in sessions.json
•Subsequent messages from same chat resume session via --resume <session_id>
•Responses are sent back to Telegram as reply or new message

Session Persistence

Claude sessions are saved to ~/.config/telegram-telethon/sessions.json:

•Each chat_id maps to a Claude session_id
•Sessions survive daemon restarts
•Track message count and last used timestamp

To reset: delete chat entry from sessions.json or configure a /reset trigger.

Example Configurations

Personal AI Assistant

Respond to all DMs to yourself:

yaml

triggers:
  - chat: "@yourusername"
    pattern: "(.+)"
    action: claude
    reply_mode: inline

Group Bot with Mention Trigger

Only respond when @mentioned:

yaml

triggers:
  - chat: "Dev Team"
    pattern: "@AssistantBot (.+)"
    action: claude
    reply_mode: inline

Multi-Action Setup

yaml

triggers:
  - chat: "*"
    pattern: "^/ask (.+)"
    action: claude
    reply_mode: inline

  - chat: "*"
    pattern: "^/ping$"
    action: reply
    reply_text: "pong"

  - chat: "Noisy Group"
    pattern: ".*"
    action: ignore

File Structure

code

~/.config/telegram-telethon/
├── config.yaml        # API credentials (api_id, api_hash, phone)
├── daemon.yaml        # Daemon triggers and Claude config
├── session.session    # Telethon session file
├── sessions.json      # Claude session persistence
└── daemon.log         # Daemon log file

Development

bash

# Install with dev dependencies
cd telegram-telethon
pip install -e ".[dev]"

# Run all tests
pytest

# Run with coverage
pytest --cov=telegram_telethon

# Run specific test file
pytest tests/unit/test_claude_bridge.py -v

Example User Requests

Mapping natural-language asks to commands:

User says	Command
"Is Telegram connected?"	`status`
"What chats do I have?"	`list`
"Find chat named X exactly"	`list --search "X"` (increase `--limit` if not found)
"Show recent messages from John"	`recent "John" --limit 20`
"Messages from the last week in Group Y"	`recent "Group Y" --days 7`
"Search Telegram for 'deadline'"	`search "deadline"`
"Unread messages from Group Z"	`unread --chat "Group Z"`
"Mark Group Z as read"	`mark-read --chat "Group Z"`
"Get thread 174 in Lab"	`thread <chat_id> 174 --limit 100`
"Send 'hi' to John" / "отправь John: hi"	`send --chat "John" --text "hi"`
"Post a markdown-formatted note to @channel"	`send --chat "@channel" --markdown --text "..."`
"Schedule this for tomorrow at 10am"	`send --chat "..." --text "..." --schedule "tomorrow 10:00"`
"Send this in an hour"	`send --chat "..." --text "..." --schedule "+1h"`
"Reply thanks to message 12345"	`send --chat "..." --text "thanks" --reply-to 12345`
"Send image.jpg to John"	`send --chat "John" --file image.jpg`
"Save a draft for John: hi" / "сделай драфт"	`draft --chat "John" --text "hi"`
"List my drafts"	`drafts`
"Send the draft for John"	`draft-send --chat "John"`
"Delete messages 123, 456 from John"	`delete --chat "John" --message-ids 123 456`
"Forward msg 789 from John to Maria"	`forward --from "John" --to "Maria" --message-ids 789`
"Edit message 76 in @channel"	`edit --chat "@channel" --message-id 76 --text "..."`
"Download last 5 voice notes from John"	`download "John" --type voice --limit 5`
"Transcribe voice message 512 from John"	`transcribe "John" 512`
"Batch-transcribe recent voices from John"	`transcribe "John" --batch --limit 10`
"Add John's messages to daily note"	`recent "John" --to-daily`
"Add messages to a person's note"	`recent "Chat" --to-person "Person Name"`
"Publish this draft to the klodkot channel"	`publish --draft "20260416-post"`
"Preview a draft before publishing"	`publish --draft "..." --dry-run`
"Publish this at 10am tomorrow"	`publish --draft "..." --schedule "tomorrow 10:00"`
"Check if @mychannel has unrendered markup"	`lint-channel --chat "@mychannel"`
"Lint message 1234 in @mychannel"	`lint-channel --chat "@mychannel" --message-id 1234`
"Start the Telegram daemon"	`python3 scripts/tgd.py start` (or `--foreground`)
"Show daemon logs"	`python3 scripts/tgd.py logs`
"Configure daemon triggers"	`daemon-config`

Saved Messages: Use "me" (not "Saved Messages") — the label is localized per user.

Troubleshooting

Issue	Solution
"Config not found"	Run `python3 scripts/tg.py setup`
"Session expired"	Delete `session.session` and re-run setup
`ModuleNotFoundError: telegram_telethon`	Run `pip install -e .` from the skill directory
"Claude timeout"	Increase `timeout` in `daemon.yaml`
"Queue full"	Reduce request rate or wait
"No trigger matched"	Check `pattern` regex and `chat` name match
Chat not found by name	Increase `--limit` on `list` (default 30); may not be in recent dialogs

Claude Behavior Guidelines

Draft vs Send: Follow User's Intent

Key Rules

Examples

Telegram Telethon Skill

Package Layout

Relationship to telegram Skill

Prerequisites

Interactive Setup (Terminal)

Non-Interactive Setup (Claude Code)

Quick Start

CLI Commands

Message Operations

Draft Operations

Media Operations

Publish a Draft to a Channel

Markdown Formatting on Send

Scheduled Delivery

Lint Published Messages

Obsidian Integration

Voice Transcription

Daemon Mode

Trigger Configuration

Trigger Fields

Claude Integration

Session Persistence

Example Configurations

Personal AI Assistant

Group Bot with Mention Trigger

Multi-Action Setup

File Structure

Development

Example User Requests

Troubleshooting

Relationship to `telegram` Skill