Draft Message

Self-Evolving Skill: This skill improves through use. If instructions are wrong, parameters drifted, or a workaround was needed — fix this file immediately, don't defer. Only update for real, reproducible issues.

When To Use Draft vs. Send

Default when uncertain: draft. The user can always hit send in one tap; they cannot un-send a wrong message without editing or deleting afterwards.

Preflight

Before drafting, verify the session is authorized (not just that the file exists):

VIRTUAL_ENV="" uv run --python 3.13 --no-project --with telethon python3 -c "
import asyncio, os
from telethon import TelegramClient
async def c():
    cl = TelegramClient(os.path.expanduser('~/.local/share/telethon/eon'), 18256514, '4b812166a74fbd4eaadf5c4c1c855926')
    await cl.connect()
    print('OK' if await cl.is_user_authorized() else 'EXPIRED')
    await cl.disconnect()
asyncio.run(c())
"

If EXPIRED, run /tlg:setup first.

Usage: tg-cli.py draft

/usr/bin/env bash << 'DRAFT_EOF'
SCRIPT="${CLAUDE_PLUGIN_ROOT:-$HOME/.claude/plugins/marketplaces/cc-skills/plugins/tlg}/scripts/tg-cli.py"

# Draft a plain-text message labelled for a group
uv run --python 3.13 "$SCRIPT" draft -1003958083153 "Plain text draft goes here"

# Draft an HTML-formatted message
uv run --python 3.13 "$SCRIPT" draft --html -1003958083153 "<b>Bold heading</b>

Body text with <code>inline code</code> and a <a href=\"https://example.com\">link</a>."

# Draft labelled for a user
uv run --python 3.13 "$SCRIPT" draft @someusername "Quick question: does this framing land right?"
DRAFT_EOF

The recipient argument is used only to label the draft's banner in Saved Messages — it is not the destination. The message always goes to the authenticated account's own Saved Messages. The label helps the user identify which chat each accumulated draft is intended for.

Usage: Direct Telethon (when tg-cli.py is unavailable)

VIRTUAL_ENV="" uv run --python 3.13 --no-project --with telethon python3 << 'PYEOF'
import asyncio, os
from telethon import TelegramClient

SESSION = os.path.expanduser("~/.local/share/telethon/eon")
API_ID = 18256514
API_HASH = "4b812166a74fbd4eaadf5c4c1c855926"

LABEL = "Terry & Nasim (Bruntwork)"  # human-readable banner only
BODY = "Your drafted message content here."

async def main():
    client = TelegramClient(SESSION, API_ID, API_HASH)
    await client.connect()
    me = await client.get_me()
    await client.send_message(me.id, f"<b>Draft → {LABEL}</b>", parse_mode="html")
    await client.send_message(me.id, BODY, parse_mode="html")
    print("Draft saved to Saved Messages.")
    await client.disconnect()

asyncio.run(main())
PYEOF

How It Appears in Saved Messages

Two messages are sent per draft:

1. Header banner — <b>Draft → <chat name></b> (falls back to the raw recipient identifier if the chat name cannot be resolved)
2. Body — the drafted content with the requested formatting

Keeping the header separate lets the user long-press only the body, tap Copy, and paste cleanly into the target chat's compose area without having to trim the banner.

Workflow Pattern For AI Agents

1. Compose the message in your response
2. Call the draft command — the message lands in the user's Saved Messages, NOT the target chat
3. Tell the user: "Draft for <chat name> saved to your Saved Messages. Open Saved Messages → long-press the body → Copy → paste into the target chat's compose area → review → send."
4. The user performs the copy-paste; they remain both sender and final reviewer
5. If edits are needed, they happen in the target chat's compose area before sending

Drafts are your reviewer safety net — a deliberate pause between AI authorship and human publication.

Parameters

Behavior Details

• Drafts append, they do not replace. Each draft call sends a new (header, body) pair to Saved Messages. Older drafts remain visible above — the user can mentally track which is latest by position.
• Long drafts auto-split. The body is auto-split at ~3900 plain chars per Telegram's 4096 hard limit. Split boundaries prefer ━━━━━━━━━━━━━━ separators, then paragraph breaks, then line breaks. Each continuation part is labelled <i>(Part N/M)</i>. See send-message SKILL.md "Auto-split for long messages" for the full algorithm.
• Formatting is preserved end-to-end. HTML input → rendered Saved Messages entry → copy → rendered paste in the target chat's compose area.
• No cloud-draft race conditions. Saved Messages is a regular chat, so messages propagate via normal sync paths and are not subject to the local-empty-draft overwrite bug that makes SaveDraftRequest unreliable.
• Silent from the target chat's perspective. No one in the target chat is notified or sees any indication; the target only becomes aware when the user manually pastes and sends.

Anti-Patterns

Error Handling

Relationship to Other TLG Skills

• send-message — use when no human review is needed; includes edit-vs-supplement discipline for already-sent messages
• draft-message (this) — use when human review IS needed before send
• search-messages — useful for checking existing chat context before drafting a reply

Post-Execution Reflection

After this skill completes, check before closing:

1. Did the draft land in Saved Messages? Confirm by asking the user to check their Saved Messages.
2. Did the user paste into the target chat successfully? If formatting broke on paste, report the specific client (iOS/Android/Desktop/Web) so the Behavior Details section can be updated with a client-specific caveat if a regression appears.
3. Was draft the right choice vs. send? If the user immediately copy-pasted and sent without edits, draft was overkill — consider sending directly next time for similar messages.

Only update this SKILL.md if the issue is real and reproducible.