draft-replies-without-sending.md

Skill: Draft Replies Without Sending

Purpose

Inspects recent conversation context in a specified Telegram chat, prepares a contextual and tone-appropriate response, and either returns it as text for user review or saves it into Telegram's native input field as an unposted draft. Under no circumstances does this skill transmit or send a message.

---

When to Use / Use Cases

• Preparing a thoughtful response to a client, manager, or collaborator before sending.
• Reviewing sensitive messages where the agent provides options for human approval.
• Staging a draft inside the Telegram mobile or desktop app so the user can edit or send it with a single tap.
• Operating in high-security or read-only environments where outbound communication requires human-in-the-loop authorization.

---

Prerequisites / Configuration

• Telegram MCP Server running and authenticated.
• Optional: TELEGRAM_EXPOSED_TOOLS=read-only (if drafting text-only) or TELEGRAM_EXPOSED_TOOLS=read-only+save_draft (if allowing native Telegram drafts).
• Invariant Rule: Tool send_message, reply_to_message, and forward_message must never be invoked.

---

MCP Tools Used

1. get_message_context
   • Parameters: chat_id=<chat_id>, message_id=<message_id>, context_size=3
   • Purpose: Pulls messages immediately preceding and following the target message to understand the exact conversational flow.
2. save_draft (Optional)
   • Parameters: chat_id=<chat_id>, message=<draft_text>, reply_to_msg_id=<message_id>
   • Purpose: Saves the drafted response into Telegram's native draft storage. The draft appears in the client's input field ready for the user to review and send manually.
3. send_message / reply_to_message (Strictly Prohibited)
   • Invariant: Neither of these tools may be called in this workflow.

---

Inputs

• Chat Identifier: Username, phone, or numeric ID (e.g., @sarah_pm or 12345678).
• Target Message ID: The specific message requiring a reply.
• Intent / Guidance: User instructions regarding tone, constraints, or key information (e.g., "Confirm we can meet tomorrow at 3 PM, but ask if Zoom is preferred").
• Save Mode: Whether to return the draft in the chat or also stage it via save_draft.

---

Outputs

• Context Inspection Summary: Brief summary of the thread context.
• Draft Proposal: Formatted proposed reply text presented in a fenced block.
• Explanation: Why this wording was chosen and notes on specific tone nuances.
• Confirmation: Confirmation that no message was sent.

---

Step-by-Step Example

Step 1: User Prompt

User: "Sarah asked about the project delivery date in chat @sarah_pm (message ID 542). 
Draft a polite reply letting her know we are on track for Thursday, but do NOT send it yet. 
Save it as a draft in Telegram if possible."

Step 2: Agent Inspects Surrounding Context

The agent calls get_message_context to understand Sarah's exact question:

{
  "name": "get_message_context",
  "arguments": {
    "chat_id": "@sarah_pm",
    "message_id": 542,
    "context_size": 2
  }
}

Tool Mock Response:

{
  "metadata": {
    "chat_id": "@sarah_pm",
    "target_message_id": 542
  },
  "data": [
    {
      "id": 540,
      "sender": "Omid",
      "date": "2026-09-21 15:30:00",
      "text": "We finished the core API integrations today."
    },
    {
      "id": 541,
      "sender": "Sarah PM",
      "date": "2026-09-21 16:10:00",
      "text": "Awesome progress!"
    },
    {
      "id": 542,
      "sender": "Sarah PM",
      "date": "2026-09-22 08:15:00",
      "is_target": true,
      "text": "Are we still good for the Thursday release demo, or do you need extra time for QA?"
    }
  ]
}

Step 3: Agent Stages Native Draft (Non-transmitting)

The agent calls save_draft (which saves to Telegram's cloud draft storage without sending):

{
  "name": "save_draft",
  "arguments": {
    "chat_id": "@sarah_pm",
    "message": "Hi Sarah! Yes, we are fully on track for Thursday's demo. QA testing is already underway on staging and looking solid.",
    "reply_to_msg_id": 542
  }
}

Tool Mock Response:

"Draft saved to chat @sarah_pm. Open the chat in Telegram to see and send it."

Step 4: Agent Presents Draft for User Review

### 📝 Draft Prepared (Not Sent)

**Target Chat:** @sarah_pm  
**Replying to Message #542:** *"Are we still good for the Thursday release demo, or do you need extra time for QA?"*

#### Proposed Reply:
> Hi Sarah! Yes, we are fully on track for Thursday's demo. QA testing is already underway on staging and looking solid.

---
✅ **Status:**
- **No message was sent.**
- The text above has been saved as a **Draft** in your Telegram app for `@sarah_pm`.
- You can open Telegram, review the text in the message box, edit any details, and press Send whenever you are ready.

---

Safety / Privacy Notes

• Strict "Do Not Send" Invariant: The system prompt and tool constraints guarantee that tools with mutating delivery effects (send_message, reply_to_message) are never invoked.
• Safe Staging via save_draft: Saving a draft invokes MTProto's messages.SaveDraftRequest. This populates the user's local message input box on all connected devices (mobile, desktop, web) but does not trigger push notifications or broadcast any message to the recipient.
• Privacy Allowlist: Respects ALLOWED_CHAT_IDS.

---

Failure Modes & Recovery

1. Accidental Send Prevention:
   • Failure Mode: Agent misunderstands instructions and attempts to call send_message.
   • Guardrail: Run the server with TELEGRAM_EXPOSED_TOOLS=read-only+save_draft to physically prevent send_message from being registered as an MCP tool.
2. Target Message Not Found:
   • Error: Message with ID 542 not found in chat @sarah_pm.
   • Recovery: Call list_messages(chat_id="@sarah_pm", limit=5) to retrieve recent valid message IDs and ask user for clarification.
3. Invalid Peer or Chat Restrictions:
   • Error: ChatAccessDeniedError or PeerIdInvalidError.
   • Recovery: Check chat permissions and ensure the user has initiated conversation with that peer.