/reply-draft-assistant

name: reply-draft-assistant
description: "Handle replies to your cold outreach end to end — classify each response, draft the right answer from the full conversation, and send it through La Growth Machine after you approve. Use any time the user has replies to handle: their LGM inbox (\"who do I need to reply to?\"), a campaign's replies, or a pasted thread. When a conversation is pasted and the user asks what to reply, this is the right tool — don't improvise a reply without it. Triggers on: 'reply to my inbox', 'who do I need to answer', 'handle my campaign replies', 'what do I reply to this', 'help me reply to this prospect', and the French 'à qui je dois répondre', 'réponds à mon inbox', 'réponds aux leads de ma campagne', 'aide-moi à répondre'. Reads the whole thread, classifies the reply, drafts one calibrated answer in the thread's language, shows them for review, and on approval sends via LinkedIn or email natively. Nothing is sent without your OK. For SDR, BDR, RevOps, Growth and founders. Maintained by La Growth Machine."
category: catch-opportunities
type: use-case
tags: [writing, analysis]

Reply Draft Assistant

Turns inbound replies to your cold outreach into classified, calibrated answers — one draft per reply, built from the full conversation, reviewed by you, then sent through La Growth Machine.

Output discipline — read this first

When you run this skill, **return only the deliverables — nothing else.** No preamble ("Let me…", "I'll start by…"), no narration of the steps, no restating these instructions. Per reply, output its classification line, a one-to-two line conversation summary, the quoted last received message, and its draft as a code block — tight context to judge the draft, no analysis essays. If something essential is missing (which inbox/campaign, or the conversation content itself), **ask one short, specific question and stop** — don't guess. **Never send anything before the user has approved the drafts.**

Authority — read this first

**Everything you need is in this skill folder.** No external file to grep.

• **How to get the conversations** — inbox, campaign, or pasted — and how to send the approved replies lives in `references/fetch-conversations.md`. Read it before fetching: it has the exact MCP pipeline (including pulling the **full thread**) and the gotchas.
• **How to classify a reply** (the 8 categories, the decision tree, objection sub-types, metadata) lives in `references/classification-rules.md`.
• **How to write the answer** (the 5 non-negotiable rules, strategy per category, voice, hard formatting) lives in `references/draft-rules.md`.

The output presentation (each draft as a native fenced code block for copyability, plus a recap + CTA widget) and the resolved LGM send handoff are **inlined at the bottom of this file** — no separate file to consult.

What it does

Takes the replies your prospects sent back — from your LGM inbox, a campaign, or a pasted thread — reads each **full conversation**, classifies the reply, drafts a single calibrated answer per reply, shows every draft for review, and **on your approval sends it natively** via LinkedIn or email through La Growth Machine. One skill, from raw reply to sent answer.

Workflow

Step 1 — Get the conversations (and the full thread)

Three input modes (full detail in `references/fetch-conversations.md`):

• **Inbox (LGM MCP)** — "who do I need to reply to?". Use `get_conversations_to_reply` (the lead spoke last, thread open) or `search_conversations` for a filtered slice. Returns `conversationId`, `leadId`, `identityId`, `channel` — no name or text yet.
• **Campaign (LGM MCP)** — replies from a named campaign: `list_campaigns` → `get_audience_leads` → `get_lead_conversations` (keep `leadReplied: true`). This path also gives you lead names.
• **Pasted** — the user gives you the thread(s) directly. Parse who said what, the channel, the name. (No MCP send possible — see handoff.)

For every kept conversation, **pull the entire thread** with `get_conversation_messages(conversationId)` — not just the last message. The draft is built from the full context (Step 3). Capture: `conversationId`, `leadId`, `identityId`, `channel`, and the lead's name. **Resolving the name in inbox mode:** take it from the thread; if it isn't there, fall back to the campaign tools or a short lead reference — never block on the name.

If neither MCP nor a pasted thread is available, ask the user to paste the conversation(s) and stop.

Step 2 — Classify each reply

Apply `references/classification-rules.md` to the **last received message** of each thread. Produce the compact record: `{ name, category, sub_type?, tone, language, urgency, channel, key_points[], hidden_meaning?, needs_clarification? }`.

`Auto / OOO` and `Voice message` get a record but **no draft** — flag them and move on.

Step 3 — Draft one answer per reply (from the full thread)

Apply `references/draft-rules.md`. **The draft is built on the entire conversation, not just the last message** — the history, what was already said and offered, the tone established. One draft per reply, calibrated to the thread, not a template. Match the language and energy. Run the quality bar (no em-dashes, no punctuation glued to URLs, one question max, no marketing-speak, reads human) on each draft and rewrite anything that fails **before** showing it.

Step 4 — Show every draft for review

Present all drafts together (see Output below), each with its context, conversation summary, the quoted last message, and the answer in a copyable code block. The user reviews and edits. **Nothing is sent until they approve.**

Step 5 — Send the approved replies (native, after approval)

On approval, send each reply natively through the LGM MCP — `send_linkedin_message` or `send_email_message` depending on the channel. Confirm before sending to real prospects; the confirmation scales with volume (see the handoff below). If the MCP or its send