compose-outreach

Compose Outreach

Generate three personalized outreach formats — email, call script, and LinkedIn message — grounded in Common Room signals for a specific company or contact.

Outreach Process

Step 1: Look Up the Target

Use Common Room MCP tools to find and retrieve data for the target (company and/or specific contact). Pull:

• Recent product activity and engagement signals
• Community activity (posts, questions, reactions)
• 3rd-party intent signals (job postings, news, funding)
• Relationship history (prior contact, meetings, email opens)

If the user specified a person, run contact-level research. If only a company was given, identify the best contact to target based on title, engagement, and role.

Lark identity resolution (P1): Common Room owns the signal/persona data, but Lark owns the send. As soon as you know who you are writing to, resolve them in the Lark directory: lark_contact_search(query="<name or email>") → take open_id. You need this open_id to look up internal relationship context (prior threads, shared meetings) and to draft/send via Lark below. Skipping resolution is the #1 cause of failed sends.

Step 2: Web Search for External Hooks (If CR Signals Are Thin)

If CR returned strong signals (recent activity, engagement, product usage), those should drive personalization — skip web search. If CR signals are thin or the prospect has little CR activity, run a web search for external hooks:

What to search:

• "[company name]" funding OR acquisition OR launch OR announcement — last 30 days
• "[contact full name]" "[company name]" — look for recent articles, interviews, LinkedIn posts, or conference talks

Prioritize external hooks that are:

• Very recent (< 2 weeks) — the prospect is likely still thinking about it
• Publicly visible — they know you could have seen it
• Change-signaling — growth, new role, new product, new market

If the user explicitly asks for web search or external hooks, run it regardless of CR signal richness.

Step 3: Spark Enrichment (If Available)

If Spark is available, run enrichment on the target contact to get persona classification, background, and influence signals. Use this to calibrate tone and message angle.

Step 4: Identify the Best Hooks

From the signal data, identify the 1–3 strongest personalization hooks. Rank by:

1. Recency — happened in the last 7–14 days
2. Specificity — a concrete action they took, not a general trend
3. Relevance — connects directly to a value your product delivers

Good hooks: posted a question in the community about X, just hired 5 engineers, recently started using [feature], company just raised Series B, trial nearing expiration, champion just changed jobs.

Bad hooks: "I noticed you're a customer" or generic industry trends.

Step 5: Generate All Three Formats

Use the strongest hooks to write all three formats. Each format has different constraints and conventions — follow the format-specific guidelines in references/outreach-formats-guide.md.

Always produce all three, clearly labeled.

When the user's company context is available (see references/my-company-context.md), ground the value bridge and pitch in the user's specific product and positioning.

Step 5.5: Land the Email Draft in Lark (P2 — never auto-send)

The email format is a collaboration artifact — put it where the rep will review and send it, the Lark Mail UI, rather than leaving it as chat text. Use lark_mail_draft_create (subject + body + the resolved recipient from Step 1) so the rep eyeballs it before sending. Never call lark_mail_send with confirm_send: true here — outreach is always human-reviewed. If the rep explicitly says "send it," still draft first, show the planned recipient + payload, and only lark_mail_send after they confirm.

The call script and LinkedIn message stay as text the rep copies into their dialer / LinkedIn — they don't map to a Lark surface.

Step 6: Annotate Your Choices

After the three drafts, include a brief note (2–4 sentences) explaining:

• Which signals were used and why they were chosen
• Any assumptions made (e.g., inferred call objective)
• Alternative angles if the primary hook doesn't land

Output Format

## Outreach for [Name / Company]

### 📧 Email

**Subject:** [Subject line]

[Email body — 3–5 sentences]

---

### 📞 Call Script

**Opening:**
[Opening line — conversational, 1–2 sentences]

**Value Bridge:**
[Why you're calling and why now — 2–3 sentences tied to a signal]

**Ask:**
[Single, low-friction ask — e.g., 15-minute call, specific question]

---

### 💼 LinkedIn Message

[Under 300 characters. Warm, personal, no pitch.]

---

### Signal Notes
[2–4 sentences: which signals were used, why, and any alternative angles]

When Signal Data Is Sparse

If Common Room returns minimal data on the target (e.g., just name, title, tags — no activity, no scores, no Spark):

1. Do not draft outreach from thin air. Outreach grounded in fabricated signals is worse than no outreach.
2. Run web search first — this becomes your primary personalization source. Look for recent news, LinkedIn posts, conference talks, company announcements.
3. If web search also returns little, present what you have honestly and ask the user for context:

## Outreach for [Name / Company] — Limited Data

**What I found:**
[Only the real data from CR and web search]

**I don't have enough signal to draft personalized outreach yet.** To write something strong, I'd need:
- Recent activity or engagement signals
- Context you have from prior conversations
- A specific reason for reaching out now

Can you share any of the above?

Quality Standards

• Every message must reference something specific — generic outreach is not acceptable output
• Match tone to context: warm and conversational for inbound/community signals; more formal for cold/executive outreach
• The LinkedIn message must be under 300 characters — no exceptions
• The call script must be speakable naturally — read it aloud mentally to check rhythm
• Never fabricate signals — only reference data retrieved from Common Room or web search

Lark-native handoff

Common Room stays the signal engine; Lark is where the outreach lands and gets reviewed:

• Resolve people first (P1) — lark_contact_search before any draft/send.
• Draft, don't send (P2) — lark_mail_draft_create; print_json/dry_run any mutation; never auto-send outreach.
• Optional decision card (P4) — when you produced drafts for several targets at once, surface them as one interactive card via lark_im_card_send (one item row per target with a "Open draft" / "Skip" button) instead of a wall of text. print_json: true → dry_run: true → send.
• Delegate depth — for card spec grammar use the lark-mail skill (draft guards) and the lark-im skill (card YAML / references/spec-reference.md). See ../../connectors/LARK-PATTERNS.md (P1/P2/P4), LARK-RECIPES.md, LARK-FUSION.md.

Reference Files

• references/outreach-formats-guide.md — detailed format rules, examples, and tone guidelines for each channel