cold-email-outreach

Cold Email Outreach

End-to-end cold outreach: research, draft, send, follow up, route replies. Strategy is grounded in proven hook frameworks (number-led / question / pain-point / benefit-first); the execution runs on Apollo, Firecrawl, the LinkedIn scraper, and Gmail through the Hyper MCP.

Out of scope — defer to other skills

Requirements

• Hyper MCP installed and connected. https://app.hyperfx.ai/mcp
• Gmail integration connected at https://app.hyperfx.ai/apps — supplies the sending account.
• Apollo integration connected — supplies prospect search and email enrichment.
• Firecrawl (bundled) — for company-page signals.
• Optional: LinkedIn scraper (bundled, runs through Apify) — for richer per-prospect personalization.

If gmail_messages_send and apollo_mixed_people_search are not in the agent's tool list, stop and tell the user to enable the Hyper MCP and connect Gmail + Apollo.

Tool surface

Critical rules

1. Never loop apollo_people_match for multiple prospects. For 2+ records always batch into apollo_people_bulk_match. Apollo's tool description warns about this explicitly — looping single-match calls burns credits and is much slower.
2. Default send mode = drafts-first for review. For any campaign with 4+ prospects, draft the first 1–3 with gmail_drafts_create, show them to the user, get explicit approval, then batch-send the rest with gmail_messages_send. Never send a full campaign without showing samples first.
3. One label per campaign. Create a cold/<campaign-name> label with gmail_labels_create at the start, apply it to every send, then track replies by searching that label. This is what makes Phase 6 reply routing actually work.
4. Stay under Gmail's send limits. ~500 messages/day per consumer Gmail account, ~2,000/day per Workspace user. Space sends out — see references/deliverability.md for warming and per-day pacing.
5. Personalization must connect to the problem. If the personalized opener could be deleted and the email still makes sense, it isn't doing any work. The opener should naturally bridge into why you're emailing.
6. One ask per email, one CTA. Interest-based (Worth exploring?) beats meeting requests on cold touch 1.
7. Honor unsubscribes immediately. Apply an unsubscribed label on any "remove me / not interested" reply and never re-target that address from the same Hyper workspace.

Workflow

Phase 1 — Define the campaign (always do this first)

Get the user to commit to:

1. ICP — Role(s), industry, company size, tech stack, geography. Concrete: "Heads of Growth at US-based pre-seed-to-Series-A B2B SaaS, 10–50 employees, using HubSpot."
2. The ask — What does a "yes" look like? (15-min call, async reply, demo, intro to someone else.)
3. Value prop in one sentence — "We help X do Y so they can Z."
4. Proof point — One specific result: "We helped Notion cut their CAC by 31% in 90 days." (Made up examples are worse than no example — get a real one.)
5. Trigger / signal (optional but powerful) — Funding round, hiring, pricing-page change, recent blog post, product launch, leadership change.
6. Sender + reply-to — Which Gmail account is sending. (Confirm with gmail_labels_list to verify the integration is live.)
7. Volume + cadence — Total prospects, max sends/day, follow-up gap pattern.

If they're stuck on any of these, push back. A campaign without proof or a clear ask will not perform regardless of how clever the writing is.

Phase 2 — Build & enrich the prospect list

# Search by ICP
apollo_mixed_people_search(
  person_titles=["Head of Growth", "VP Growth", "Director of Growth"],
  organization_num_employees_ranges=["11,50"],
  person_locations=["United States"],
  per_page=50,
)

Then for the prospects you actually want to contact, batch-enrich for emails:

# CORRECT — one bulk call for many prospects
apollo_people_bulk_match(
  details=[
    {"first_name": "...", "last_name": "...", "domain": "..."},
    ...up to 10 per call...
  ],
  reveal_personal_emails=True,
)

Only fall back to apollo_people_match for single-prospect lookups (e.g., the user pastes one LinkedIn URL).

For deeper company-level context (industry, revenue range, tech stack), call apollo_mixed_companies_search by organization name on the companies you want to enrich. Note that person search results already include core company fields (headcount, industry, location) — only reach for apollo_mixed_companies_search when you need data beyond what the person search returns.

Phase 3 — Per-prospect signals (the personalization layer)

For each prospect, gather one specific observation that connects to the problem you solve. Use the cheapest signal that works:

Personalization tiers (use the highest tier you can afford for this campaign):

• Tier 1 (mass / low-effort) — first name + role + company + industry. Acceptable only when the value prop is sharp enough to carry the email on its own. Reply rates: low.
• Tier 2 (signal-based) — the prospect is in a role/stage where the problem you solve is acute (e.g., a new Head of Growth in their first 60 days). Reply rates: meaningfully better.
• Tier 3 (observation-based) — references something from the company site, pricing page, careers page, or a recent product launch. This is the sweet spot.
• Tier 4 (deep) — references a recent LinkedIn post, blog post they wrote, or talk they gave. Reserve for high-value targets.

Anything below Tier 2 should be treated with suspicion — {{FirstName}} swaps don't count as personalization.

Phase 4 — Draft emails (drafts-first by default)

Pick a framework that matches the situation. The four shapes that consistently work:

• Observation → Problem → Proof → Ask — "You're hiring SDRs. That usually means meetings/SDR ratio is the bottleneck. We helped [company] hit X. Worth exploring?"
• Question → Value → Ask — "Struggling with [problem]? We do [Y]. [Company Z] saw [result]. Worth a look?"
• Trigger → Insight → Ask — "Congrats on [funding/launch]. That usually creates [Y challenge]. We've helped similar teams with that. Curious?"
• Story → Bridge → Ask — "[Similar company] had [problem]. They [solved it this way with us]. Relevant to you?"

See references/frameworks.md for full examples and when each shape works best.

Subject lines. Short, lowercase, internal-looking. 2–4 words. No emojis, no first names, no urgency tricks. Targets: looks-like-a-colleague-sent-it. Examples: quick question, reply rates, hiring ops, q3 forecast, for {{company}}. Avoid: Increase your revenue 10x!, John, are you free Thursday?, [URGENT] follow-up.

Voice rules.

• Write like a peer, not a vendor. Use contractions. Read it aloud — if it sounds like marketing copy, rewrite it.
• "You/your" should outnumber "I/we" by ≥2:1.
• Every sentence must move the reader toward replying. The best cold emails feel like they could have been shorter, not longer.
• Calibrate to seniority: C-suite → ultra-brief and peer-level. Mid-level → more specific value. Technical → precise, no fluff, respect their intelligence.

What to avoid (these are the AI-tells reviewers immediately spot):

• "I hope this email finds you well." / "I came across your profile."
• "leverage", "synergy", "best-in-class", "leading provider", "circle back"
• Feature dumps. One proof point beats ten features.
• HTML, images, multiple links.
• Fake Re: / Fwd: subject lines.
• Identical templates with only {{FirstName}} swapped.
• Asking for a 30-minute call on touch 1.

Drafts-first send pattern (default for any 4+ prospect campaign):

# 1. Create label for the campaign once — capture the returned id
label = gmail_labels_create(name="cold/q3-growth-leads")
campaign_label_id = label["id"]

# 2. Draft the first 1-3 prospects for user review
for p in prospects[:3]:
    gmail_drafts_create(
      to=p["email"],
      subject="quick question",
      body=render_email(framework="observation", prospect=p),
    )

# 3. Show drafts to user, await explicit approval

# 4. After approval, send and label each message
for p in prospects[3:]:
    result = gmail_messages_send(
      to=p["email"],
      subject="quick question",
      body=render_email(framework="observation", prospect=p),
    )
    gmail_labels_add(message_id=result["id"], label_ids=[campaign_label_id])

If the user wants every email reviewed, use gmail_drafts_create for all of them and send via gmail_drafts_send after approval. If the user is confident and the templates are pre-approved (e.g., they've run this campaign shape before), you can skip directly to gmail_messages_send from prospect 1. Default behavior is drafts-first.

Phase 5 — Run the follow-up cadence

3–5 total touches with widening gaps. Each follow-up adds something new — a different angle, fresh proof, a useful resource. "Just checking in" gives the reader no reason to respond.

Default cadence (adjust to the user's situation):

Always reply in the original thread (gmail_reply_to_message with the message_id returned from the touch-1 send) — preserves context and improves deliverability. See references/follow-up-sequences.md for angle rotation, breakup-email templates, and how to prune prospects mid-sequence.

Phase 6 — Track replies and route them

# Pull all replies on the campaign label from the last 7 days
gmail_messages_list(query="label:cold/q3-growth-leads is:unread newer_than:7d")

For each reply, read the body with gmail_get_message(message_id=...), classify it, and label:

Apply classification and clear unread with two separate calls:

gmail_labels_add(message_id=..., label_ids=[classification_label_id])
gmail_labels_remove(message_id=..., label_ids=["UNREAD"])

Create the sub-labels once with gmail_labels_create and capture their IDs before the routing loop.

Quality check (before any send)

Read each draft against this gut-check. Reject any that fail more than one:

• Does it sound like a human wrote it? (Read it aloud.)
• Would you reply if you got this?
• Does every sentence serve the reader, not the sender?
• Is the personalization connected to the problem you solve — not just a generic compliment?
• Is there one clear, low-friction ask?
• Does the subject line look like it came from a colleague?
• Is the email under ~120 words on touch 1?

Reference workflows

For long-form material — read on demand: