Positive Reply Scoring

Positive Reply Scoring

Reply rate tells you if people are paying attention. Positive reply rate tells you if they want what you're selling. This skill computes the second.

Why this exists

A campaign can get 5% reply rate and still be a disaster. If 90% of those replies are "unsubscribe" and "not a fit," you're burning your domains for nothing.

The metric that matters is:

positive_reply_rate = positive_replies / total_sent

Compared side-by-side:

• Campaign A: 1% reply rate, 70% positive → 0.7% positive reply rate
• Campaign B: 5% reply rate, 10% positive → 0.5% positive reply rate
• Campaign A wins.

Classification schema

Every reply is classified into exactly one bucket:

| Label | Meaning | Count as "positive"? | |---|---|---| | positive_interested | "Yes, tell me more" or booked a meeting | ✅ | | positive_soft | "Send more info" / "reach out in Q3" / info request | ✅ | | positive_referral | "Not me, but talk to X" | ✅ (referral is high-value) | | neutral_question | Clarifying question, no commitment yet | ❌ (optional — some score as half) | | negative_notnow | "Not right now, maybe later" | ❌ | | negative_notfit | "Not a fit" / "we don't need this" | ❌ | | negative_hostile | Angry reply, complaint, report | ❌ (and track separately as risk signal) | | unsubscribe | Explicit opt-out | ❌ | | ooo | Out-of-office auto-reply | ❌ (exclude from denominators) | | bounce | Technical bounce | ❌ (exclude from denominators) | | other | Can't tell | ❌ |

Positive reply rate = (positiveinterested + positivesoft + positivereferral) / totalsent

Inputs

• Smartlead API key (env: SMARTLEAD_API_KEY)
• Campaign ID to score
• Optional: client_id (if using a sub-client setup)
• Optional: date range (defaults to full campaign)

Steps

1. Fetch all leads + replies from the campaign

Run the fetch script:

npx tsx scripts/fetch-campaign-replies.ts --campaign-id=12345 --out=/tmp/replies.json

This walks /campaigns/{id}/leads paginated, identifies leads with replies (has_reply = true), then fetches /campaigns/{id}/leads/{lead_id}/message-history for each, and writes them to a JSON file with one object per reply.

Output schema per reply:

{
  "lead_id": "...",
  "email": "...",
  "lead_first_name": "...",
  "company": "...",
  "reply_time": "ISO timestamp",
  "reply_subject": "...",
  "reply_body": "... full text ...",
  "sequence_step": 1
}

2. Classify replies in the Claude Code conversation

Once the JSON is written, Claude (the one running this skill) reads the file and classifies each reply. For speed, fan out in batches of 20-30 via the Task tool (see personalization-subagent-pattern skill for fan-out mechanics).

Classification prompt (per batch):

Classify each reply as one of:
- positive_interested, positive_soft, positive_referral
- neutral_question
- negative_notnow, negative_notfit, negative_hostile
- unsubscribe, ooo, bounce, other

For each reply, output: { lead_id, label, confidence: 0.0-1.0, one_line_reason }

Rules:
- OOO auto-replies ("I'm out of office") → ooo
- Bounces (delivery failure messages) → bounce
- "Take me off your list", "unsubscribe", "STOP" → unsubscribe
- "Not interested", "not a fit" → negative_notfit
- "Not right now, circle back in Q3" → negative_notnow
- "Try [other person]" → positive_referral
- "Send more info" or "Tell me more" → positive_soft
- "Yes, let's book a call", "what times work" → positive_interested
- Insults, reports, legal threats → negative_hostile

If confidence 0.3% or unsub >2% → deliverability risk, pause campaign

4. Save to disk

Write aggregate results to:

~/cold-email-ai-skills/profiles//scores/-.json

This builds a history so you can trend positive reply rate over campaigns.

5. Flag action items

At the end, surface:

• Positive replies that need a human response — list the top 10 positive_interested leads and their reply bodies. The user should reply to these within 30 seconds of seeing this report.
• Referrals that need follow-up — positive_referral labels. Add the referred contacts to a new outreach list.
• Hostile flags — any negative_hostile replies. Read them manually; consider pausing the inbox if someone is genuinely angry.
• Unsubscribes — confirm they're globally suppressed (Smartlead does this automatically, but double-check).

When to use this skill

• After a campaign has run for at least 14 days (otherwise sample is too small)
• When comparing two campaigns in an experiment (use the same cutoff date for both)
• Weekly as a quality check on running campaigns
• Before deciding to kill or scale a campaign

Common gotchas

• Don't trust reply rate alone. A 5% reply rate from spam-trap replies and unsubscribes is worse than a 2% reply rate from real buyers.
• Exclude OOO + bounce from denominators. They're not real replies. The script does this automatically.
• Smartlead's built-in AI categorization exists but is less controllable. This skill uses Claude directly for transparency and prompt-tunable classification.
• Small samples lie. Below ~500 sent, the positive reply rate has too much noise. Wait for more volume before declaring winners/losers.
• Classify only FIRST reply per lead. If a lead replied, you replied, they replied again — only the first reply is the signal. Later messages are the conversation, not the scoring.

What to do next

Respond to every positive_interested reply within 30 seconds of seeing it. Then /experiment-design to plan the next iteration based on what worked.

If positive reply rate is <1% after 200+ sends: the 1% rule failed. Run /email-deliverability-audit (are you reaching the inbox?) (check for vague CTAs, generic first lines, em dashes).

Or wait: this skill is the Wednesday task in /cold-email-weekly-rhythm. Run it weekly going forward.

Related skills

• /experiment-design — uses positive reply rate as the success metric
• /email-deliverability-audit — if hostile + unsub are elevated, run this next
• /cold-email-starter-kit → 10-reply-handling.md for what to do with the positive replies once flagged

Scripts

• scripts/fetch-campaign-replies.ts — pulls replies via Smartlead API
• scripts/aggregate-scores.ts — computes rates from classified JSON

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: growthenginenowoslawski
• Source: growthenginenowoslawski/coldoutboundskills
• License: MIT

Install and usage instructions live in the source repository linked above.