personalization-at-scale-skill

Personalization at Scale

Generate hundreds of unique, researched first lines in minutes instead of hours. Takes a prospect list and finds personalization angles from company news, LinkedIn activity, funding rounds, hiring signals, and mutual connections to make cold outreach feel warm.

<quick_start> Trigger: "Personalize outreach for [N] prospects" or "Generate unique first lines for my prospect list" Input: CSV or list with First Name, Last Name, Title, Company, LinkedIn URL (email if available) Output: 2-3 first-line options per prospect with confidence scores, delivered as a review CSV at ~/.claude/skill-state/personalization-at-scale/<date>.csv (default) or as Gmail drafts when Tim asks for drafts — never sent.

Tool order (batch mechanics in Stage R):

1. qualify_lead (Epiphan AI) per contact — Gate Q on the whole list before any research.
2. apollo_organizations_enrich + apollo_organizations_job_postings per unique company (hiring + firmographics); Clay find-and-enrich-company for news / tech-stack gaps; apollo_people_match for role changes; intent-signal-aggregator-skill when wired.
3. search_product_catalog once per use case (Gate S) if any line carries an Epiphan capability claim.
4. get_writing_style once, then check_my_copy over the lines in batches (Gate B).
5. Deliver: CSV for review, or Gmail create_draft per prospect (Gate D). </quick_start>

<house_rules> Customer-facing skill. Before output, run the gates in order (details: ~/.claude/skills/epiphan-house-rules-skill/reference/gates.md): S: search_product_catalog (Epiphan AI) or Epiphan Knowledge search before any product claim. B: check_my_copy on every customer-facing string. Q: qualify_lead right after the candidate pull, before tiering/enrichment/drafts. D: outbound text becomes a Gmail draft, never a send. Identifiers (owner IDs, portal, emails, seats) come from reference/identifiers.md in that skill — do not restate them here. On any gate failure follow reference/self-healing.md (retry, degrade with a flag, alert Tim, halt) and append a line to ~/.claude/skill-runs/.jsonl. </house_rules>

<success_criteria>

• 70%+ prospects have unique, specific personalization found
• Each first line is recent (within 30-60 days), role-relevant, and couldn't be copy/pasted to another prospect
• Confidence scores assigned (High/Medium/Low) to every first line
• Fallback strategies provided for prospects with no personalization found
• Every input contact passed through qualify_lead before research; excluded / BTL / NEVER counts reported
• Any Epiphan capability claim in a line comes from Gate S; every line passed check_my_copy before delivery
• Output delivered as a review CSV (or Gmail drafts, never sent) at a named path
• Outcome sidecar + run-log line written </success_criteria>

Stage Q — qualify_lead gate (replaces the former inline Golden Rules and suppression stages)

Before researching or personalizing ANY contact, run mcp__claude_ai_Epiphan_Ai__qualify_lead (email, else name) on every row, writeToHubSpot: false. The verdict is authoritative for customer / device owner, channel partner, product-only engager, AE-owned active vs STALE AE LEAD, suppressed, geo (USA/Canada default), junk, and the ATL / GRAY / BTL / NEVER tier. Rules and owner IDs: ~/.claude/skills/epiphan-house-rules-skill/reference/gates.md and reference/identifiers.md.

• keep + ATL / GRAY → personalize
• keep + BTL → personalize only if Tim asked for the full list; otherwise list as "deprioritized"
• NEVER / exclude → no research spend; STALE AE LEAD → listed for Tim, not personalized

Report: [N] removed by Gate Q (X customers, Y channel, Z AE-active, W suppressed, V out of geo, U BTL/NEVER). qualify_lead unavailable after one retry → halt before research, tell Tim, status error (stage: gate_q).

---

Stage R — Research (wired tools, batched)

Source	Tool	Batch mechanics
Hiring signals, firmographics	apollo_organizations_enrich, apollo_organizations_job_postings (Apollo plugin)	Once per unique domain, not per contact; cache by domain for the run
Company news, funding, tech stack changes	Clay find-and-enrich-company (data points: Latest Funding, Tech Stack, News)	Only for domains where Apollo returned no signal; Clay credits
Role change / promotion, tenure	apollo_people_match	Per contact; reveal_personal_emails: false
Prior Epiphan touches, reference accounts	search_account_text, search_companies_by_description (Epiphan AI)	Per domain; Epiphan-native before any web search
Intent / page visits	intent-signal-aggregator-skill when wired	Per domain
LinkedIn posts, podcasts, conference talks	Web search	Gaps only, after the tools above; skip if nothing recent (30-60 days)

Pick the best angle per contact from these results; a line with no tool-sourced signal gets a Fallback Strategy (below) and a Low confidence score.

---

Stage G — Gate S + Gate B, then delivery

1. Gate S: if any line names an Epiphan product or capability, search_product_catalog once per use case and quote it; documentationGap → remove the claim or write [spec unverified], status partial (stage: spec).
2. Gate B: get_writing_style once per run; then check_my_copy over the generated lines in batches of ~40 lines (stay under the 20,000-character limit); apply every error before delivery. Brand MCP down → prefix the file/drafts with [brand check skipped], partial (stage: brand).
3. Delivery (Gate D): write the review CSV to ~/.claude/skill-state/personalization-at-scale/<YYYY-MM-DD>.csv (columns: first_name, last_name, email, company, tier, angle_type, source, source_date, line_1, line_2, line_3, confidence). When Tim asks for drafts, Gmail create_draft per prospect from [email protected]; never send_message. Tell Tim the path and the count; lines enter sequence position 1 only after his review.

---

Instructions

You are an expert sales development researcher who specializes in finding personalization angles for outbound prospecting at scale.

Research Sources

Company news, hiring patterns, funding, tech-stack changes, role changes, LinkedIn activity, conference / podcast appearances, thought leadership, mutual connections — each wired to a concrete tool in Stage R above. Do not describe a source you did not query.

Personalization Styles

1. Congratulations - Recent achievement or announcement
2. Observation - Noticed something specific about their company/role
3. Shared Interest - Common connection, interest, or experience
4. Insight - Industry trend relevant to their situation
5. Question - Ask about their approach to a challenge
6. Compliment - Genuine praise for their work/content
7. Problem Call-Out - Identify a pain point they're likely experiencing

Quality Standards

Good Personalization:

• Specific and unique to them (couldn't copy/paste to anyone else)
• Recent (within last 30-60 days ideally)
• Relevant to their role or business
• Natural and conversational (not creepy-stalker)
• Easy to verify (they can remember this happening)

Avoid:

• Generic compliments ("I love your company!")
• Fake personalization ("I was on your website...")
• Stale information (from 6+ months ago)
• Information they'd be uncomfortable you know
• Obvious automation ("I saw your recent LinkedIn post" x 100)

Output Format

# Personalization at Scale: [Campaign Name]

**Campaign**: [Campaign name/description]
**Prospect Count**: [Number]
**Target Persona**: [Job title/role]
**Industry**: [Industry or vertical]
**Research Date**: [Date]
**Personalization Success Rate**: [X]% (prospects with unique personalization found)

---

## Campaign Summary

**Personalization Breakdown**:
- [X] prospects: Company news/press mention
- [X] prospects: Recent LinkedIn activity
- [X] prospects: Funding or growth signals
- [X] prospects: Mutual connections
- [X] prospects: Hiring/tech stack signals
- [X] prospects: Recent job change
- [X] prospects: Content/thought leadership
- [X] prospects: No personalization found (fallback needed)

**Time Saved**: Manual ~5 min/prospect vs AI ~10 sec/prospect = [X] hours saved

---

## Personalized First Lines

### Prospect #1: [Name]

**Details**: [First Last] | [Title] | [Company] | [LinkedIn URL]

**Personalization Found**:
- **Type**: [Congratulations/Observation/Shared/etc.]
- **Source**: [LinkedIn post / Company news / Funding round / etc.]
- **Date**: [When this happened]
- **Context**: [Brief description of what you found]

**Option 1 (Direct)**:
> "Hi [First Name], congrats on [specific achievement]! I noticed [additional observation]. [Transition to value prop]"

**Option 2 (Question)**:
> "[First Name], I saw [specific thing]. Curious - are you [question related to their situation]?"

**Option 3 (Insight)**:
> "Hi [First Name], given [their situation/news], I imagine [relevant challenge]. [Transition to value prop]"

**Confidence Score**: [High/Medium/Low]
- High: Recent, specific, highly relevant
- Medium: Relevant but older, or less specific
- Low: Generic personalization, may not resonate

---

### Prospect #2: [Name]

[Repeat structure for each prospect]

---

## Personalization by Type

### Congratulations
Prospects with recent achievements, funding, promotions, or launches. First line pattern:
> "Congrats on [specific event]! With that kind of [growth/change], [likely pain point you solve]..."

### Observations
Prospects who posted content, made comments, or showed LinkedIn activity. First line pattern:
> "Loved your take on [topic]. The point about [specific thing] really resonated - we see that with [similar companies]..."

### Mutual Connections
Prospects with 1st or 2nd degree connections you can reference. First line pattern:
> "Hi [Name], I noticed we're both connected with [Mutual Connection]. [Context]. Thought I should reach out about [topic]..."

### Company News
Companies with recent press mentions, launches, or announcements. First line pattern:
> "[Name], saw [Company] is [news event]. That kind of [change] usually creates [specific challenge you solve]..."

### Hiring Signals
Companies with job postings indicating growth, tech changes, or priorities. First line pattern:
> "Noticed you're hiring [X+ roles]. Scaling that fast usually creates [specific problem you solve]..."

### Thought Leadership
Prospects on podcasts, webinars, published blogs, or conference speaking. First line pattern:
> "Really enjoyed your [content type] on [topic]. Your point about [specific insight] was spot-on..."

---

## No Personalization Found — Fallback Strategies

**Role-Based**: "Hi [Name], most [job titles] I talk to are dealing with [common pain point]. Is that on your radar?"

**Company-Stage**: "Hi [Name], companies at [their stage/size] typically face [challenge]. How are you handling [specific aspect]?"

**Industry**: "Hi [Name], with [industry trend], I imagine [company] is thinking about [related topic]..."

**Competitor Reference**: "Hi [Name], we work with [competitor 1], [competitor 2], and [competitor 3] to solve [problem]. Worth a conversation?"

---

Usage Instructions

Step 1: Upload Prospect List

Provide a CSV or list with at least:

• First Name, Last Name, Job Title, Company Name
• LinkedIn URL (if available), Email (if available)

Optional: Company website, Industry, Company size, Location

Step 2: Specify Preferences

Personalization Style (pick 1-3): Congratulations | Observations | Mutual connections | Company news | Hiring signals | Thought leadership

Tone: Professional | Casual | Direct | Consultative

Avoid: Anything older than [X] days | Personal information | Sensitive topics

Step 3: Review & Customize

• Review first 10 personalizations and adjust tone if needed
• Flag any that feel "off"
• Add company-specific context and modify CTAs

Step 4: Export & Use

Formats: CSV with personalization columns | Merge fields for Outreach/Salesloft | Individual email drafts

Workflow: Generate → Upload as custom fields → Use in sequence position 1 → Track response rates by type → Double down on what works

---

Performance Benchmarks

Metric	Generic Cold Email	With Personalization
Response Rate	1-3%	8-15%
Lift	Baseline	5-10x improvement

Time: Manual 5-10 min/prospect vs AI 10-30 sec/prospect = 8-16 hours saved per 100 prospects

Quality Threshold: Aim for 70%+ with unique personalization. Below 50% = consider different prospect list.

---

Best Practices

1. Mix Personalization Types: Don't just use LinkedIn posts for everyone
2. Keep It Natural: Should sound like you'd say it in person
3. Update Regularly: Refresh every 30 days as news/activity changes
4. Track What Works: Note which types get best response by persona
5. Quality Over Quantity: 100 well-personalized > 500 generic
6. Don't Be Creepy: If it feels stalker-ish, skip it
7. Don't Fake It: "I was on your website" when you clearly weren't
8. Always Verify: Spot-check first 10 personalizations manually

Common Use Cases

Trigger Phrases:

• "Personalize outreach for 300 prospects"
• "Generate unique first lines for my prospect list"
• "Find personalization angles for these LinkedIn profiles"
• "Research these 500 companies and prospects"

Response Approach:

1. Ingest prospect list (CSV or manual input)
2. Research each prospect across multiple sources
3. Identify best personalization angle per prospect
4. Generate 2-3 first line options per prospect
5. Provide confidence scores and fallback options
6. Export in requested format

Remember: Good personalization should feel like you actually researched them, because you (or AI) did!

<self_healing> partial = lines were delivered but a stage degraded: an Apollo or Clay call failed after retry (continue with the other sources, more Fallback lines), Gate S returned documentationGap ([spec unverified]), Brand MCP down ([brand check skipped]), or Gmail quota stopped drafting (remaining rows stay in the CSV). error = the input list could not be parsed, or qualify_lead was unavailable (halt before any research). Ladder per ~/.claude/skills/epiphan-house-rules-skill/reference/self-healing.md: retry once → degrade with a flag → tell Tim → halt. One JSON line per run to ~/.claude/skill-runs/personalization-at-scale-skill.jsonl. </self_healing>

Emit Outcome Sidecar

As the final step, write to ~/.claude/skill-analytics/last-outcome-personalization-at-scale-skill.json:

{"ts":"[UTC ISO8601]","skill":"personalization-at-scale-skill","version":"<frontmatter version>","variant":"default",
 "status":"[success|partial|error]","runtime_ms":[estimated ms from start],
 "metrics":{"prospects_in":[n],"excluded_by_qualify":[n],"prospects_personalized":[n],"first_lines_generated":[n],"avg_confidence_pct":[n],"sources_used":[n],"brand_errors_fixed":[n],"spec_unverified":[n]},
 "error":null,"session_id":"[YYYY-MM-DD]"}

Use status "partial" if some stages failed but results were produced; error must name the stage. Use "error" only if no output was generated.