cold-email-sequence-generator-skill

Turn a qualified contact list plus a value proposition into a multi-touch cold email sequence whose output is a set of reviewable Gmail drafts, not chat text. Each email has A/B subject lines, one CTA, timing, and a breakup touch. Product facts come from the Epiphan catalog, voice from the Epiphan Brand MCP, and the recipient list from `qualify_lead`. Keywords: `cold email sequence`, `email campaign`, `outbound sequence`, `nurture`, `re-engagement`, `breakup email`

<quick_start> Tools in call order:

1. Inputs — contact list (HubSpot list id, CSV, or named contacts), vertical, sequence type, value prop, proof points. Missing list → ask; missing vertical → infer from domain, flag.
2. Variant — read config.json; if variants.enabled, run bash scripts/variant-assigner.sh cold-email-sequence-generator-skill $(date +%s) and apply the assigned prompt_suffix.
3. Gate Q — qualify_lead per contact. Excluded contacts drop; STALE AE LEAD rows are listed for Tim, not drafted; ATL tier from the verdict sets ordering.
4. Gate S — search_product_catalog then Epiphan Knowledge search for every product, spec, or competitor claim the sequence will carry. Unverified → [spec unverified] in the draft.
5. Gate B — get_writing_style once, then check_my_copy on every subject line and body.
6. Gate D — mcp__claude_ai_Gmail__create_draft per contact per touch, from [email protected], subject prefixed [Seq N/M]. Never send_message.
7. Deliver — a summary table (contact / tier / touches drafted / gate flags) plus the sidecar and run-log line below. Sending is batch-send-drafts-skill's job. </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>

<variant_support> This skill supports A/B variants via config.json.

1. Read this skill's config.json — check if variants.enabled is true
2. If enabled, run: bash scripts/variant-assigner.sh cold-email-sequence-generator-skill $(date +%s)
3. Apply the prompt_suffix from the assigned variant's definition to modify the workflow
4. Record the variant name in the outcome sidecar (see final stage below)

Current experiment: cold-email-subject-style-001

• control: Standard 7-email framework
• concise: 5-email framework, 50-word max per email </variant_support>

<success_criteria>

• qualify_lead ran on every contact before any draft; excluded and stale-AE rows are reported, not drafted
• Sequence type selected (Classic 7, Fast-Track 5, Long-Play 12-14, Event-Based, Re-Engagement)
• Every product or competitor claim traces to an Epiphan Knowledge search result or carries [spec unverified]
• Every subject line and body passed check_my_copy; corrections applied before drafting
• One Gmail draft per contact per touch exists, from [email protected], with [Seq N/M] subject prefix
• Each email has A/B subject lines, one CTA, send-day offset; final touch is a breakup
• Outcome sidecar and run-log line written with gate flags </success_criteria>

Stage 1: Inputs and candidate pull

Accept any of: a HubSpot list (hubspot_list_members), a HubSpot search (hubspot_search_contacts), a CSV, or named contacts. Capture per contact: email, first name, title, org, vertical, HubSpot id. Capture per sequence: type, value prop, proof points Tim wants used, personalization hooks.

Stage 2: Gate Q — qualify_lead

Call qualify_lead per contact. Apply the verdict as returned: exclude customers, channel partners, product-only engagers, suppressed, and AE-owned-active; surface STALE AE LEAD rows in the summary for Tim's decision. Keep the ATL/GRAY/BTL tier for ordering (ATL first). If qualify_lead is unavailable after one retry: halt, write the state, tell Tim — do not draft to an ungated list.

Stage 3: Sequence design

Type	Touches	Span	Use for
Classic	7	2 weeks	Net-new cold list
Fast-Track	5	1 week	Warm or event-adjacent list
Long-Play Nurture	12-14	4-6 weeks	Budget-cycle or RFP-timed accounts
Event/Trigger	3-5	event-specific	Extron SMP EOL, new building, failed recording, conference
Re-Engagement	5	2 weeks	Dead-deal or >90-day-stale contacts

Email roles (Classic 7):

#	Role	Goal	Length	CTA
1	Introduction	Problem reframe, no product name	50-100 words	Soft question
2	Value Proof	Verified proof point	75-125 words	Specific time
3	Different Angle	Alternative pain	50-75 words	Yes/no question
4	Social Proof	Peer validation (verified)	60-90 words	Simple reply
5	Resource Share	Give before asking	40-60 words	"Useful?"
6	Direct Ask	Straight meeting request	30-50 words	Day/time
7	Breakup	Permission to close	25-40 words	"Should I close your file?"

Timing: Day 0, 2, 4, 6, 8, 10, 14. Tue-Thu mornings (recipient timezone) outperform; avoid Monday inbox pile-up and Friday afternoons. Record the day offset in each draft's subject-line note so batch-send-drafts-skill can order them.

Stage 4: Gate S — spec and claim verification

Before writing a single body: search_product_catalog for every product named (confirms the SKU is current; EOL products get an acknowledgement, never a spec), then Epiphan Knowledge search for the exact spec, capability, or competitor fact. Proof points and case references must come from the knowledge result or from Tim's supplied inputs; never invent a customer, number, or quote. Unverified claim → write [spec unverified] inline and set outcome partial naming Stage 4.

Stage 5: Gate B — brand voice, then write

get_writing_style once per session. Write each touch from the templates below, then run check_my_copy on every subject line variant and every body; apply corrections before Stage 6. If Brand MCP is down after one retry: draft anyway, prefix each body [brand check skipped], outcome partial naming Stage 5.

Stage 6: Gate D — Gmail drafts

mcp__claude_ai_Gmail__create_draft per contact per touch, from [email protected]:

• Subject: [Seq N/M] <winning A/B subject> — the alternate subject goes in the first body line as A/B alt: … so Tim can swap it in review.
• Body: the touch as approved by check_my_copy, signature per identifiers.md.
• Never call send_message or reply. Gmail write failure → retry once, then report the touch as undrafted in the summary (outcome partial, Stage 6).

Stage 7: Summary

SEQUENCE STAGED — <type>, <M> touches, <date>
 Contact              Tier   Drafts   Flags
 J. Smith (u.edu)     ATL    7/7      —
 R. Lee (city.gov)    GRAY   7/7      [spec unverified] x1
 K. Ng (hosp.org)     —      0        STALE AE LEAD (Ron Epstein, 112d) — Tim decides
Excluded by qualify_lead: 4 (2 customer, 1 partner, 1 AE-active)
Next: batch-send-drafts-skill to review and send.

Email Templates

Brackets are filled from Stage 1 inputs and Stage 4 verified facts only.

Email #1: The Introduction

Subject Lines (A/B Test):

• A (Curiosity): Quick question about [org]'s [specific challenge]
• B (Value): [Verified outcome] for [org type]
• C (Personal): [Name], saw your post about [specific topic]

Hi [First Name],

I noticed [specific observation about their org/role] and thought you might be facing [specific challenge].

Teams like [verified peer org type] have [verified outcome] without [common objection].

Worth a quick 15-minute conversation to see if the same applies at [org]?

[Signature]

P.S. - [Personalized one-liner from research]

Email #2: The Value Proof

Subject: "How [peer org type] achieved [verified result]"

[First Name],

Following up — wanted to share how this worked for a [peer org type].

They were [situation]. In [timeframe], they:
- [Verified result #1]
- [Verified result #2]

They got started without [common objection].

Would [Day] at [Time] work for 15 minutes?

[Signature]

Email #3: The Different Angle

Subject: "Different thought about [org]"

Hi [First Name],

I realize [original pain point] might not be top of mind right now.

But what about [alternative pain point]?

Most [their role]s tell me [common complaint], which is why [mini value prop, no spec unless verified].

If this hits closer to home, happy to share how [quick win].

[Signature]

P.S. - If neither is relevant, just let me know and I'll stop.

Email #4: The Social Proof

Subject: "[Mutual connection] suggested I reach out" OR "How [peer] handles [challenge]"

[First Name],

I was speaking with [title] at a [peer org type] about [challenge].

Here's what they said after [solution]:
"[Quote supplied by Tim or from knowledge base — never invented]"

Open to a quick call to hear what's working in [vertical]?

[Signature]

Email #5: The Resource Share

Subject: "Thought you might find this useful"

[First Name],

No ask here — just sharing something helpful:

[Resource description]: [Link]

We put this together after hearing [their role]s struggle with [pain point]. Useful even if you never buy anything from us.

Hope it helps.

[Signature]

Email #6: The Direct Ask

Subject: "Let's cut to the chase"

[First Name],

Let me be direct: I think we could help [org] [achieve outcome] based on [observation].

I'd like to show you:
1. [Specific thing #1]
2. [Specific thing #2]
3. [How peers in your role use it]

15 minutes. No pressure. How's [day/time]?

[Signature]

Email #7: The Breakup

Subject: "Should I close your file?"

[First Name],

I'll assume [topic] isn't a priority right now — totally fine.

I'll close your file unless I hear otherwise. If you want to revisit, [time-sensitive reason, e.g. budget cycle], so a quick chat now might save you a step later.

No worries either way — appreciate your time.

[Signature]

P.S. - If someone else at [org] should hear about this, happy to redirect.

Breakup variations: Permission ("Not relevant / not a priority / bad timing — which? If #3, when should I check back?") · Referral ("Clearly I'm not reaching the right person — who owns [topic]?"). Avoid FOMO lines that name a competitor's customer unless Gate S verified it.

Vertical Quick-Starts

• Higher Ed (ICP 90): finals-week reliability → peer-campus proof → distribution pain → resource (CMS integration doc) → direct ask → breakup
• Courts (ICP 85): gap in the record → county proof → unattended 8-hour sessions → resource → direct ask → breakup
• Government (ICP 80): council capture without AV headcount → city proof → posting lag → resource → direct ask → breakup
• Healthcare (ICP 75): OR capture without clinical IT load → med-ed proof → learner access → resource → direct ask → breakup
• Corporate AV / Worship / K-12: same arc; swap the reframe for the vertical's top pain from Tim's inputs

A/B Testing and Benchmarks

Test in order: subject line (question vs statement, personalized vs generic), body length (50 vs 100 words), CTA style, send time. Method: 50/50 split on the first 100 contacts, 48-hour read, winner to the rest. Benchmarks (Email 1 open 35-55%, sequence reply 8-25%, meeting booked 1-6%) are planning ranges, not claims to put in copy.

Pre-Launch Checklist

• Gate Q report reviewed (excluded + stale-AE rows acknowledged by Tim)
• No [spec unverified] or [brand check skipped] flags left in drafts Tim intends to send
• Personalization variables all filled; no literal brackets remain
• Links tested; no attachments
• Daily send limit agreed with Tim before batch-send-drafts-skill runs

<self_healing> Failure = no drafts produced; partial = a stage degraded (name it: Q halt, S unverified, B skipped, D write failed). Ladder: retry once → degrade (draft with inline flag, keep core) → alert Tim (summary line, Slack DM if unattended) → halt when qualify_lead is unavailable for the list. Full contract: ~/.claude/skills/epiphan-house-rules-skill/reference/self-healing.md. </self_healing>

<outcome_sidecar> At the end of every run write ~/.claude/skill-analytics/last-outcome-cold-email-sequence-generator-skill.json and append one line to ~/.claude/skill-runs/cold-email-sequence-generator-skill.jsonl:

{"ts":"<UTC ISO8601>","skill":"cold-email-sequence-generator-skill","version":"<frontmatter version>",
 "variant":"<assigned variant or default>","status":"success|partial|error","runtime_ms":0,
 "metrics":{"contacts_in":0,"contacts_qualified":0,"contacts_excluded":0,"stale_ae":0,
            "emails_generated":0,"drafts_created":0,"spec_unverified":0,"brand_skipped":0},
 "error":null,"session_id":"<date>"}

status partial/error must name the failing stage in error. The SessionStart sweep ingests it. </outcome_sidecar>