Recursive 10-expert-panel cold-email optimizer

A two-mode Claude Code skill — audit existing campaign or start from scratch — that designs and tunes cold-outbound sequences for the Instantly sending platform. The signature move is recursive expert-panel scoring: a configurable panel of 10 expert personas scores each sequence draft 0–100 across explicit lenses (subject curiosity, first-sentence pattern interrupt, body brevity, CTA softness, deliverability risk, personalization believability, sequence flow). Each round produces a 10-row scorecard, an aggregate score, a ranked top-3 weaknesses list, specific copy edits, and an updated draft. Rounds iterate until the panel aggregate hits 90/100. Paired with a Python audit script that pulls live Instantly state — campaign health, warmup scores per sending account, domain inventory, capacity math.

The default failure mode of LLM copy review is unearned 9/10 — models flatter their own drafts and converge on "looks fine." This skill encodes an explicit guardrail against that pattern ("Scores must be brutally honest. No padding to 90 without earning it") and forces a multi-round score-edit-rescore loop. Output is a strategy doc with capacity math wired to real warmup data: accounts with score <80 or <14 days warmup are excluded from "ready" count, dramatically deflating projected pipeline for under-warmed setups — which is the right number, not the encouraging one. Concrete weekly targets ship with the skill (open 40–60%, reply 3–7%, positive reply 1–3%, meeting 0.5–1.5%).

1. Triage mode. Ask: (a) existing Instantly account, or start from scratch? (b) is the Instantly v2 API key available? If yes → audit mode; if no → copy work proceeds but infrastructure section is skipped.
2. Infrastructure audit (audit mode). Run python3 scripts/instantly-audit.py --api-key <KEY>. Report active campaigns (name / status / reply rate / open rate), sending accounts (count / warmup score / daily limit), domain inventory, and warmup gaps. Flag any account with warmup <80 or <14 days warmup as NOT ready.
3. Define ICP and business context. Fill the ICP template (titles, industries, company size, revenue floor, anti-ICP) and a one-sentence what-you-sell, primary offer, real URLs, proof points.
4. Configure the expert panel. Default roster is 10 personas defined in references/expert-panel.md; confirm or edit before scoring starts.
5. Run the recursive scoring loop. Each round emits a score table (10 panelists × 0–100 + rationale), aggregate score, ranked top-3 weaknesses, specific copy edits, updated copy. If aggregate <90 → next round addresses the top 3 weaknesses. If ≥90 → finalize. Hard rule: panelists score brutally; no aggregate may pad above weaknesses.
6. Compute send capacity. Use the verbatim chain: accounts-ready × 30 (conservative) or × 50 (aggressive) emails/day × 22 working days × reply rate × qualification rate = pipeline opportunities.
7. Deliver the strategy doc. 10 sections — Brutal Truth, ICP Summary, Infrastructure Status, Scoring Rounds, Final Email Copy, Implementation Plan, Capacity Math, Weekly Metrics Targets, STOP List, START List.
8. Human review gate before sending. The skill's hard rule is "Do NOT push anything to Instantly automatically" — the strategy doc is for review; explicit human approval is required before any API write to Instantly.

Infrastructure audit command (verbatim):

python3 scripts/instantly-audit.py --api-key <KEY>

Capacity math chain (verbatim from skill):

Accounts ready (score ≥80, ≥14 days warmup) × 30 emails/day = conservative daily volume
Accounts ready × 50 emails/day = aggressive daily volume
Daily volume × 22 working days = monthly send capacity
Monthly sends × expected reply rate = expected replies
Expected replies × qualification rate = pipeline opportunities

Weekly targets (verbatim):

Open rate: 40%+ good / 60%+ great
Reply rate: 3%+ good / 7%+ great
Positive reply rate: 1%+ good / 3%+ great
Meeting rate: 0.5%+ good / 1.5%+ great

Panel scoring lenses (verbatim):

- Subject line curiosity / open rate potential
- First sentence pattern interrupt
- Body clarity and brevity
- CTA softness and specificity
- Sequence flow and follow-up logic
- Deliverability risk signals (spam words, link density)
- Personalization believability

Install (verbatim — keep author path runnable):

git clone https://github.com/ericosiu/ai-marketing-skills.git \
 && cd ai-marketing-skills/outbound-engine \
 && pip install -r requirements.txt \
 && cp SKILL.md <your-project>/.claude/skills/cold-outbound-optimizer.md

• Unearned 90s. Without the "brutally honest, no padding" guardrail, panels converge on flattery. Keep that line in the skill prompt verbatim.
• Auto-push to the sender. Never wire scoring output directly into Instantly campaign creation. Human approval gate is load-bearing — a panel-approved sequence is still a draft until a human reads it.
• Under-warmed accounts inflating capacity. Operators routinely include accounts with warmup <80 in capacity math; the skill correctly excludes them. If your projected pipeline halves after audit, that's the math, not a bug.
• No API key, no infrastructure picture. Audit mode is the demo-able half; copy mode alone still ships a strategy doc but you lose the capacity grounding.