lead-scoring

---
name: lead-scoring
description: Rank leads by fit, intent and timing, every point sourced, hottest called first; use when leads outnumber time or someone asks who to call first.
---

# Know which leads to call first, and why

A sourced, tiered list, proven on your lost deals, Hot leads called within the hour.

## Claude does
1. Collect {LEADS}, {SIGNALS} (dated), {CHANNEL}, {TODAY}, and {ICP} or {DEALS_CSV} (last 20-50 closed-won plus last quarter's closed-lost). Ask once. No ICP and no deals, no scoring; never guess the customer.
2. No {ICP}? Derive one from {DEALS_CSV}: drop referral, M&A and odd-contract wins, keep attributes of wins not losses, sort into must-have, nice-to-have, disqualifier; one ICP per motion; under 20 wins the user writes it.
3. Write {SCORING_KEY} for CODEX.md's eleven-line rubric: full, partial, zero per line; the disqualifier list; which trigger events and urgency words count; no twelfth input until the back-test passes.
4. Rule now on ambiguous cases (unfamiliar sector, borderline title) and name strategic flags (former champion, sector influencer): one routing tier up, number unchanged.
5. Confirm {ROUTING}: Hot 75-100 called within 1 hour, Warm 50-74 within 4, Cold 25-49 within 24, Disqualified under 25 archived, and who owns Hot; no owner, no run.

## Then Codex does
Scores closed-lost deals, then every live lead, line by line against the key with source, quote and confidence per point. Bulk arithmetic, no judgement.

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

Fill first: {ICP}, {SCORING_KEY}, {LEADS}, {SIGNALS}, {DEALS_CSV}, {CHANNEL}, {TODAY}, {ROUTING}.

## Claude checks
1. Back-test: most closed-lost land below Hot, else tighten the key and rerun before live leads count.
2. Recompute the top ten sums, one wrong total fails; rows in equal rows out.
3. Disqualified rows stay Disqualified, naming why, whatever their intent; mostly Disqualified means a weak source.
4. Five random non-zero points show source, quote and confidence; one missing fails.
5. No points from opens, clicks or downloads alone; gaps scored 0, marked Inferred, never filled.

Failed check: correct CODEX.md, rerun Codex, never hand-patch.

## Rules
- Show the top ten's arithmetic so a human can disagree with it.
- High intent never rescues a disqualified account. State the disqualified count; it grades the list source.
- Keep fit and intent visible; one number hides a perfect fit with no buying motion. Never score engagement alone; curious subscribers are not buyers.
- Every point needs a source; missing data lowers confidence, never inflates. Size is not budget, a title is not a decision-maker, urgency needs a trigger.
- The ICP excludes outlier wins, contains disqualifiers, is who to hunt today not the whole market, and lies once a year old. Few inputs; no twenty-factor formula.
- Every tier has an action, response time and owner. Leads are not opportunities; never forecast from a tier.
- Re-score on triggers and a schedule; re-derive the ICP quarterly.

## If Codex is not installed
Claude does that half, says so, still checks.

## Built from
- sales-qualify (ai-sales-team-claude): https://github.com/zubair-trabzada/ai-sales-team-claude/blob/main/skills/sales-qualify/SKILL.md (1,093 stars)
- lead-researcher (borghei/Claude-Skills): https://github.com/borghei/Claude-Skills/blob/main/personal-productivity/lead-researcher/SKILL.md (702 stars)
- qualify-lead (lead-qualification-plugin): https://github.com/BayramAnnakov/lead-qualification-plugin/blob/main/skills/qualify-lead/SKILL.md (21 stars, MIT)
- Clay lead-scoring playbook: https://www.clay.com/guides/how-to-set-up-lead-scoring-in-hubspot (first-party, 30 May 2026)
- MadKudu Customer Fit docs: https://help.madkudu.com/docs/customer-fit-scoring (vendor documentation)
- Common Room on behavioural scoring: https://www.commonroom.io/blog/behavioral-lead-scoring/ (19 Jun 2026)

# Know which leads to call first - Codex task

## You are given
- {ICP}: must-have, nice-to-have, disqualifier attributes.
- {SCORING_KEY}: Claude's ruling per rubric line, disqualifier list, counted trigger events and urgency words, edge cases, strategic flags.
- {LEADS}: live leads, CSV or rows. {SIGNALS}: dated per-lead signals if separate.
- {DEALS_CSV}: closed deals with an outcome column.
- {CHANNEL}: outreach channel. {TODAY}: reference date. {ROUTING}: action, response time, owner per tier.

## Produce
Score every closed-lost {DEALS_CSV} row, then every {LEADS} row, applying the key exactly. A line no ruling covers scores 0 and flags the row Unclear; never invent a ruling.

Rubric (key overrides):
- Fit 0-40: size 0-10 (50-200 staff 10, 201-500 7, 20-49 5, 501-1000 3, else 0); industry 0-10; geography 0-10; tech stack 0-5; decision-maker 0-5.
- Intent 0-40: request type 0-15 (Demo 15, Pricing 12, Contact 10, Content 3); engagement signals 0-10; message specificity 0-10; content buying signals on {CHANNEL} 0-5.
- Timing 0-20: urgency 0-10 (timeline words); trigger events 0-10 (funding, leadership change, hiring surge, news, dated).

Every non-zero point cites a source (column or URL), a quote of at most 125 characters and a confidence label (High, Medium, Low, Inferred); no source, no points.

Check disqualifiers (personal email with no company, competitor employee, .edu, wrong geography, below minimum size, explicit non-commercial intent, plus the key's list) before the tier is final; any hit: Tier Disqualified, Fit 0, Reason naming it.

Total = Fit + Intent + Timing. Hot 75-100, Warm 50-74, Cold 25-49, Disqualified under 25. Routing = the {ROUTING} row; a strategic flag lifts it one tier, Total unchanged.

Files:
1. `backtest.md`: closed-lost rows with Total and Tier, then "N of M land below Hot".
2. `lead-scores.csv`: input columns plus size, industry, geography, tech_stack, decision_maker, request_type, engagement, specificity, content_signals, urgency, trigger_events, Fit, Intent, Timing, Total, Tier, Routing, Flag, Evidence (source, quote, confidence per line), Reason; Total descending, Disqualified rows kept.
3. `lead-scores.md`: Markdown table (Company, Contact, Fit, Intent, Timing, Total, Tier, Routing, Reason); "Arithmetic, top ten" with each sub-score, evidence and sum; "Disqualified: N of M" naming each; "Unclear" with what was missing.

## Rules
- Fit, Intent and Timing stay separate columns, combined only in Total.
- Age every signal from {TODAY}; undated signals score 0.
- Missing data scores 0, labelled Inferred; never guess.
- Opens, clicks and downloads earn nothing beyond Content 3. Size is not budget, a title alone is not a decision-maker, urgency needs timeline words or a trigger.
- Do not drop, merge or add rows.

## Return
Print: rows in and out per list; back-test line; counts per tier, Disqualified, Unclear; top ten with sub-scores and Total; the three file paths.