Lead Scoring

Lead scoring

Evaluate accounts through three layers — fit (structural), signals (temporal), interpretation (synthesized) — to produce a routing recommendation per account. NOT a composite numeric score. Compressing fit, timing, and context into one number destroys the information operators need to act.

When to run

• User asks "score this lead", "is X a good fit", "should we pursue X", "prioritize these accounts", "rank these prospects"
• ABM campaign needs account tiering input
• Pre-discovery account brief, sales pipeline qualification gate

Skip when: user wants full research without assessment angle (/company-context), ICP definition (/icp-research), ABM tactics on already-tiered accounts (/abm-campaign), outreach copy (/outreach-emails).

Inputs

Required: at least one company identifier (URL, LinkedIn URL, or name).

Recommended (lift quality): client ICP doc (/icp-research), CRM engagement history, competitor list, prior /company-context output.

Mode detection: single account → deep assessment. Batch (5+ accounts) → lightweight pass + priority matrix. >15 accounts → calibration round + parallel waves (see the premium reference).

If ICP doc missing: proceed with generic B2B SaaS criteria, flag as "generic ICP" in output, suggest /icp-research upstream.

Steps

1. Validate input — confirm company identifier(s), determine mode (single/batch), identify ICP reference.
2. Fit assessment (Phase 1) — score firmographic, technographic, use case, negative-fit dimensions per the premium reference. Output verdict: STRONG_FIT | MODERATE_FIT | WEAK_FIT | NO_FIT with evidence + confidence per dimension.
3. Signal detection (Phase 2) — catalog leadership, growth, intent, operational, engagement signals per the premium reference. Tag each with category, recency (strong/moderate/weak/expired per decay table), source URL, confidence level.
4. Apply recency decay — drop expired signals from active inventory; weak-recency signals provide background only, don't drive routing. Decay table in the premium reference.
5. Interpret signal clusters (Phase 3) — identify reinforcement, contradictions, dominant story. Write 2-4 sentence situation hypothesis: "Based on [cluster], [company] appears to be [situation]. This suggests [implication]. The window is [timeframe] because [decay reasoning]."
6. Confidence assessment — rate HIGH (dense + fresh + diverse) | MODERATE (2 of 3) | LOW (sparse or stale).
7. Routing recommendation (Phase 4) — apply fit × signals matrix in the premium reference. Output: SALES | MARKETING | MONITOR | EVALUATE | DEPRIORITIZE | DISQUALIFY + 2-3 sentence rationale + 1-3 specific next actions.
8. Optional — activation score — if client wants auditable math: signal_activation = strength × recency × fit × tier_weight, sum top-3 per account, bucket into Hot/Warm/Nurture/Cold. Formula in the premium reference.
9. Optional — tier mode (numeric) — if client CRM needs a lead_score field or sales ops wants a single-column sort: compute weighted tier_score (0-5) and bucket Tier 1 / 2 / 3 / Disqualify. Formula + alignment-with-routing check in the premium reference.
10. Self-evaluation gate — every signal has source + recency tag, fit dimensions have evidence (not assumption), interpretation reads as narrative not list, routing follows fit×signals matrix, gaps marked [UNAVAILABLE], confidence levels per ontology.
11. Format output — single account: full template in the premium reference. Batch: priority matrix template.
12. Review gate (Level 1) — present fit verdict, signal summary, situation hypothesis, routing recommendation. Actions: [Approve] [Challenge fit] [Add signals] [Change routing].
13. Suggest chain — if SALES routing → /outreach-emails. If batch → /abm-campaign. If fit uncertain → /company-context. If no ICP → /icp-research.

Scoring validity — before fit-rules count as predictive

When the fit rubric or ICP rules are derived from a client's own customers (won deals, a "good-fit" list), they fit those examples by construction and routinely fail on the wider universe. Before treating derived rules as predictive — or locking a numeric tier/activation model (steps 8–9) tuned on a small known-positive set — clear the pre-lock gate in .claude/rules/scoring-validity.md: ground-truth provenance, base rate + discriminative ratio (≥2.0), holdout, backwards-reasoning check. Below the sample floor (our client set is often N≈8), ship a directional hypothesis, not a locked rule. Scoring a single account against an already-validated rubric doesn't trigger this — deriving the rubric from outcomes does.

What good looks like

• projects/research/taste-library/resources/0626-sales-qualification-frameworks/health-rubrics.md — deal-health (10-dim) + account-health (9-dim) rubrics; bolt onto fit+signal scoring when the account is an open opportunity, not just a prospect (re-weight per client motion)

Examples: none baked into skill (every assessment is account-specific). Pull patterns from projects/consulting/{client}/sales/ lead-assessment outputs when present.

Evaluations — output passes if:

• Fit verdict cites evidence per dimension (not asserted)
• Every signal tagged with category + recency + source URL + confidence
• Situation hypothesis reads as 2-4 sentence narrative, not a list
• Routing follows fit×signals matrix (divergence flagged with rationale)
• No invented data; gaps marked [UNAVAILABLE]
• If tier mode active: tier and routing align (or divergence has 1-sentence rationale)
• Decay applied: weak-recency signals don't drive routing, expired excluded

Anti-patterns (auto-fail):

• Compressing into single composite score without preserving fit/signal/interpretation layers
• Binary STRONG_FIT or NO_FIT verdicts (full 4-tier range required)
• Score with no routing recommendation
• Months-old signals treated as fresh
• Employee count as primary fit indicator (revenue model, tech stack, growth trajectory often matter more)

Final ship gate

Run /premortem --output before ship. See /premortem skill for the 5 execution domains (will-it-resonate / will-it-convert / will-it-stay-on-brand / will-stakeholder-push-back / will-it-degrade-over-time) and output template.

Trivial-case escape: ## Premortem\nNo failure modes — trivial change satisfies the contract for genuinely trivial outputs.