score-accounts

What this skill does

# Score Accounts

Rank a list of accounts by ICP fit + intent + trigger signals. Calls `get_gtm_context(detailed: true)` unconditionally, resolves mixed-identifier inputs explicitly surfacing ambiguity, scores each account on four axes, and presents both the ranking and the weight set as iteratively-refinable artifacts.

## The bar

1. **Resolution accuracy 100%** — every input auto-resolved / verified / ambiguous / failed. Nothing silently picked.
2. **Every score explainable** — composite is a transparent weighted sum, never an opaque number.
3. **"Why now" cites a specific signal** — not the composite restated.
4. **Every tier comes with a recommended action.**
5. **Weights and axes are exposed and overridable.**

Sellers reject black-box scores. Transparency + per-account "why now" are what make this skill trusted.

## Scope

Scores **company-level accounts**, not contacts. Persona-aware ranking is a chain target via `personalize-email` after tier-A is produced.

## Input

- **Accounts (required)** — list of ZI IDs / company names / domains / mixed CSV.
- **Use case (default `prospecting`)** — `prospecting`, `abm`, `territory_planning`, `pipeline_acceleration`. Affects tier thresholds + recommended actions.
- **Weight overrides (optional)** — `{fit, intent, trigger, engagement}` summing to 100.
- **Tier thresholds (optional)** — `{A, B}`. C is the remainder.
- **ICP override (optional)** — natural-language refinement on top of `get_gtm_context.icp`.
- **Intent topics (optional)** — explicit list overriding GTM-derived defaults.

## Four-axis framework

| Axis | Question | Source |
|---|---|---|
| **Fit** | Does this match our ICP? | `enrich_companies` vs `get_gtm_context.icp` |
| **Intent** | Are they actively researching the topic? | `enrich_intent` with curated topics |
| **Trigger** | Fresh event creating a window? | `enrich_news` + `enrich_scoops` (last 90d) |
| **Engagement** | Already interacting with us? | `account_research` narrative for known accounts. If absent, weight redistributed. |

Each axis 0–100 independently. Composite is the weighted sum — never collapsed to an opaque number.

## Default weights

```
fit:        45%
intent:     25%
trigger:    25%
engagement:  5%   (redistributed if unavailable)
```

User overrides accepted. Weights are exposed in every output. Cache per-axis scores; recompute only the composite when weights change.

## Tier thresholds

| Tier | Composite | Recommended action |
|---|---|---|
| **A** | ≥ 75 | Route to AE for 1:1 outreach within 24h. Chain to `personalize-email`. |
| **B** | 50–74 | SDR sequence; ABM retargeting; nurture-to-meeting. |
| **C** | < 50 | Watchlist; monitor for tier-promotion signals. |

Use-case adjustments: `abm` → A=80/B=55 · `territory_planning` keeps defaults · `pipeline_acceleration` → A=65/B=40.

## Workflow

### 1. Pull GTM context (always)
`get_gtm_context(detailed: true)`. Capture ICP, personas, competitors, offerings, strategic priorities. ICP = fit-axis target; strategic priorities → intent-topic curation.

### 2. Honor input data first
Use user-supplied weights / thresholds / ICP refinements / intent topics. Fall back to GTM defaults only for missing fields.

### 3. Resolve identifiers (four-bucket routing)

- **Auto-resolved** — top match dwarfs alternatives. Score without confirmation.
- **Verified** — clear top match BUT plausible alternatives exist. Score; surface verification note.
- **Ambiguous** — no dominant match. Pause scoring; surface candidates.
- **Failed** — no match. List separately.

Routing by input type:
- **Numeric ZI ID** → auto-resolved.
- **Domain** (`.com` / `.io` / `.co` / `.ai`) → `search_companies(companyWebsite)`. Single match → auto-resolved. Multiple → ambiguous.
- **Name** → `search_companies(companyName)`.
  - Top match's size/revenue dwarfs alternatives → auto-resolved.
  - Clear top but 3+ plausible alternatives → verified with note.
  - No dominant match → ambiguous.
- **No match** → failed.

**Surface rule.** Never silently pick a winner. Present top 5 with attributes; ask user to confirm. Use GTM context as soft tiebreaker for `verified` (e.g., a B2B SaaS context defaults an ambiguous name to the SaaS-industry candidate over an unrelated-industry candidate, with a flag).

**Domain-confirmation gate (mandatory for high-collision names).** When `search_companies(companyName=X)` returns >100 matches AND no strong GTM tiebreaker exists, require domain confirmation. Surface top match's domain and ask. Never silently auto-pick — cost of getting it wrong is scoring the wrong company entirely.

**Duplicate-record detection (mandatory).** If top candidates share the same domain root (e.g., `acmeco.com` and `acmecoinc.com`) AND ≤20% revenue diff AND same metro/country → flag suspected duplicate. Surface both records and offer to union. For signal-heavy workflows, scoring both and unioning is the right default — signals may be split across records.

Resolution path must hit 100% accuracy. Score auto-resolved + verified immediately; pause ambiguous; list failed separately.

### 3.5. Relationship-context pre-flight (mandatory)

Tag each resolved account against GTM context. Tag visible on the row before the tier letter — sellers see relationship status BEFORE running the play.

- **`competitor`** ⚔️ — in `get_gtm_context.competitors`. Don't exclude from ranking (competitive intel matters) but make it impossible to miss visually.
- **`customer`** 🤝 — in `get_gtm_context.customers` / `proof_bank`. Shift recommended action to expansion / renewal.
- **`partner`** 🔗 — in `get_gtm_context.partners` / `integration_partners`. Shift to co-sell / integration angle.
- **`prospect`** — default; no tag.

In the row label: `⚔️ [Account] (B 62)`. Skill never silently produces "pursue this competitor" rankings.

### 4. Curate intent topics
From `get_gtm_context.strategicPriorities` (or user-supplied list), derive 5–10 topics. `lookup intent-topics fuzzyMatch=<theme>` — one call per theme (multi-field fuzzyMatch fails silently). If no topics resolve, set intent weight = 0 and redistribute.

### 5. Fetch data per account (parallel, batched ≤10; chunked for large lists)

For each resolved account, in parallel:
- `enrich_companies(companyId, fields: industries, employeeCount, revenue, country, metroArea, businessModel, employeeCountByDepartment, foundedYear)`.
- `enrich_intent(companyId, topics: curated list, signalScoreMin: 60, signalStartDate: 30d ago, pageSize 25)`.
- `enrich_news(zoominfoCompanyId, last 90d, categories: PERSON,MERGER_OR_ACQUISITION,FUNDING,PRODUCT,FINANCIAL_RESULTS, pageSize 15)`.
- `enrich_scoops(zoominfoCompanyIds, last 90d, pageSize 10)`.

**Hard batch limit: ≤10 concurrent per MCP tool type.**

**Batch + context-window discipline.** For lists >25 accounts, process in **chunks of ~25 accounts** end-to-end (resolve → fetch → score → compose row → write chunk → discard raw payloads) before moving to the next chunk. Don't accumulate full raw enrichment payloads for hundreds of accounts in working context — once per-axis scores + the winning trigger event + the winning intent topic are captured per account, drop the rest. For >100-account lists, summarize completed chunks into running totals (tier distribution, top-A list, multi-product anomalies, duplicate-suspected flags, missing-axes counts) and discard the per-account breakdowns from context. Output is built incrementally chunk-by-chunk so a long list doesn't blow context.

Skip `account_research` here; fire selectively in §7.5 for tier-A.

### 6. Score each axis

**Fit (0–100)** — compare `enrich_companies` to `get_gtm_context.icp`:

| Dimension | Max | Banded scoring |
|---|---|---|
| Industry / sub-industry | 25 | Primary = 25 · secondary = 15 · adjacent = 8 · none = 0 |
| Employee count band | 20 | In band = 20 · one off = 12 · two off = 4 · outside = 0 |
| Revenue band | 20 | Same banding |
| Geography | 15 | ICP country = 15 · in continent = 8 · outside = 0 |
| Business model | 10 |