Leads automatisch anreichern mit Clay + Claude

---
name: lead-enrichment
description: Enrich a Clay table row with company summary, ICP fit score, and a personalized cold-email opener. Use this skill from Clay AI columns or from a Claude Code session driving the Anthropic Batch API over an exported lead list. Returns structured fields plus a draft opener for rep review — never auto-sends.
---

# Lead enrichment

## When to invoke

Whenever a Clay table or exported CSV of leads needs three derived fields in one pass: a 2-sentence company summary, an ICP fit score with one-line reasoning, and a sub-50-word cold-email opener that cites a real signal from the company's public footprint. Take a domain (and optionally a LinkedIn URL or a free-text ICP rubric) as input and produce a structured JSON-shaped Markdown record per row.

Do NOT invoke this skill for:

- Auto-sending email. The opener is a draft for rep review. Wiring the output directly into a sequence-send step bypasses the guardrail this skill exists to enforce.
- Enriching leads that have not consented to processing. If the source list was scraped from a region with prior-consent rules (EU/UK GDPR, Quebec Law 25, etc.) without a documented lawful basis, stop and surface the concern to the operator. This skill will not silently process unconsented data.
- Account-level research briefs for discovery calls — use the `account-research` skill instead. This one optimizes for batch volume and cost-per-row, not depth.
- Public-company financial scoring or any decision that needs licensed data (S&P, Pitchbook). This skill reads public web only.

## Inputs

- Required: `domain` — the company's primary domain (e.g. `acme.com`). Must resolve and serve HTML; parked or dead domains are returned with `status: unreachable` and skipped.
- Required at config time: a Clay table reference (table ID + column map) OR a CSV path if running outside Clay. The skill assumes the table has at least `domain`; `first_name`, `title`, and `linkedin_url` improve opener quality if present.
- Required at config time: `references/icp-rubric.md` populated with your real rubric. Without it, `icp_fit_score` is omitted (not guessed).
- Optional: `references/opener-style-guide.md` — your team's voice, banned phrases, max length override.
- Optional: `references/source-quality-matrix.md` — vendor preference order when multiple enrichment columns upstream of this skill disagree.
- Optional: `recent_news` — pre-fetched by an upstream Clay column. When present, the opener step uses it instead of re-fetching.

## Reference files

Always read the following from `references/` before generating any row. They encode the operator's positioning and quality bar — without them output is generic.

- `references/icp-rubric.md` — the rubric the score uses. Replace template content with your actual ICP definition before first run.
- `references/opener-style-guide.md` — voice, length cap, banned phrases (e.g. "I noticed", "love what you're doing", any superlative).
- `references/source-quality-matrix.md` — preference order for resolving conflicts between upstream enrichment vendors (Apollo vs. Clearbit vs. ZoomInfo vs. Clay's native People/Company API).

## Method

Run these four sub-tasks in order per row. Steps 1-3 produce the structured fields; step 4 only runs when the ICP score clears the configured threshold.

### 1. Resolve and fetch

Hit `https://{domain}` with a 10-second timeout and one redirect hop. On non-2xx, parked-domain heuristics, or empty body, mark `status: unreachable` and skip steps 2-4. Do not retry indefinitely — parked domains are a meaningful share of any scraped list and the skill should fail fast on them rather than burn API spend.

If the homepage resolves, additionally fetch `/about`, `/company`, and `/customers` (best-effort, ignore 404s). Concatenate cleaned text up to ~8K tokens; truncate from the end of the longest section, not the front, to preserve the homepage hero copy.

### 2. Extract structured snapshot

In one Claude call, extract:

- `industry` — single noun phrase, derived from copy not guessed
- `size_signal` — one of `solo | smb | mid | ent`, justified by a cited copy fragment (employee count claim, customer-logo density, funding mention) or `unknown`
- `value_prop` — one sentence in the company's own framing
- `recent_signal` — at most one verifiable fact (a launch, hire, funding round, customer win) with a URL anchor. If none found in the fetched pages, return `null` rather than inventing one.

Engineering choice: extraction is a single structured Claude call, not three. Each additional round-trip multiplies cost-per-row at 100K-row scale; the extraction prompt is small enough to stay reliable in one pass.

### 3. Score against ICP rubric

Load `references/icp-rubric.md`. Score the snapshot 1-10 with a one-line justification that names which rubric dimension drove the score. If the rubric file still contains template placeholders (`{...}` literals), return `score: null` and `reason: "rubric not configured"` rather than scoring against nothing.

Engineering choice: ICP scoring runs before opener generation, not after. Rationale: opener generation is the most expensive step (longer output, more careful prompt). Skipping it for sub-threshold rows is where the cost savings live. A common misconfiguration is to generate openers for every row "in case the rep wants to override" — that defeats the filter.

### 4. Generate opener (only if score >= threshold)

Default threshold: 6/10. Configurable per run.

Inputs to the opener step: `value_prop`, `recent_signal` (if non-null), the lead's `first_name` and `title` if present, and the style guide.

Hard rules baked into the opener prompt:

- Max 50 words.
- Must reference exactly one specific thing from `recent_signal` or `value_prop`. If both are null, return `opener: null` rather than writing flattery.
- No superlatives ("amazing", "incredible", "love").
- No invented company claims. The opener may only reference facts present in the snapshot. The prompt includes the snapshot inline and explicitly forbids facts not in it.
- No question close that pretends to know an internal pain ("how are you handling X?" without evidence X is happening).

The opener is always returned as a draft with a `confidence` field (0-1) reflecting how strong the cited signal was. Reps see the confidence; low-confidence openers are surfaced for rewrite, not auto-sent.

## Output format

One record per row, emitted as a fenced JSON block inside Markdown so Clay's AI column can parse it and a human reading the log can scan it.

```markdown
### acme.com

```json
{ "domain": "acme.com", "status": "ok", "industry": "Workforce management software", "size_signal": "mid", "size_signal_evidence": "About page cites '350+ employees across 4 offices'", "value_prop": "Schedules and pays hourly workforces for restaurants and retail.", "recent_signal": { "summary": "Launched a payroll module in March 2026.", "url": "https://acme.com/blog/payroll-launch" }, "icp_fit_score": 8, "icp_fit_reason": "Mid-market, hourly-workforce vertical — direct match on rubric line 2.", "opener_draft": "Saw the March payroll launch — folding pay into the same surface as scheduling is the move most ops leaders ask us about. Worth a 20-min on whether the multi-state tax piece is hitting the same wall others have?", "opener_confidence": 0.78, "sources": [ "https://acme.com/", "https://acme.com/about", "https://acme.com/blog/payroll-launch" ] }
```
```

For unreachable rows the record collapses to `{ "domain": "...", "status": "unreachable" }` with no other fields — no placeholder strings.

## Watch-outs

- **Source-quality drift across vendors.** When a Clay table has both Apollo and Clearbit enrichment columns upstream and they disagree on headcount or industry, this skill's snapshot can flip-flop run to run. Guard: `references/source-quality-matrix.md` declares a preference order; the snapshot step cites which vendor (or homepage-derived value) it used per field, so drift is auditable.

- **Opener inventing claims that aren't in the data.** Without strict prompting, openers drift toward confident-sounding fabrications ("congrats on your Series C" when there was no Series C). Guard: the opener prompt receives the snapshot inline and has an explicit "facts not in the snapshot are forbidden" rule; `recent_signal` carries a URL so a reviewer can verify in one click; openers with `opener_confidence` under 0.5 are flagged for rewrite, not sent.

- **Cost-per-row escalation if the ICP filter is loose.** A rubric that scores most rows 7+ defeats the threshold gate; opener generation runs on every row and per-row cost rises 3-4x. Guard: the skill logs a `score_distribution` summary at the end of each batch (`{ "1-3": N, "4-6": N, "7-10": N }`); if more than 60% of rows land 7+ across a 1K-row sample, the skill prints a warning to tighten the rubric before the next batch.

- **Rate limits on homepage fetches.** Hitting one root domain per row is fine; hitting `/about` and `/customers` triples the request count and some hosts throttle. Guard: per-host concurrency capped at 2, per-host minimum delay 250ms, and 429s are honored with a single back-off retry before the row is marked `status: unreachable`.

- **Stale enrichment.** A row enriched 90 days ago is not the same signal as one enriched today; reps treating old `recent_signal` values as fresh write embarrassing openers. Guard: every record carries an `enriched_at` timestamp, and the Clay column is configured to re-run when `enriched_at` is older than 30 days.

# ICP rubric — TEMPLATE

> Replace every `{...}` placeholder with your team's real values before
> the first run. The lead-enrichment skill checks for unsubstituted
> placeholders and refuses to score against a template — it returns
> `score: null, reason: "rubric not configured"` instead of guessing.

## Firmographics

| Dimension | In-ICP | Stretch | Out |
|---|---|---|---|
| Industry | {list, e.g. "B2B SaaS, fintech, vertical SaaS"} | {adjacent industries} | {hard out, e.g. "consumer, agencies, edu"} |
| Headcount | {range, e.g. "200-2000"} | {range, e.g. "100-200, 2000-5000"} | {below/above, e.g. "<100, >5000"} |
| Revenue | {range, e.g. "$25M-$500M ARR"} | {range} | {below/above} |
| Geo | {regions, e.g. "US, Canada, UK"} | {regions, e.g. "EU, ANZ"} | {regions, e.g. "China, Russia, sanctioned"} |
| Funding stage | {stages, e.g. "Series B-D, public"} | {stages} | {stages, e.g. "pre-seed, bootstrapped <$5M"} |

## Technographics

Tools whose presence on the prospect's stack signals fit (we win when they have these because the integration story is short or the pain is acute):

- {Tool 1 — e.g. "Salesforce + Outreach"}
- {Tool 2 — e.g. "dbt + Snowflake"}
- {Tool 3 — e.g. "Segment"}

Tools whose presence signals misfit (we lose to them, or their presence implies a build-not-buy culture):

- {Tool A — e.g. "Hightouch (we lose to)"}
- {Tool B — e.g. "in-house reverse-ETL (build-not-buy)"}

## Pain ranking

When the skill produces an ICP score, it weights against this priority order. Higher = more important to surface first.

1. {Pain category — e.g. "outbound data quality blocking pipeline"} — weight 5
2. {Pain category — e.g. "rep ramp time eating quota"} — weight 4
3. {Pain category — e.g. "RevOps reporting backlog"} — weight 3
4. {Pain category — e.g. "tooling sprawl / consolidation push"} — weight 2
5. {Pain category — e.g. "compliance / data-residency"} — weight 1

## Score interpretation

The skill returns `1-10`. Suggested interpretation (override per team):

- **9-10** — direct ICP, multiple positive signals, no disqualifiers. Send.
- **7-8** — strong fit on firmographics, signal is plausible. Send after rep glance.
- **5-6** — adjacent. Default threshold cuts here. Opener may not generate.
- **3-4** — stretch fit, weak signal. Skip unless rep has specific reason.
- **1-2** — out of ICP. Suppress.

## Disqualifiers

Single signals that drop a row out regardless of other fit. The skill flags these with `disqualifier: "..."` in the output and forces score to 1.

- {Disqualifier 1 — e.g. "uses {Competitor} as system of record"}
- {Disqualifier 2 — e.g. "regulated industry without our SOC 2 + HIPAA"}
- {Disqualifier 3 — e.g. "headcount under 50 — wrong motion"}

## Last edited

{YYYY-MM-DD} — bump on every material change. The skill prepends a "rubric may be stale" note when this date is more than 90 days old.

# Opener style guide — TEMPLATE

> Replace the placeholders with your team's actual voice rules. The
> lead-enrichment skill loads this file on every opener generation.
> Treat it as the single source of truth — do not duplicate rules into
> the prompt itself, where they will drift.

## Voice

One sentence on register: {e.g. "direct, plain, no jargon, no exclamation marks, no questions in the close unless the question is specific"}.

One sentence on stance: {e.g. "we're peers solving the same problem, not vendors selling at"}.

## Length

- Hard cap: 50 words. The skill enforces this and truncates with an ellipsis warning if the model overshoots.
- Soft target: 30-40 words. Three sentences max.
- No greeting line ("Hi {name},"). The sequence step adds the greeting; baking one into the opener double-greets.

## What every opener must contain

1. A specific reference to one thing from the snapshot — preferably `recent_signal`. If there is no recent signal, the opener references the `value_prop` and labels itself low-confidence.
2. A reason the reference matters to the operator's persona. Not "I noticed X" — "X is the move ops leaders ask us about".
3. A single, specific close. Either a 20-min ask or a question that would be answerable only by someone inside the company.

## Banned phrases

The prompt rejects any opener containing these. Add team-specific additions over time.

- "I hope this email finds you well"
- "I noticed"
- "I came across your"
- "love what you're doing"
- "quick question"
- "circling back" (in a first touch)
- Any superlative — "amazing", "incredible", "best-in-class", "leading"
- Any em-dash trio that reads as an LLM tic — "X — Y — Z"
- Compliments on growth/funding without a cited source

## Banned structures

- The "I was just on your website" opener. Always reads false.
- The "we work with companies like {logo}" name-drop in sentence one.
- The "are you the right person?" close. Wastes the reply.

## Confidence signal

The skill returns `opener_confidence: 0.0-1.0`:

- **0.8+** — opener references a dated `recent_signal` with a URL and the rep's persona is in `references/1-icp-rubric.md` buying committee.
- **0.5-0.8** — opener references `value_prop` only, or `recent_signal` is older than 60 days.
- **<0.5** — opener is generic; rep should rewrite or skip the row.

Threshold for auto-queueing into a sequence: configure per-team. Default recommendation: **never auto-queue**. The opener is a draft.

## Worked example

Snapshot: industry "scheduling for restaurants", recent_signal "March payroll launch", value_prop "schedules and pays hourly workforces".

Opener (35 words, confidence 0.78):

> Saw the March payroll launch — folding pay into the same surface as
> scheduling is the move most ops leaders ask us about. Worth a 20-min
> on whether the multi-state tax piece is hitting the same wall others
> have?

Why it scores 0.78 not 1.0: the close assumes a pain ("multi-state tax piece is hitting the same wall") that the snapshot does not directly evidence. A 1.0 opener would tie the close to a fact in the snapshot.

## Last edited

{YYYY-MM-DD}

# Source quality matrix — TEMPLATE

> Replace placeholders with the vendors actually wired into your Clay
> table. The lead-enrichment skill consults this matrix when two
> upstream sources disagree on the same field, and cites which source
> won in the per-row output for audit.

## Why this exists

Clay tables routinely stack 2-3 enrichment vendors per field — Apollo on industry, Clearbit on headcount, ZoomInfo on funding. They disagree often: Apollo will say a 200-person company is 50, Clearbit will assign the wrong industry, ZoomInfo will lag a recent acquisition by a quarter.

Without a declared preference order, the skill's snapshot step flip-flops between vendors run to run, ICP scores oscillate, and operators cannot reproduce the same row twice.

## Field-by-field preference order

Per field, list vendors top-to-bottom in trust order. The skill takes the highest-ranked vendor that returned a non-empty value. The homepage-derived value (extracted by the skill itself in step 1) is always the tiebreaker if all vendors disagree with each other.

### Industry / sector

1. {e.g. "Homepage About-page extraction (skill step 1)"}
2. {e.g. "Clearbit Reveal"}
3. {e.g. "Apollo Company"}
4. {e.g. "ZoomInfo Industry"}

### Headcount

1. {e.g. "LinkedIn employee count via Clay's Linked-In Profile column"}
2. {e.g. "Apollo employee_count"}
3. {e.g. "Clearbit metrics.employees"}
4. Skill homepage extraction (last resort — copy claims often inflated)

### Revenue / ARR

1. {e.g. "ZoomInfo revenue (paid only)"}
2. {e.g. "Clearbit metrics.annualRevenue (estimated, low confidence)"}
3. Omit (do not extract from homepage; vanity metrics are not revenue)

### Funding stage / last round

1. {e.g. "Crunchbase via Clay's native column"}
2. {e.g. "Press-release fetch in skill step 1, only if dated within 90d"}
3. Omit

### Recent signal (launches, hires, news)

1. Skill homepage / blog fetch in step 1 — must carry a dated URL
2. {e.g. "Clay's Built-With column for tech-stack changes"}
3. {e.g. "Google News column, capped to last 60 days"}

Important: a vendor returning "we don't know" is treated as missing, not as a contradicting value. The next vendor in the list is consulted.

## Conflict logging

When two vendors return different values for the same field, the skill adds a `conflicts` block to the per-row output:

```json
"conflicts": [
  { "field": "headcount", "values": { "linkedin": 380, "apollo": 120 }, "chose": "linkedin" }
]
```

Operators should sample 10-20 conflict rows weekly. If one vendor is consistently wrong, demote it in this matrix. The matrix is the control surface, not the prompt.

## Vendor cost ceiling

Each vendor has a per-row cost (Clay credits or external API spend). Set a soft cap here — if a row's combined enrichment cost exceeds the cap, skip the lowest-ranked vendor and accept lower confidence.

- Soft cap per row: {e.g. "12 Clay credits"}
- Hard cap per row: {e.g. "20 Clay credits — abort row"}

The skill respects these caps and emits `cost_ceiling_hit: true` when the hard cap fires, so the batch summary surfaces the count.

## Last edited

{YYYY-MM-DD}