Abm account brief skill

ABM Account Brief Skill

Generates structured, evidence-based account briefs for B2B outbound. The output is designed to feed directly into a Smartlead / Apollo / LinkedIn Sales Navigator sequence with minimal human editing.

When to use

• You are about to add a new account to a target list and need a brief before first touch.
• You are refreshing an account that has gone cold and need a new angle.
• You are running a batch (10-500 accounts) and want consistent, scoreable output.

When NOT to use

• The account is already in active sales conversation — use the CRM record, not a fresh brief.
• The account is outside the defined ICP (see references/icp-criteria.md). Do not generate; return out_of_icp: true with the reason.
• You do not have at least one recent signal (news, hiring, funding, product launch, leadership change). Without a signal, hooks are generic and rep-time is wasted. Return insufficient_signal: true and surface what signal is missing.

Inputs

Required:

• company_domain — root domain (e.g. acme.com)
• persona — one of: CTO, VP Engineering, Director Marketing Technology, Head of RevOps, VP Sales

Optional but recommended:

• recent_signals — list of {date, type, source_url, summary}; types: funding, hiring, product_launch, leadership_change, press, linkedin_post
• prior_touches — list of {date, channel, outcome}
• account_owner_notes — free text from the rep

Output schema (JSON)

{
  "company": {
    "name": "string",
    "domain": "string",
    "industry": "string",
    "size_band": "1-50 | 51-200 | 201-1000 | 1001-5000 | 5000+",
    "hq_country": "string",
    "one_line_what_they_do": "string"
  },
  "icp_fit": {
    "in_icp": true,
    "score_0_to_10": 8,
    "rationale": "string"
  },
  "hooks": [
    {
      "rank": 1,
      "angle": "string (one sentence, what the hook is)",
      "evidence_url": "string (must be a real URL from recent_signals or public web)",
      "why_it_lands_for_persona": "string"
    }
  ],
  "recommended_channel": "email | linkedin_dm | linkedin_inmail",
  "channel_rationale": "string",
  "cold_email_draft": {
    "subject_line": "string (max 50 chars, no emojis, no salesy verbs)",
    "body": "string (max 60 words, plain text, single CTA, no jargon)"
  },
  "confidence_0_to_10": 7,
  "confidence_rationale": "string",
  "guardrails_triggered": []
}

How the model should reason

1. Verify ICP fit first. Read references/icp-criteria.md. If the company fails any hard criterion (industry exclusion, sub-scale headcount, wrong geo), return out_of_icp: true and stop. Do not generate hooks for non-ICP accounts.
2. Find the strongest signal. Rank recent_signals by recency (last 30 days > last 90 days > older) and specificity (a named person's LinkedIn post > a corporate press release). The top signal anchors hook #1.
3. Generate three distinct hooks. Each must cite a different evidence point. No hook may be a paraphrase of "I saw your company is growing" — that is generic and will be rejected by the eval.
4. Match channel to persona and signal type. Leadership change or LinkedIn post → LinkedIn DM. Funding or product launch → email with a clear business framing. Hiring signal for a senior role → LinkedIn InMail to the hiring manager.
5. Write the email. 60 words MAX. Subject line cannot contain: quick, touching base, circling back, synergy, leverage, unlock. The body must reference the evidence URL's content (not just exist), state one specific value claim, and end with a single low-friction CTA (a question, not a meeting ask).
6. Score confidence honestly. If you had to stretch on evidence, say so. Confidence below 5 means the rep should not send this without human review.

Guardrails

The Skill enforces these before returning output. If violated, add to guardrails_triggered and either fix or refuse:

• No fabricated evidence. Every evidence_url must come from recent_signals or be a URL you have been given. Do not invent URLs.
• No PII inference. Do not guess personal details (family, health, politics, religion) from the persona's profile.
• No competitor name-drops unless they are in the public signal itself.
• Subject line length ≤ 50 chars. Body ≤ 60 words. Hard caps.
• Banned subject-line phrases (see step 5 above).
• Output must be valid JSON matching the schema. No prose outside the JSON.

Eval loop

This Skill is regression-tested against golden_set.jsonl. Run python scripts/score_output.py after any prompt change. The eval scores on:

• Schema validity (binary)
• Evidence grounding (every hook's URL appears in the input signals)
• Hook diversity (no two hooks paraphrase the same signal)
• Subject/body length compliance
• LLM-as-judge rating of hook quality vs the golden hook (1-5)

See prompts/v1.md and prompts/v2.md for version history. v2 added the "no generic growth hooks" instruction after v1 failed the diversity check on 3/10 golden cases.