Sales Outreach

Sales Outreach

One job: what do I say — and can they tell I actually looked?

The copy is the easy half. The reason a touch lands is that it opens with something specific, recent, and true about the person reading it: a post they wrote last week, the round they closed last month, the role they just opened, the exec they just hired. This skill pulls that from Crustdata first, then writes around it.

The hook comes before the copy. Always. A beautifully written email built on "I saw you're in fintech" is a template. A blunt three-line email built on "your Tuesday post about the on-call rotation" gets replies. If no real signal exists for a target, say so and offer a different angle — never invent a hook.

Code Mode ground rules (read once, apply everywhere)

All Crustdata calls run inside the execute tool of the Crustdata MCP server (install.crustdata.com/mcp) as a plain-JavaScript script. Every snippet below follows these rules:

• Plain JavaScript only. Author against the typed tool surface from get_schema, but the script body carries zero type annotations — a : Type, as, or generic is a parse error that fails the whole run before any tool call, at zero spend.
• Open every script with a source-labeled comment: // user query: ... for the user's literal ask, // model query: ... for a derived step. Scripts without one are rejected before running.
• One I/O primitive: const r = await callTool(name, params) → { ok: true, data } or { ok: false, status, errorType, message }. Always branch on r.ok — a failed call does not abort the script, so an unchecked failure silently proceeds on empty data and looks like "they have no recent activity". That is exactly the failure that produces a fabricated hook.
• fields is a response whitelist. The result carries only the groups you list; an omitted group reads as undefined later and looks like missing data. List every group you read.
• Return the smallest projection. Only what the script returns reaches the model. For outreach that means hook candidates — date, one line of text, a link — never raw posts or whole profiles.
• Fan out independent work with await parallelMap(items, fn); chunk first where there is a per-call cap. Never parallelize cursor pagination or dependent stages. checkpoint(acc) after a costly stage so a timeout still returns the hooks you paid for.

Step 0 — Intake (always, before writing a word)

1. Config backbone, never blocking. If config/persona-profile.md or config/gtm-config.md exist in the working directory, read them: the rep's voice (tone, sentence length, sign-off, never-do list), what they sell, who buys. If they don't exist, ask one inline question ("What do you sell, and what's your sign-off?") or point at the icp-builder skill, which writes both files from one LinkedIn URL. A missing config never blocks a draft — write in a plain, direct voice and say you did.
2. Confirm the ask in one line: WHO (person, account, or list) — CHANNEL (email / call / LinkedIn) — MOMENT (first touch / follow-up / reply / objection) — GOAL (a meeting? an answer? an intro?).
3. Pull the hook (next section). One sharp hook beats five soft ones.

THE HOOK LAYER (this is the skill)

Four sources, cheapest and most specific first. Pull one; stop when you have something good. Running all four on one person is usually a waste of credits.

Hook 1 — what they said (their posts)

The best hook there is, because it's theirs and it's dated. Keyed on the person's professional network profile URL. 1 credit per post, so limit is the spend dial — 5 is plenty to find one hook.

// user query: write a cold email to https://www.linkedin.com/in/example — find a hook first
const url = "https://www.linkedin.com/in/example";
const cutoffMs = Date.now() - 60 * 24 * 3600 * 1000; // hooks older than ~60 days read as stale

const r = await callTool("social_post_list_live", {
  professional_network_profile_url: url, // person key; company_domain / crustdata_company_id key the company feed
  limit: 5,                              // 1 CREDIT PER POST — set this deliberately
  fields: ["text", "date_posted", "post_type", "engagement", "hyperlinks", "share_url"], // whitelist
});
if (!r.ok) return { hooks: [], error: r.message };

const parse = (d) => { const t = Date.parse(d ?? ""); return Number.isNaN(t) ? null : t; };
const posts = (r.data.posts ?? []).map(p => ({
  date: p.date_posted,
  ms: parse(p.date_posted),
  type: p.post_type,
  url: p.share_url, // the link to the post itself — the rep needs it to verify in five seconds
  text: (p.text ?? "").slice(0, 400),
  reactions: p.engagement?.total_reactions,
  comments: p.engagement?.total_comments,
  links: p.hyperlinks?.other_urls ?? [],
}));
return {
  fresh: posts.filter(p => p.ms !== null && p.ms >= cutoffMs),
  stale: posts.filter(p => p.ms !== null && p.ms < cutoffMs).length,
  undated: posts.filter(p => p.ms === null), // never silently drop these — read them yourself
};

Reading the result: prefer a post they wrote over a reshare with no commentary, and an opinion over an announcement. A post with a real take gives you a sentence you can agree or disagree with, which is what earns a reply.

Hook 2 — what happened to the company (funding, news, growth)

company_identify is free but fuzzy — one domain can match several companies. Take the top confidence_score, then enrich by id so you pay for exactly one match (2 credits). Enriching by domain or name instead can fuzzy-match several companies and bill 2 credits per returned match, which exact_match: true narrows but does not guarantee down to one.

// model query: recent company events at acme.com worth opening an email with
const idr = await callTool("company_identify", { domains: ["acme.com"] }); // free, one identifier type per call
if (!idr.ok) return { error: idr.message };
const matches = idr.data[0]?.matches ?? [];
const best = matches.slice().sort((a, b) => (b.confidence_score ?? 0) - (a.confidence_score ?? 0))[0];
if (!best) return { error: "no company match for acme.com" };
const companyId = best.company_data?.basic_info?.crustdata_company_id ?? best.company_data?.crustdata_company_id;

const er = await callTool("company_enrich", {
  crustdata_company_ids: [companyId],                       // one id = one match = exactly 2 credits
  exact_match: true,                                        // narrows the fuzzy match; by-name/by-domain can still return >1 -> N x 2 cr
  fields: ["basic_info", "funding", "headcount", "news"],    // whitelist: list EVERY group you read
});
if (!er.ok) return { error: er.message };
const cd = er.data[0]?.matches?.[0]?.company_data ?? {};
const news = (firstArray(cd.news) ?? []).slice(0, 10).map(n => ({
  title: n.article_title, url: n.article_url, date: n.article_publish_date,
}));
return {
  name: cd.basic_info?.name,
  lastRound: cd.funding?.last_round_type,
  lastRaise: cd.funding?.last_fundraise_date,
  lastRoundUsd: cd.funding?.last_round_amount_usd,
  investors: cd.funding?.investors,
  headcount: cd.headcount?.total,
  growth12m: cd.headcount?.growth_percent?.["12m"],
  news,
};

Funding is a hook with a shelf life. A round closed 3 weeks ago is "congrats + here's what usually breaks next"; a round closed 14 months ago is background, not an opener. News articles carry a real publish date — use it in the sentence and link the article if you reference a specific claim.

Hook 3 — what they're hiring for

Job posts are the most honest thing a company publishes. The roles they're opening are the problems they've agreed to spend money on.

// model query: what acme.com is hiring for, and whether their posts name our category
const domain = "acme.com";
const term = "<the tool you displace>"; // a brand/product/tech name -> ALWAYS [.] exact token, never (.)
const [fresh, named] = await parallelMap([
  {
    filters: eq("company.basic_info.primary_domain", domain),
    sorts: [{ field: "metadata.date_added", order: "desc" }],
    fields: ["job_details", "location", "metadata"],
    limit: 15,
  },
  {
    filters: and_(
      eq("company.basic_info.primary_domain", domain),
      exactToken("content.description", term) // [.] guarantees the literal token; (.) is typo-tolerant and pulls lookalikes
    ),
    fields: ["job_details"],
    limit: 1, // count query — read total_count, don't download rows
  },
], async (params) => await callTool("job_search", params));
if (!fresh.ok) return { error: fresh.message };
return {
  newest: fresh.data.job_listings.map(j => ({
    title: j.job_details?.title,
    city: j.location?.city,
    added: j.metadata?.date_added,
  })),
  total_open: fresh.data.total_count,
  postings_naming_term: named.ok ? named.data.total_count : null,
};

Scope roles by title, not job_details.category — the category field is coarse and most postings land in the catch-all. If you want to say "you've got four data engineering roles open", count titles.

The hook writes itself from the shape: three SDR reqs open means a pipeline target they just raised. A first security hire means they're about to buy security tooling. Name the actual role and when it was posted.

Hook 4 — a signal that already fired

The best case: the work is already done. sales-prospecting's champion tracker, expansion radar, and account ranking all emit dated signal rows. Hand one straight in and skip the pull entirely.

A handed-over row looks like: person — was [role] at [your customer] — now [title] at [new company] — landed [date]. That IS the hook, and it's already dated. Same for "raised Series B on 2026-04-12" or "hired a VP Data 6 weeks ago".

Do not re-pull what you were handed. Spend a credit only to add colour (their first post in the new job, say) and only if the draft needs it.

Signal → angle

No hook, no fabrication

If the pulls come back empty — no recent posts, funding two years old, no open roles — say that out loud and offer the alternatives instead of inventing a reference:

• Write from a role-level insight ("every VP Eng at 300 people hits this") and label it as such — a good generic beats a fake specific.
• Pull a hook on a different person at the same account who is active — if you already have their profile URL (from the user's list, or from sales-prospecting / account-research, which do the person discovery), run Hook 1 on them instead.
• Use the account's own words: social_post_list_live keyed on company_domain gives the company feed when the individual is quiet.
• Recommend waiting for a signal and setting up the watcher in sales-prospecting's champion tracker.

A hook must pass three tests before it goes in a draft: is it theirs (not their industry's), is it dated (post ≤60 days, funding/news ≤90 days), and would they recognize it in one read.

THE TOUCH TYPES

1. Cold email

Under 90 words. Subject 3-5 words, lowercase-ish, no clickbait, no "quick question".

• Line 1 — the hook. Their thing, dated, in their language. The reader should think "yes, I did that".
• Line 2 — the bridge. One sentence connecting their thing to the pain you solve.
• Line 3 — one concrete outcome. A number or a named customer situation, not a value prop.
• Line 4 — one ask. The lowest-friction version that still moves ("worth 15 minutes Thursday?" or "want me to send the two-line version?").
• Their sign-off, if the persona config carries one. The config written by icp-builder records tone and style but not a sign-off, so ask for it inline the first time and reuse it after.

One hook, one idea, one ask. Offer 2-3 angle variants for testing (different hook, or same hook different pain) and say which one you'd send.

2. Call script

• Opener + permission line ("caught you cold — 30 seconds and you can tell me to go away?").
• Reason for the call, tied to the same hook. Cold calls die on "just reaching out".
• Value hypothesis in one sentence, framed as a guess they can correct.
• 2-3 open discovery questions — how they handle it today, what it costs them.
• The ask: a meeting, not a sale.
• Objection branches, written out: not interested / send me an email / we already use something / no budget / gatekeeper. Each is one acknowledgment plus one question that keeps the call alive.
• 15-second voicemail and a one-line follow-up text, both referencing the same hook so the touches stack.

3. LinkedIn touches

• Connection note — 300 characters max, no pitch. The hook plus a reason to connect.
• First DM — value or a smart question. Still no pitch.
• Follow-up — a soft ask, only after a reply or a signal.
• Break-up — one line, gracious, leaves the door open.

Before any of it, suggest 1-2 of their recent posts worth a genuine comment — you already have them from Hook 1, at no extra cost. A real comment three days before a connection request does more than any first line.

4. Reply handler

Classify first, then advance. Match their length and their energy — a two-line reply gets a two-line answer.

5. Objection talk-tracks

Diagnose the real objection first. "Too expensive" is usually "value unclear" or "no internal champion". "Bad timing" is usually "not a priority". Say what you think the real one is before answering it.

Then: acknowledge → reframe → one real proof point → soft re-ask. One proof point, and it must be real — a named situation, a number you can defend. Never invent a case study.

Give two variants: direct (for a buyer who's blunt) and soft (for a relationship you're still building).

Playbook mode: the team's top 8-10 objections in one document, each with the diagnosis, both variants, and the trap to avoid. This one is worth rendering as a page — see the artifact rule under Rules.

6. Sequence builder

1. The cadence table first — day, channel, angle, personalization level. Vary the angle across touches: problem → proof → a different persona's pain → social proof → break-up. Never send the same idea five times in different words.
2. Then write every touch using types 1-3.
3. Mark each touch 1:1 (rep writes a custom line) vs merge-field (scales across the list), and flag exactly where the rep must add something only they know.
4. Hook freshness: touch 1 carries the pulled hook. By touch 4 that hook is two weeks old — either re-pull or switch to an angle that doesn't depend on recency. Say which in the table.

Typical shape (adapt to the motion, don't ship this as-is):

7. Signal-triggered drafts

Input: a signal that just fired (new funding, exec hire, champion moved, hiring surge), usually handed over from sales-prospecting. Output: a draft that references the actual trigger with its date, mapped through the signal → angle table above.

The rule that makes these work: the draft must be sendable the week the signal fired. A "congrats on the round" email six weeks late is worse than nothing. If the signal is already stale, say so and re-angle.

To keep signals flowing automatically, point the user at sales-prospecting's champion tracker, which sets up the recurring watcher.

THE NO-SLOP RULE (every draft, no exceptions)

Run this pass on every piece of copy before showing it. A touch that smells like AI is worse than no touch — it tells the reader they were on a list.

Kill on sight:

• Em dashes. Use a period or a comma.
• "I hope this finds you well", "I wanted to reach out", "quick question", "circling back", "just following up".
• "delve", "leverage", "streamline", "unlock", "elevate", "seamless", "robust".
• Rule-of-three constructions ("faster, cheaper, and more reliable").
• Negative parallelism ("not just X, but Y").
• Filler adverbs: "really", "actually", "truly", "incredibly".
• Fake familiarity ("Hope you're crushing it!") and fake urgency.
• Any sentence that would survive a find-and-replace of the company name. If it works for any company, it works for none.

Then read it out loud. It should sound like the rep texting a smart colleague. If a draft reads like a template, rewrite it — do not ship slop.

LIST MODE — many accounts, one run

When the input is a list (usually a scored list from sales-prospecting), the order is always: hooks → copy → export. Not the reverse.

Step 1 — fan out the hooks. Independent per person, so parallelMap them. State the cost before running: posts are 1 credit each, so limit: 3 across 40 people is up to 120 credits.

// user query: draft outreach for these 40 accounts — pull a hook for each first
const people = inputs.people; // [{ name, profileUrl, company, domain }]
const cutoffMs = Date.now() - 60 * 24 * 3600 * 1000;

const hooks = await parallelMap(people, async (p) => {
  const r = await callTool("social_post_list_live", {
    professional_network_profile_url: p.profileUrl,
    limit: 3,                                        // 1 cr/post x 3 x N people — quote this before running
    fields: ["text", "date_posted", "post_type", "share_url"],
  });
  if (!r.ok) return { name: p.name, hook: null, reason: r.message };
  const fresh = (r.data.posts ?? [])
    .map(x => ({ date: x.date_posted, ms: Date.parse(x.date_posted ?? ""), url: x.share_url, text: (x.text ?? "").slice(0, 300) }))
    .filter(x => !Number.isNaN(x.ms) && x.ms >= cutoffMs);
  return { name: p.name, company: p.company, hook: fresh[0] ?? null, reason: fresh.length ? null : "no recent post" };
});
checkpoint(hooks);
return {
  withHook: hooks.filter(h => h.hook),
  needCompanyFallback: hooks.filter(h => !h.hook).map(h => ({ name: h.name, company: h.company, reason: h.reason })),
};

Everyone in needCompanyFallback gets Hook 2 or Hook 3 at the account level, or an honest merge-field template. Never let an empty pull turn into a made-up first line.

Step 2 — contact enrichment is opt-in and cost-confirmed. Quote the worst case before you run it, then run it:

"Business emails for 40 people: at most ~40 credits (1 per matched person). Adding personal emails and phones raises the ceiling to 5 per person, so ~200. Business only — go?"

// user query: get business emails for the 40 people on the list (worst case ~40 credits, confirmed)
const urls = inputs.profileUrls;
const batches = chunk(urls, 25); // hard cap: 25 URLs per call
const results = await parallelMap(batches, async (batch) => {
  const r = await callTool("person_contact_enrich", {
    professional_network_profile_urls: batch,
    fields: ["contact.business_emails"], // narrow the request; see the cost note below
  });
  return r.ok ? r.data : batch.map(u => ({ matched_on: u, matches: [], error: r.message }));
});
const rows = results.flat();
return {
  contacts: rows.map(m => ({
    url: m.matched_on,
    email: m.matches?.[0]?.person_data?.contact?.business_emails?.[0]?.email ?? null,
    status: m.matches?.[0]?.person_data?.contact?.business_emails?.[0]?.status ?? null,
  })),
  unmatched: rows.filter(m => !m.matches?.length).map(m => m.matched_on),
};

Omitting fields requests all three tiers (business email, personal email, phone) and takes the ceiling to 5 credits per matched person. Narrow the whitelist to the tiers you actually need, and read credits_remaining on the response for the real spend. Report unmatched rows honestly — never pad a list with guessed email patterns.

Step 3 — hand off. Two artifacts, no credentials, no auto-send:

• Campaign CSV: email, first_name, last_name, company, title, linkedin_url, hook_text, hook_date, hook_url, personalization_line, tier. The hook, its date, and the link to it ride in the file so the rep can sanity-check every row before it sends.
• Campaign spec the user pastes into whatever sequencer they run: the sequence steps with the full copy per step, the merge fields used, send window and timezone, daily ramp (start at 20-30/day on a new domain), threading (follow-ups in-thread), and stop-on-reply ON.
• Deliverability guardrails in the spec: warmed domain, at most one link in touch 1, plain text over HTML, unsubscribe honored instantly.

Rules

• The hook comes first, and it is real or it is absent. Never fabricate a post, a quote, a round, a mutual connection, or a case study. "I couldn't find a recent signal for this person, here's a role-level angle instead" is a valid, honest answer.
• Date every reference in the draft and in your notes to the rep, so they can check it in five seconds.
• One hook, one idea, one ask per touch. Shorter beats clever.
• Never auto-send anything. Email drafts stay drafts. This skill never writes to a CRM; any log entry is yours to make before writing. Sequencer setup is a spec plus a CSV that the rep loads and starts themselves.
• Respect opt-outs completely and immediately. Never volume your way past a compliance limit; if the user wants that, it's theirs to own explicitly.
• No-slop rule on every deliverable (full list above): no em dashes, no "delve"/"leverage"/"streamline", no filler, nothing that survives a find-and-replace of the company name.
• Costs are real: state expected spend before post pulls, list fan-outs, or contact enrichment. Every execute response carries credits and credits_remaining; account_credits (free) reports the balance.
• Every deliverable ends with the handoff question: copy in chat, a CSV plus campaign spec for your sequencer, or a rendered doc.
• Adapt the layout to the content — never let it hide anything. The brand system is fixed; the layout is not. If real content doesn't fit — a long company or person name, a 12-word title, 200 rows — change the layout, not the content: let the card grow, wrap instead of truncating, drop to one column, widen the column, raise the cap, or give the wide thing its own scroll container. Never solve a fit problem by clipping a card, ellipsing a name, or silently dropping rows. Where a cap really is unavoidable, say so in the UI ("showing the top 50 of 214"). Look at the rendered output and fix what's cut off before you hand it over.
• Icons in rendered output: Lucide, the dashboard's icon set, inlined as SVG with a currentColor stroke. No emojis in artifact UI.
• Photos and logos are free — use them in rendered output. basic_profile.profile_picture_permalink (person) rides in the basic_profile group of any person record handed to you; basic_info.logo_permalink (company) comes free from company_identify and from company_enrich's basic_info. Neither costs an extra credit. Base64-inline both as data:image/jpeg;base64,... URIs — the media CDN serves them as binary/octet-stream, so a remote <img src> renders blank. Monogram fallback when an image is missing.
• Artifact branding: copy is chat-native by default, and the CSV stays a plain data file — never render an artifact just to render one. But IF a deliverable is rendered as a page or document (an objection playbook, a sequence spec, a campaign one-pager), it carries the Crustdata brand lockup in the header or footer: a small uppercase "Powered by" eyebrow plus the official Crustdata wordmark, linking to crustdata.com. The wordmark pair ships in this skill's assets/ — crustdata-logo-light.png (dark text, for light backgrounds) and crustdata-logo-dark.png (white text, for dark backgrounds). Base64-inline the theme-appropriate variant at ~17px height — never hotlink; rendered artifacts cannot fetch remote images. Brand accent: #5547E2 (#8387FF on dark grounds). Body font: Geist when embeddable, else the system stack.

Error handling

• A failed call is the dangerous case, not the obvious one. r.ok === false on a post pull looks identical to "this person doesn't post". Branch on it, report it as an error, and fall back to Hook 2 or Hook 3 — never let it become a generic first line dressed up as personalization.
• 403 naming a field = a plan-gated projection. Drop that field from fields and re-run. 403 naming the endpoint (contact enrichment is the common one) = the plan doesn't include it; say so plainly and hand off the list without emails rather than guessing address patterns.
• Empty posts, empty news, no open roles is a normal outcome, not a failure. Report which sources came back empty and take the no-hook path above.
• company_identify returns several matches for one domain — that's expected, it's fuzzy. Take the top confidence_score; if two are close, name both and ask.
• Unmatched rows in a list run are reported by name, never dropped and never silently swapped for a template row.

Cost cheat sheet

A single well-hooked cold email typically costs 3-5 credits. A 40-person sequence with hooks and business emails lands around 150-200.

Works with

• sales-prospecting — builds the list and fires the signals this skill writes on. Champion-tracker rows are the highest-converting input here.
• account-research — deep-dive an account before a big-ticket touch; its account plan gives you the pain map the bridge line needs.
• icp-builder — writes config/persona-profile.md (voice) and config/gtm-config.md (what you sell) from one LinkedIn URL, so drafts sound like the rep from run one.
• warm-path-workspace — when there's a warm route in, use it: a referral opener beats the best cold email you'll ever write.

Tool dependencies

This skill requires:

• Crustdata MCP server (install.crustdata.com/mcp): a single Code Mode MCP exposing list_tools, get_schema, and execute. All Crustdata data tools are reached inside an execute({ code }) plain-JavaScript script via await callTool(name, params) — author against the typed surface from get_schema, but the script body carries zero type annotations (a type annotation is a parse error that fails the whole run). Tools used here: social_post_list_live, company_identify, company_enrich, job_search, person_contact_enrich (opt-in, cost-confirmed), account_credits.
• Nothing else. No sequencer credentials, no email-sending access, no CRM write access. Outbound leaves this skill as copy, a CSV, and a written campaign spec — the rep loads and starts it.