Contact Enrichment

Contact Enrichment

Build a research brief on one sales contact: who they are, what they've said recently, what's happening at their company, and how to open a conversation. Free web search does most of the work; a handful of paid calls fill the gaps where free search is measurably weaker.

Three rules governing every step

1. Always prompt, never guess. Ask for each step's parameters; offer concrete examples so answering is one word.
2. Announce, then wait. Before any paid step, state what it does and what it costs, then wait for explicit OK. Report the live running total, never a fixed promise up front.
3. Company-agnostic. Assume nothing about the user's own company or product. The "why us" angle comes only from their intake answer — never a built-in ICP.
4. Write explicit negatives. A slot you checked and found empty gets "none found (checked YYYY-MM-DD)". Silence is the one unusable answer: the next run cannot tell an empty slot from an unchecked one.

Prerequisites

• taskfuel CLI connected (taskfuel whoami), balance ≥ ~$0.10 for a default run.
• Base taskfuel skill covers payment mechanics — quote with --dry-run, pay with --max-amount, never suppress stderr on a paid call. Read it first if you haven't.
• Use catalog URLs exactly as taskfuel discover prints them; Orthogonal endpoints live under mpp.orthogonal.com.

Spending rules specific to this skill.

• --dry-run every call and pay with --max-amount set to the quote. Every price in this file is indicative; the quote is authoritative.
• Any call over $0.10 needs explicit OK — that's PDL ($0.28) and userInsights ($0.50).
• Report the live running total, and read taskfuel balance before and after the run. For authoritative billing, point the user at their taskfuel.ai dashboard.

Intake

Two entry points, both $0.01 — pick by what the rep has, not by price.

• Email — slightly preferable when available: unique, so it skips confirm-the-match, and it carries the company domain free.
• Name + company — the common case (referral, conference, LinkedIn browse). Requires a confirm-the-match step before Step 3.

Free web search runs on top of both, always — not as a fallback.

Reject name alone. Nothing here disambiguates a bare name. Anything else the rep knows is a tiebreaker, never assumed.

Also ask (optional): "What are you selling, and what's your angle with this person?" Drives "why us" and sharpens openers. Blank ⇒ omit that section.

Deliverable

briefs/<first-last-company>-YYYY-MM-DD.md. If one exists, ask: new dated file or overwrite. Create it right after intake — before any paid call — with Status: in-progress, so a run survives interruption. If identity resolution fails, slug whatever the input was so failed runs stay traceable.

Section order (most are Step 7 synthesis over data already bought, $0 extra):

header (name/title/company/confidence + buying-role + outreach verdict) · handle with care (placed here on SOFTEN/HOLD; otherwise in its usual slot below — never omitted silently) · trigger events / "why now" · conversation-starter hooks · suggested openers · relevance angle / "why us" (only if an angle was given) · possible common ground · role & responsibilities · company context · online presence · personal technical footprint · recent posts/talks/activity · interests · shared connections & internal relationship history · free-search completeness check · call logistics · sources & cost.

Pipeline

Paid extras that no run takes by default — career history, phone, hiring signals, firmographics, the LinkedIn fallback — live in ADD-ONS.md, with the condition that justifies each. Reach for it only when a step comes up short.

Step 1 — Doc shell (free)

Write the file: header, input, Status: in-progress.

Step 2 — Resolve identity

One paid call, either entry point, then free web search on top — always. Announce the cost first. There is no paid fallback chain: if the call comes up empty, free search is the whole recovery; if that also fails, the contact is unresolvable (terminal state below).

# email path
taskfuel call https://mpp.orthogonal.com/apollo/api/v1/people/match \
  --method POST --body '{"email":"<email>"}' --max-amount 0.01
# name+company path
taskfuel call https://mpp.orthogonal.com/apollo/api/v1/people/match --method POST \
  --body '{"first_name":"<F>","last_name":"<L>","organization_name":"<Co>"}' \
  --max-amount 0.01

Returns in one shot: name · title · linkedin_url · city/state/country · email · seniority + departments (buying-role as typed fields, no inference) · employment_history · and a full organization object (domain, industry, headcount, founded year, company LinkedIn, keywords) which covers most of Step 3a at $0 extra.

⚠ Check employment_history before anything else. A current: false on the company the rep named means the contact has moved — that outranks every other field, and the rep's premise (and angle) may be aimed at the wrong company. Depth varies by contact: sometimes one untyped entry, sometimes five with start/end dates. Read it before buying Exa or PDL for career dates.

⚠ Firmographics are estimates and have been wrong — a founded_year five years off, a headcount of 1 for a two-person team. Cite source + date; prefer a primary page.

Confirm the match with the user before Step 3 on the name+company path.

Websearch on top — free, mandatory, both paths

Not a fallback. It runs even when the paid call succeeds, because that's what catches paid-source errors — a mis-geocoded location, a dead X handle, an unverified LinkedIn URL, a founded_year five years off.

Search "{name}" {company} / "{name}" {domain}; on the email path also the address, its local part and its domain (a company's own about/team page often names the person outright).

Accept a websearch-derived identity only if corroborated ≥2× independently, at least one source primary — their LinkedIn, the company's own site, their GitHub, a conference speaker page — not RocketReach-style aggregators, which scrape the same upstream and corroborate nothing — and with no same-name ambiguity.

⚠ Websearch confabulates rather than returning null. Paid sources return null when they don't know; a search summarizer invents a plausible answer. Observed: a summary dating an article "March 2026" that its own page dates 2025-11-26; and a past job implied as current.

Conflict precedence

• Typed fields come from the paid source: seniority/departments, the organization object, employment_history, workHistory/educationHistory (Exa), phone (PDL). Websearch gives prose, not typed data.
• Except firmographics — those lose to a company's own page or a funding database.
• On conflict, websearch wins only when corroborated against a primary page — WebFetch that page and cite it. A summary is confabulation risk, not evidence.
• Location: no source is authoritative. Cross-check company HQ (and the phone country code if the escalation was bought).
• Record an unresolved conflict rather than silently picking a winner.

Terminal state — paid call empty and websearch empty

The contact is unresolvable. Stop and ask: different input, manual research, or abandon. There is no further endpoint to try, by design. Never build a brief on a guessed identity.

Writes: name, title, company + domain, career history, match confidence, which source resolved it. From the same response at $0 extra: role start date → trigger events · location → timezone · seniority/departments → buying-role · organization → Step 3a · employer + education → common ground.

Step 3 — Company context (free; the paid call is a rare fallback)

a) Profile — free search. "{company}" employees revenue founded, "{company}" funding investors, "{company}" founders. Funding databases (Crunchbase / PitchBook / Tracxn) and the company's own site cover most fields. Aggregator headcount/revenue are dated estimates — cite source + date. Free search returns exact figures where paid returns coarse buckets, and buckets hide source conflicts.

b) Tech stack — free, from live infrastructure (beats every paid source):

curl -sI https://<domain>   # server:/cf-ray: → CDN · x-powered-by: → runtime
dig +short MX  <domain>     # aspmx.l.google.com → Google Workspace
dig +short TXT <domain>     # SPF includes: zendesk, sendgrid, protonmail, …

Cross-check the runtime against the contact's own skills — a match is a hook.

c) Company GitHub — free search. "{company}" GitHub (distinct from the contact's, Step 4). Prefer over PredictLeads' paid github_repositories ($0.04), which returned a false negative. Pull primary languages, product-relevant repos, protocol/spec forks. No org is an explicit negative, not a skipped slot.

d) Paid fallback — Apollo organizations/enrich ($0.01) is an opt-in add-on; Step 2's organization object usually makes it unnecessary. See ADD-ONS.md.

Writes: industry, headcount, revenue, funding, description, tech stack, repo findings — each with source + date. Funding/launch → trigger events; headcount + seniority → buying-role.

Step 4 — Social/profile surface (free)

Paid discovery is removed — it returned nothing free search missed and produced two false negatives.

Search free, skipping anything already known: LinkedIn URL · GitHub profile · personal site/blog · talk/conference/podcast mentions.

Verify every identifier — supplied or found — before use. Any can be silently wrong: a dead X handle, an unverified LinkedIn URL, a denied GitHub account that exists. Run a free search ("{name}" {company} {network}); accept only when the result corroborates the known name and company. Mark "confirmed" only once that search has happened — never on a paid source's say-so. Verify before any paid call that consumes the identifier (twit.sh, Fiber).

Personal technical footprint — free, and the strongest hook for a technical contact. Distinct from the company org. Search "{name}" {company} GitHub and "{name}" blog OR Medium OR author. A protocol-level fork or a how-to they wrote proves individual depth a job title cannot. An explicit negative here is load-bearing: it steers openers away from a technical hook for a non-technical contact.

Email — out of scope to discover, in scope to record.

• Never spend to find one. No pattern construction, no email-finder, no PDL. Finding and verifying an address needs role-mailbox rejection (mgt@, info@) plus a verify→retry loop — a different job from this skill, and one a dedicated email-finding skill should own.
• If a call already made hands one over, keep it — Apollo's Step 2 response carries email free. Record it, labelled unverified (Apollo doesn't prove deliverability).
• Otherwise write Email: not resolved.
• ⚠ A ⚡name@domain is a lightning address, not an inbox.

Writes: verified LinkedIn · personal site/blog · verified GitHub · verified X handle · talks; explicit "not found" notes for empty results.

Step 5 — Recent activity: posts, talks, interests, news

Announce each sub-source's cost separately; any can be skipped.

Interests / writings / talks — free search. Never buy Happenstance (free search returns its entire advertised output, sourced). Search author pages (Medium, company blog, trade press) and interviews — interviews are the richest hobby source. Sourced is the normal path; tag "(inferred from public activity)" only when nothing personal surfaces.

News — POST https://mpp.orthogonal.com/serper/news ($0.002), a default step. Returns items with date + source. At ~3% of a run it is rounding error, and what it catches — a breach, layoffs, a lawsuit — is the most expensive thing to miss in outbound. Free search runs alongside.

• ⚠ Use the Orthogonal host, not stableenrich.dev/api/serper/news ($0.04) — same Google News, 20× the price. (search1api.com/news is $0.003 but returns no dates, which is the whole point — don't swap.)
• ⚠ Every date comes off a source page — summaries confabulate (Step 2), and a wrong date in a cold email is worse than a missing one. WebFetch the article or read a dated URL slug before promoting anything to a trigger event.
• Read news both ways: positive/neutral → trigger events + hooks; negative → handle with care.

⚠ The 90-day recency gate — run before Step 7, every time. Ask explicitly: "has anything happened to this company in the last 90 days?" Search "{company}" news and "{company}" incident OR outage OR breach OR layoffs OR lawsuit, plus the company's blog/status page. This is a distinct check, not a by-product of the news call — a topical query can rank an evergreen feature above a three-week-old incident.

• Anything material inside 90 days sets the outreach verdict (Step 7) and goes in handle with care, directly under the header.
• Nothing found? The explicit negative is what lets the verdict be SEND.

LinkedIn posts — Fiber ($0.04), the strongest personalization signal:

taskfuel call https://mpp.orthogonal.com/fiber/v1/linkedin-live-fetch/profile-posts \
  --method POST --body '{"identifier":"<verified slug>"}' --max-amount 0.04

Per item: caption (their own words), resharedPost (the amplified post, a separate field), postUrl (permalink), author, postedAt, engagement. Because reshares are a separate field, attribution is unambiguous — quote caption as theirs; cite resharedPost.author as "{Contact} amplified {Actor}'s post: …". Up to 50/page; continuations bill $0 via parentRequestId with unchanged filters. Authored ratio varies a lot by contact (20/38 and 42/50 observed).

⚠ The date-bucket trap — neither Fiber nor Aviato returns a post date. Both return a noEarlierThan/noLaterThan window: ~30 days for the last year, 365 days beyond.

• Write the window as a window. "He posted about hiring in March 2025" is fabrication when the source supports only "sometime in 2024–2025".
• To promote a post to a trigger event, WebFetch its postUrl and read the real date.
• No real date? Keep the item, drop the date — "undated, LinkedIn window 2024-08→2025-08". A hook survives undated; a trigger event does not.
• The deep tail is old — good for interests and common ground, useless for "why now."

Twitter/X — use the verified handle from Step 4. GET /users/by/username ($0.005) + GET /tweets/user ($0.01, ≤20 tweets). A 404/empty means the handle is wrong — go back to web search, don't keep paying against a bad handle. GET /workflows/userInsights ($0.50) is explicit opt-in only. Never call twit.sh write/action endpoints (follow/like/retweet/tweet/ delete/setProfile/bookmark) — read-only research.

Personal site/blog — GET https://x402.ottoai.services/web-extract?url=… ($0.005), clean LLM-ready markdown. It renders JS and PDFs itself, so reach for POST /api/firecrawl/scrape ($0.0126) only when web-extract returns an empty or boilerplate body. Neither may be pointed at LinkedIn — Firecrawl refuses it outright.

Writes: recent posts/talks/activity (source + date tagged), hiring signals → trigger events, interests (sourced; tagged inferred only when synthesized).

Step 6 — Shared connections & internal relationship history (free)

Neither is queryable in this stack — two honest manual-check notes, deliberate choices, not placeholders:

• Shared connections — tell the rep to check LinkedIn "mutual connections" from their own account.
• Internal relationship history — no CRM in the stack; tell them to check theirs / ask the team before cold-emailing someone a colleague knows.

Step 6.5 — Free-search completeness check (free, mandatory)

Paid structured data crowds out patient free search — measured, on the same contact, as a thinner landmine section in a cheaper run. This step is the guard.

Before Step 7, each slot holds a result or an explicit negative:

A blank fails the brief — every slot holds a result or an explicit negative.

Step 7 — Synthesize + finalize (free)

Pure reasoning over data already bought. Each item carries a citation or an explicit "(inferred)" tag.

The outreach verdict — decide FIRST; the openers inherit it

Handle-with-care must be able to overrule the openers, not sit beside them. Set one in the header:

• SEND — nothing material in 90 days; openers stand.
• SOFTEN — outreach is fine, but every opener must acknowledge the event or deliberately route around it. Say which.
• HOLD — recommend not sending yet, and say what to wait for. Still write the openers, marked for later.

The verdict is a recommendation to a human, never a refusal to do the work. Produce the full brief regardless; the rep may overrule you.

A verdict can also be about the premise — if employment_history shows the contact left the company the rep named, say so up front. The brief is still worth writing; the angle needs re-aiming.

Sections to synthesize

• Trigger events / "why now" — new role / funding / launch / hiring / recent posts, most-recent first. Every date read off a source page.
• Conversation-starter hooks — 3–5 bullets, source + date tagged.
• Suggested openers — 2–3 draft lines grounded in tagged hooks; tie one to the "why us" angle if given. Drafts, not send-as-is. Each must be consistent with the verdict — on SOFTEN/HOLD, an opener that ignores the landmine is a defect.
• Relevance angle / "why us" — omit entirely if no angle was given. Never invent a pitch.
• Possible common ground.
• Handle with care — the 90-day gate and durable landmines (public friction, stated allergies, topics to avoid). Under the header on SOFTEN/HOLD; write "none surfaced (checked YYYY-MM-DD)" rather than omitting it.
• Buying-role framing — decision-maker / champion / influencer, taken straight from seniority/departments rather than inferred → header.
• Call logistics — timezone always; phone only if the PDL escalation was bought, else Phone: not resolved.
• Sources & cost — every call, its cost, run total, balance delta. Set Status: complete.

End the run: print the full brief inline, then a last line with the saved path (Saved to: briefs/…). The rep acts on it now; the file is the durable copy.

Cost

Default run — $0.067, either entry point

Step 2 is flat with no tail — one paid call, no escalation chain. A miss costs $0 extra; an unresolvable contact stops the run rather than spending more.

Opt-in add-ons — priced, with the condition that justifies each, in ADD-ONS.md. In-line extras not listed there: ottoai extract +$0.005 (Firecrawl +$0.0126 if it returns empty).

• Steps 3–4 and Step 5's interests are free — this halved run cost and improved quality. But free ≠ optional: Step 6.5 fails the brief on a blank.
• Fiber is the largest line item at $0.04. Skip it (→ ~$0.027) only when the contact has no LinkedIn presence.
• The phone is the biggest cost lever — $0.28 for one field, 4× the whole default run. Only buy it when the rep says they will dial.