Account Research

Account Research

Turn one target company domain into a decision-ready dossier: the resolved entity tree, the buyers across every buying unit, the live signals, and the angle to enter on. Breadth lists come from list-building; this skill goes deep on a single account.

The engine is Extruct Deep Research: you write one brief describing the account and the decision it supports, Extruct runs 10/25/50 research agents over the web and its index, and returns a single fully-cited report. You assemble that report into the dossier. There is no table or column setup — one brief, one report.

Related Skills

list-segmentation → account-research → email-search → campaign-sending

Reads the company context file for ICP and target roles. Hands a buyer map plus signals to the outreach skills. Optional email/phone enrichment is delegated to email-search.

Engine

Extruct Deep Research API: POST /v1/deep_research_tasks. The endpoints, auth, depth levels, the poll gotcha, and the report shape are in references/deep-research-engine.md — read it first.

Inputs

Workflow

Step 1: Confirm access

~/.ext auth user — must be the project account (danny@extruct.ai for extruct-research). Note available_credits (Deep Research bills 1 credit per executed agent, up to 50 at xhigh). Capture the slug (kebab-case), the seller, the region focus, and the decision the dossier must support — they all go into the brief.

Step 2: Compose the brief

One brief string (<=20,000 chars) carries the seller context, the target, the decision, the region, and a numbered list of every layer the dossier needs. Bake in the ICP / target roles from the context file. Ask for a source URL per claim and to flag low-confidence items and gaps. Cover:

1. Overview — what they do, HQ, founding year, headcount, mission.
2. Entity tree (the differentiator) — parent → divisions / subsidiaries / operating units, deduped, with a verified count of legal entities and countries. A flat database shows one record; the real account is a tree, and every operating unit is a separate buyer.
3. Numbers — funding rounds (amount, date, investors), valuation, revenue/ARR + growth, customers/locations, retention. Reconcile disputed figures; label estimates.
4. Signals — current OPEN and recently FILLED roles in the target functions; leadership changes; dated news (M&A, launches, partnerships, funding, expansion); the tech stack (CRM + the incumbent tooling the offer sits beside); what each division is prioritizing.
5. Decision-makers — named buyers across the entity tree, not just the parent, in the target functions, with exact titles and LinkedIn URLs.
6. Entry angle + honest verdict — the buying committee, the wedge, a concrete first move, and an objective read on whether this is a strong fit (say so plainly even if it is weak).

Step 3: Run it

Create the task at the chosen depth, then poll to done (in the background; mind the strict=False JSON gotcha). See the engine doc.

Step 4: Assemble the dossier

Read report.markdown + report.sources. Tag each signal to the buyer who owns it with the tier model, then cluster the linked signals into <=4 entry angles — see references/methodology.md. Write findings to claude-code-gtm/accounts/{slug}/:

• overview.md — what they do, the resolved entity tree, the footprint
• buyer-map.md — decision-makers by buying unit, with the angle per person
• signals.md — dated, sourced events and what each implies
• account-brief.md — the distilled brief: the angle, the people, the why-now

Save the raw cited report (report.md + sources.md) alongside as provenance. Emit a CRM-syncable account map (companies + people + signals) for campaign-sending or the CRM. Keep the report's honest verdict — do not inflate a weak-fit into a slam-dunk. A visual dashboard is out of scope; the dossier is the deliverable.

Step 5: (Optional) enrich buyer emails

Delegate the named buyers to email-search for verified emails and phones. Drop anyone whose current employer resolves off-target.

Ground Rules

• Extruct is the engine. The whole account — entity tree, signals, people, verdict — comes from one Deep Research report. Don't reach for a separate signal vendor.
• Source every claim. The report cites [n]; carry the source URL through to the dossier. Confidence-gate over guessing; keep single-source / low-confidence items labelled as such. Never fabricate to fill the dossier — a smaller, fully-sourced result beats a padded one.
• Verify before claiming. Read the actual report, not memory. If the task failed, or sits in in_progress with agent/source counts that never advance, STOP and say so rather than backfilling with web search.
• Stay vertical-agnostic. Read who and what to target from the context file; don't assume an industry.

Reference

• references/deep-research-engine.md — the Extruct Deep Research API workflow (endpoints, depth levels, brief template, poll gotcha, report shape).
• references/methodology.md — the tier model for linking signals to buyers, entry-angle rules, and the dossier shape.