# Company Deep Dive > Use this skill when the ask is about one specific company: "tell me about [company]", "research [company]", "what does [company] do", "who is [company]", "look up [company]", "company deep dive", "due diligence on [company]", "background on [company]", "dig into [company]", "analyze [company]" — or when evaluating a company as a sales target, partner, or investment. Research the live web instead of answering from memory: funding, leadership changes, and product launches move faster than any knowledge base, so run it even for well-known companies. Produces a sourced 360-degree report covering funding, leadership, product and technology, market position, news, and strategic outlook — with a date and URL on every claim. - Skill: `swan-gtm/company-deep-dive` (Agent Skill, multi-file: 5 files) - Install (CLI): `npx skillmds@latest add swan-gtm/company-deep-dive` - Raw SKILL.md: https://api.skillmd.com/api/skills/swan-gtm/company-deep-dive/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Docs & Writing - Author: swan-gtm (https://skillmd.com/u/swan-gtm) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/swan-gtm/company-deep-dive --- Runs when someone needs real, current intelligence on one specific company — not a vibe from stale general knowledge. Produces a 360-degree company report: quick assessment up top, then funding, leadership, product, market position, and news, closed by a strategic outlook — every factual claim carrying a date and a source URL. ## Scope the run Three decisions before any research: 1. **Which company, exactly.** If the name is unambiguous ("research Stripe"), run two quick web searches — `"[Company] official site"` and `"[Company] company overview"` — to lock the legal name and primary domain, and confirm in one line: "Researching **[Company]** ([domain])". If the name collides ("research Mercury" — bank? auto brand?), ask one clarifying question with the top candidates. If no company was named, ask which one. 2. **Full or quick.** Default to the full deep dive. If the ask says "quick overview", "brief", or "summary", run quick mode: search-snippet research only, skip the deep-extraction step, produce a shorter report. 3. **First run or refresh.** Keep a running file per company in your workspace (format in `references/company-memory.md`). If a file for this company exists, load it and run in **refresh mode**: tell the reader "I have prior research on [Company] from [date] — refreshing with what's new", pass the known facts into every research thread so they hunt only for new signals, and lead the report with a "What's new since [date]" section. No file → full mode across all dimensions. ## The play ### 1. Research five dimensions in parallel Fan out one research thread per dimension — run them concurrently, not sequentially; a deep dive is five narrow investigations, not one broad search: | Dimension | Hunting for | |---|---| | **Funding & financials** | Rounds, total raised, valuation, revenue signals, investors, IPO chatter | | **Product & technology** | Core products, launches, tech-stack signals, engineering blog, open source | | **Leadership & team** | Founders, C-suite, key hires and departures, headcount, culture signals, executive interviews | | **News & events** | Press coverage, partnerships, acquisitions, awards, conferences, social chatter | | **Market position** | Competitors, market share, analyst coverage, customer reviews, case studies | The per-dimension query patterns — including which queries to date-filter, which to run against the company's own domain, and which review and analyst sources to target — are in `references/dimension-queries.md`. Read it before fanning out. Two searches run outside the fan-out, first, to frame everything: an "about" search restricted to the company's own domain, and a search for the company's profile on the major business registries (Wikipedia, Crunchbase, Pitchbook). These give founding date, HQ, headcount, and mission — the skeleton the dimensional findings hang on. Every finding comes back as a structured signal — description, article date, event date, source URL, source type — not prose. The distinction between when a page was published and when the event actually happened is where most research goes wrong; the dating rules, source-quality hierarchy, and the drop rules for stale or unverifiable signals are in `references/signal-dating.md`. Read it before recording the first signal. If a research thread fails or comes back empty, rerun its queries directly — never leave a dimension blank without trying twice. ### 2. Extract the pages that matter From all findings, pick the **5–8 most informative URLs** and extract each full page (in parallel). Prioritize, in order: funding announcements with specific amounts; official product or feature pages; executive interviews, podcasts, or conference talks; in-depth analyst or journalist profiles; the company's own about/team page. Full pages yield the numbers, direct quotes, and context that snippets flatten. Quick mode skips this step entirely and reports from snippets. ### 3. Synthesize the report Assemble the 360-degree report — exact section order, per-section content rules, and the refresh-mode variant are in `references/report-format.md`. The non-negotiables: - Lead with a 2–3 sentence **Quick Assessment** — what the company is, where it stands, the one thing that matters most right now. Most readers stop there; make it earn the stop. - Every claim dated and sourced. A dimension with nothing found says "no public data found" — never speculation dressed as findings. - Confirmed facts and inferred signals are labeled differently. - Direct executive quotes (from interviews, earnings calls, talks) beat paraphrase wherever found. ### 4. Update memory and offer next steps Write the full report to your workspace, and update (or create) the company's running memory file: structured key facts plus dated signal entries, so the next run dedups against it instead of re-reporting old news. Format and dedup rules in `references/company-memory.md`. Then offer the natural follow-ups: go deeper on one dimension, compare side-by-side with another company, set the company up for ongoing tracking over time, or research the specific people ahead of a meeting. ## What this is not - **Not multi-company monitoring.** This goes deep on ONE company; tracking a set of competitors over time is a different, recurring motion. - **Not people research.** It profiles companies, not the individuals attending your next meeting. - **Not investment advice.** It is intelligence gathering from public sources — say so if asked. - **Not live monitoring.** The report is point-in-time; rerun it (in refresh mode) when currency matters. ## What good looks like - The best operators check the **event date, not the article date**, before calling anything new — a syndicated recap published this week about a funding round from last year is the single most common way a "fresh" report embarrasses its author. - The second tell: single-sourcing a big claim from an aggregator. Funding amounts, M&A, and leadership changes get corroborated against a primary or wire source before they appear — every time. - The mediocre version is a Wikipedia-shaped summary: undated claims, no URLs, "recent developments" that are neither, dimensions padded with speculation where the search came up empty, and a refresh that re-reports everything already known instead of only what changed. - Good output test: a reader can click through any claim to its source, the Quick Assessment could stand alone as a briefing, and a refresh run reads like a changelog — short, dated, only genuinely new or updated signals. ## Rules - MUST research the live web — never answer about a specific company purely from prior general knowledge, however famous the company. - MUST attach a date and source URL to every factual claim in the report. - MUST corroborate funding, M&A, and leadership-change claims sourced from derivative or aggregator sites against a primary or major source before reporting them (see `references/signal-dating.md`). - MUST write "no public data found" for an empty dimension — NEVER fill the gap with speculation. - MUST dedup against the company's memory file in refresh mode and surface only NEW or UPDATED signals. - NEVER include a signal whose event date cannot be established — dropped means dropped, not footnoted as "background". - NEVER present an estimate or inference with the same confidence as a confirmed fact — label which is which.