Research

Instructions

Research is target-driven and doesn't require org config to function — there's no state-hint paragraph and no Setup.md. The agent just routes to the right sub-page and runs.

Step 1 — Parse the ask, pick the sub-page

Route based on what the user is asking about. Load the sub-page via swan-get-skill.

If the ask is ambiguous (e.g. just a company name with no further context), default to Company. If the ask explicitly chains multiple facets ("research acme — company plus the buying committee"), it's fine to load two sub-pages — but always pick a primary.

Step 2 — Pick the depth

Most asks want a one-paragraph answer; some want a dossier. Read the user's phrasing:

• "/research acme" / "what does acme do" → brief read (3–5 lines, the essentials).
• "deep brief on acme" / "full dossier" / pre-meeting prep with a VP+ → run the sub-page end-to-end.
• "quick — is acme an ICP fit" → micro-answer, one paragraph.

Default to brief. Scale up only when the user signals depth, or when the situation (board prep, founder intro, executive outreach) demands it. Over-researching when a sentence would do is the most common failure mode.

Step 3 — Load the sub-page and execute

The sub-page owns the procedure for its facet. Follow it.

Across all sub-pages, the same cheap-before-expensive pattern applies:

• check what Swan already has (swan-search-companies, swan-get-memory, prior briefs)
• check the CRM if connected
• reach for free preview tools (swan-fetch-scraped-url, swan-fetch-businesses, swan-website-traffic, swan-fetch-business-events) before paid enrichment
• enrich (swan-enrich-company, swan-enrich-contact) only when the cheap signals left a real gap

Step 4 — Output shape

Default to a structured short summary, not raw scraped data. Every sub-page ends with a composed block — follow its shape. Always:

• Cite sources inline (e.g. "via swan-website-traffic", "from LinkedIn post 2026-05-03", "press").
• Flag confidence on soft signals (high / medium / low based on signal density).
• Surface the highest-signal finding first. Don't bury the lede in a wall of context.
• Tag the company in Swan via swan-update-company when the research produced a durable fact worth reusing (financial state, tech stack, motion type) so the next play inherits it.

What good looks like

• Right depth for the ask. A one-paragraph answer when a paragraph is enough; a full dossier when the stakes justify it. Reading the room beats running every facet.
• Sources cited, confidence flagged. Every claim points to a tool result, a URL, or a CRM record. Soft signals (inferred revenue, motion type, runway) carry an explicit "est" or "looks like" — never stated as fact.
• Lede first. The user sees in 30 seconds what they came for. The strongest hook is line one of the output, not buried below firmographics.
• Knows when to stop. If business events covered the surface area, no more web fetches. If a person has no public profile, the dossier says so and ends.

What gets overlooked:

• Over-researching. Running every sub-page when only one matters. Running the deep flow when a brief was the right call.
• Treating all signals as equal-confidence. Headcount × industry-avg as a revenue figure stated like a fact. Inferred motion stated without hedge.
• Raw dumps. Pasting 40 LinkedIn posts, 200 keywords, or the full enrichment payload into the chat. Synthesize, then summarize.

Rules

• MUST pick a single sub-page per call unless the user explicitly asks for multi-facet research.
• MUST cite sources inline and flag confidence on soft signals.
• MUST check the free preview tools (swan-fetch-scraped-url, swan-fetch-businesses, swan-website-traffic, swan-fetch-business-events) before chaining paid enrichments.
• MUST tag the company in Swan with any durable fact (motion type, financial state, detected stack, ICP fit) the research produced.
• NEVER produce a raw dump. Always synthesize into the sub-page's composed output shape.
• NEVER invent quotes, beliefs, financials, or priorities not present in the source.
• NEVER chain swan-enrich-company or swan-enrich-contact calls without first checking what Swan and the CRM already have.
• If a tool result is truncated, read the JSON from files/tool-outputs/<toolName>_<callId>.json in swan-execute-code and summarize from there.

Specificity sub-pages this skill will grow

Shipped today:

• Company — general company research (the default landing for "/research acme")
• Person — full dossier on a named person
• AccountTeam — buying-committee mapping inside one account
• GTMmotion — how the target sells today
• News — last-30-days signal scan
• Domain — digital footprint, traffic, SEO posture
• Financial — funding history, ownership, runway, pressure
• TechStack — detected tools by category
• CompetitorMentions — public mentions of a named competitor

Accreted over time as use cases sharpen:

• AcquisitionTargets — research for M&A or partnership shortlists
• CustomerHealth — research framing for an existing customer (expansion / churn risk)
• IndustryDeepDive — a whole vertical, not one account
• Investor — partner / fund research
• Board — board prep on a director or chair