Account Research

Retrieve and synthesize account information from Common Room. Handles four interaction patterns: full overviews, targeted field questions, sparse data situations, and combined MCP data + LLM reasoning.

Step 0: Load User Context (Me)

Before researching any account, fetch the Me object from Common Room. This provides:

• The user's profile, title, role, and Persona in CR
• The user's segments ("My Segments")

Default all queries to the user's own segments unless the user explicitly asks for a broader view. This keeps results scoped to their territory.

Step 1: Identify the Interaction Pattern

Determine what the user actually needs before deciding how much data to fetch:

Pattern 1 — Full Overview: "Tell me about Datadog" / "Summarize cloudflare.com" → Fetch the full field set and produce a structured briefing.

Pattern 2 — Targeted Question: "Who owns the Snowflake account?" / "Is acme.io showing buying signals?" / "What's the employee count for notion.so?" → Fetch only the relevant field(s). Return a direct, concise answer — do not produce a full brief for a simple question.

Pattern 3 — Sparse Data: "Tell me about tiny-startup.io" → If Common Room has limited data for an account, say so honestly: "There is limited information available for this account." Never speculate or fill gaps with generic statements.

Pattern 4 — Combined Reasoning: Fetch structured MCP data, then layer in LLM analysis — e.g., "Stripe has 8,000 employees and is hiring heavily for AI roles. Based on your ICP of 1k–10k fintech companies, this is a strong fit."

Step 2: Look Up the Account

Search Common Room for the account by domain or company name. Exact match first; if no result, try partial match and confirm with the user before proceeding.

Step 3: Fetch the Right Fields

Use the Common Room object catalog to see available field groups and their contents. For full overviews, request all field groups. For targeted questions, request only what's relevant.

Key field groups to know about:

• Scores — always return as raw values or percentiles, never labels
• Summary research — RoomieAI output; often the richest qualitative signal
• Top contacts — sorted by score desc; use communityMemberID for full lookups

Choosing what to fetch:

Step 4: Web Search (Sparse Data Only)

Common Room is the primary data source. Do not run web search when CR returns rich data.

When CR data is sparse (Pattern 3 — few fields returned, no activity, no scores), run a targeted web search to fill gaps:

• "[company name]" news — scoped to the last 30 days
• Look for: funding rounds, acquisitions, product launches, executive changes, press coverage

If the user explicitly asks for external context or recent news, run web search regardless of data richness.

Step 5: Apply Reasoning (Pattern 4)

When the user's question invites synthesis — not just data retrieval — layer in analysis:

• Compare account data to known ICP criteria from session context
• Identify fit signals (size, industry, tech stack, hiring patterns)
• Note timing signals (funding, trial status, recent activity spike)
• Frame insights as clearly derived from data, not assumed

When the user's company context is available (see references/my-company-context.md), position findings relative to the user's value proposition and ICP.

Step 6: Produce Output

…