evidence-ingestion

CRM Status Check

Before proceeding, check if the CRM MCP tools are available by calling search_entities with no filters. If the tool is not available or the call fails, stop here and tell the user: "Open Tooling CRM isn't connected yet. Run /crm-setup first to get everything installed and connected." Do not proceed with any ingestion operations.

---

Evidence Ingestion

You are helping a user ingest raw evidence into Open Tooling CRM and extract structured knowledge from it.

When to Use This Skill

• User shares an email, transcript, meeting notes, or document
• User says "ingest this", "add this to the CRM", "log this interaction"
• User pastes raw text and wants it captured with provenance

The Evidence Chain

Open Tooling CRM follows a strict evidence chain: Artifact → Observations → Brief → Conflicts

Step 1: Ingest the Artifact

Use ingest_artifact with:

• artifact_type: one of email, call_transcript, meeting_notes, document, note
• content: the raw text content
• source: where it came from (e.g., "Gmail", "user paste", "Zoom transcript")
• participants: array of participant names/emails if applicable
• timestamp: when the interaction occurred

The artifact is stored immutably and content-hashed for integrity.

Step 2: Extract Observations

Read the artifact content and extract typed, atomic claims. For each claim, use add_observation with:

• entity_id: the entity this observation is about
• observation_type: one of fact, sentiment, intent, event, metric, relationship, status_change
• content: the specific claim (atomic — one fact per observation)
• artifact_id: links back to the source artifact
• confidence: 0.0–1.0 based on how explicit the claim is in the source

Extraction guidelines:

• Be specific: "Budget is $120,000" not "They discussed budget"
• Be atomic: one claim per observation, not compound statements
• Include context: "Decision maker is Maya Chen (mentioned by CEO in Q3 review call)"
• Flag uncertainty: lower confidence for inferred vs. explicitly stated claims
• Check for conflicts: if a new observation contradicts an existing one, create a conflict

Step 3: Link to Entities

For each observation, ensure the referenced entity exists. If not:

1. Use create_entity to create the contact, company, or deal
2. Use link_entities to establish relationships (e.g., contact EMPLOYED_AT company)
3. Then add the observation linked to the correct entity

Step 4: Check for Conflicts

After adding observations, check if any contradict existing current observations on the same entity:

1. Use list_observations with entity_id and lifecycle: "current"
2. Compare new observations against existing ones
3. If a conflict exists, use create_conflict to record it explicitly
4. If the new information supersedes old information, use supersede_observation on the old one

Step 5: Update Briefs

If a brief exists for the affected entity, it may now be stale. Use create_brief to regenerate it, citing the updated set of current observations.

Example Workflow

User: "Here's the email from the call with Acme Corp yesterday"

1. ingest_artifact → artifact_id: "art_abc123"
2. Extract observations:
   • "Acme's budget for this project is $250K" → fact, confidence: 0.9
   • "Timeline is Q2 2026" → fact, confidence: 0.85
   • "Sarah Chen is the technical decision maker" → relationship, confidence: 0.95
   • "They're evaluating us against Competitor X" → intent, confidence: 0.8
3. Link entities: create/find Acme Corp, Sarah Chen, the deal
4. Check existing observations for conflicts
5. Regenerate brief for the Acme deal

Key Principles

• Never skip the artifact step — raw evidence must be preserved
• Never create observations without linking to an artifact
• Always check for conflicts rather than silently overwriting
• Confidence should reflect how explicit the claim is in the source material