Success Criteria Extraction

Given a discovery call transcript, deal context, and product capabilities, use the Claude API to extract prospect-stated goals and generate specific, measurable success criteria. The output is a structured set of criteria the prospect can agree to, each with a metric, target, timeline, and measurement method.

API Call

POST https://api.anthropic.com/v1/messages
Headers:
  x-api-key: {ANTHROPIC_API_KEY}
  anthropic-version: 2023-06-01
  content-type: application/json

Body:
{
  "model": "claude-sonnet-4-20250514",
  "max_tokens": 3000,
  "temperature": 0.2,
  "messages": [{
    "role": "user",
    "content": "You are a success criteria analyst for a B2B sales process. Analyze the following inputs and generate specific, measurable success criteria that the prospect can agree to.\n\nDISCOVERY TRANSCRIPT:\n{transcript}\n\nDEAL CONTEXT:\n- Company: {company_name}\n- Industry: {industry}\n- Headcount: {headcount}\n- Current tools: {current_tools}\n- Deal size: {deal_size}\n- Champion: {champion_name}, {champion_title}\n\nPRODUCT CAPABILITIES:\n{product_capabilities}\n\nHISTORICAL SUCCESS RATES (similar deals):\n{historical_success_json}\n\nExtract and generate success criteria. For each:\n1. The prospect's stated goal (use their exact words when possible)\n2. A specific metric to measure it\n3. A quantified target (use the prospect's number if stated; otherwise derive a conservative estimate from their pain description)\n4. Timeline to achieve it\n5. Who measures it (prospect-side stakeholder)\n6. Baseline value (current state, if mentioned or estimable)\n7. Achievability score (0-100): how likely this criteria is to be met based on historical data and product capabilities. Score 80+ = safe to commit. Score 50-79 = achievable with effort. Score <50 = risky, flag for expectation management.\n8. Category: efficiency, revenue, cost_savings, quality, time_to_value, adoption\n\nReturn JSON:\n{\n  \"success_criteria\": [\n    {\n      \"prospect_goal\": \"\",\n      \"metric\": \"\",\n      \"target\": \"\",\n      \"timeline\": \"\",\n      \"measured_by\": \"\",\n      \"baseline\": \"\",\n      \"achievability_score\": 0,\n      \"category\": \"\",\n      \"source_quote\": \"\",\n      \"measurement_method\": \"\"\n    }\n  ],\n  \"missing_info\": [\"list of things the prospect didn't mention that would strengthen the criteria\"],\n  \"risk_flags\": [\"any criteria that seem unrealistic or unmeasurable\"],\n  \"recommended_follow_up_questions\": [\"questions to ask in the next call to refine criteria\"]\n}"
  }]
}

Input Requirements

• transcript: Discovery or follow-up call transcript from Fireflies or Gong (use fireflies-transcription or gong-call-recording fundamental)
• company_name, industry, headcount, current_tools, deal_size, champion_name, champion_title: From Attio deal record
• product_capabilities: A JSON summary of what your product can measurably deliver
• historical_success_json: Aggregated data from past deals showing which criteria were achieved and at what rates (query from Attio)

Output

JSON with structured success criteria. Each criterion includes an achievability score that guides whether to commit to it in the mutual success plan.

Post-Processing

1. Store each criterion in Attio as a custom attribute on the deal record
2. Fire PostHog event success_criteria_extracted with properties: deal_id, criteria_count, avg_achievability_score, categories
3. Flag any criterion with achievability_score < 50 for founder review before including in the mutual success plan

Guardrails

• Never generate success criteria the product cannot measure. Every criterion must have a measurement_method that maps to an observable data point.
• Never inflate achievability scores. When in doubt, score lower — it is better to under-promise.
• If the transcript contains fewer than 2 identifiable goals, return missing_info noting that additional discovery is needed before criteria can be defined.
• Cap the number of criteria at 5. If more goals are identified, rank by achievability and prospect emphasis, and include the top 5.

Alternatives

| Tool | Method | Notes | |------|--------|-------| | Claude (Anthropic) | Messages API | Best structured extraction quality | | GPT-4 (OpenAI) | Chat Completions API | Alternative LLM | | Gemini (Google) | Generative AI API | Alternative LLM | | Gong AI | Deal insights | Built-in extraction, less customizable | | Manual | Spreadsheet template | Fallback for enterprise deals requiring human judgment |