Generate Improvement Hypotheses

Given a play's current metrics, historical trends, and context, use the Claude API to generate ranked hypotheses for what to change to improve performance.

API Call

POST https://api.anthropic.com/v1/messages
{
  "model": "claude-sonnet-4-20250514",
  "max_tokens": 2000,
  "messages": [{
    "role": "user",
    "content": "You are an optimization agent for a GTM play. Analyze this data and generate improvement hypotheses.\n\nPlay: {play_title}\nLevel: Durable\nMotion: {motion}\n\nCurrent metrics (last 2 weeks):\n{metrics_json}\n\nHistorical trend (8 weeks):\n{trend_json}\n\nAnomaly detected: {anomaly_type} — {anomaly_description}\n\nCurrent configuration:\n{current_config_json}\n\nGenerate exactly 3 ranked hypotheses. For each:\n1. What to change (specific, actionable — e.g., 'change email subject line from X to Y', not 'improve messaging')\n2. Why this might work (based on the data)\n3. Expected impact (quantified estimate)\n4. Risk level (low/medium/high — high means it could make things worse)\n5. How to test it (specific A/B test or experiment design)\n\nRespond in JSON: {\"hypotheses\": [{\"change\": \"\", \"rationale\": \"\", \"expected_impact\": \"\", \"risk\": \"\", \"test_design\": \"\"}]}"
  }]
}

Input Requirements

• metrics_json: Current KPIs from PostHog (use posthog-custom-events fundamental)
• trend_json: 8-week trend data (use posthog-anomaly-detection fundamental)
• anomaly_type: Output from anomaly detection (drop/plateau/spike/normal)
• current_config_json: Current play parameters (email copy, targeting, cadence, etc. from CRM/automation)

Output

JSON with 3 ranked hypotheses. Store in Attio as a note on the play's campaign record. Each hypothesis becomes a candidate for the next optimization experiment.

Guardrails

• Never generate hypotheses that require budget increases > 20% without human approval
• Never suggest changes to more than 1 variable at a time (isolate for clean testing)
• If risk is "high" on the top hypothesis, flag for human review before proceeding
• Rate limit: max 1 hypothesis generation per play per week