Generate Workflow Optimization Suggestions

Given a user's behavioral data (event sequences, feature usage patterns, time-on-task metrics), use the Claude API to generate specific, actionable workflow improvement suggestions tailored to that user's actual usage.

API Call

POST https://api.anthropic.com/v1/messages
{
  "model": "claude-sonnet-4-20250514",
  "max_tokens": 1500,
  "messages": [{
    "role": "user",
    "content": "You are a workflow optimization agent for a SaaS product. Analyze this user's behavior data and generate improvement suggestions.\n\nProduct context:\n{product_description}\n\nAvailable features the user has NOT discovered:\n{undiscovered_features_json}\n\nUser behavior data (last 30 days):\n- Event sequence (most recent 50 actions): {event_sequence_json}\n- Feature usage frequency: {feature_usage_json}\n- Average time per workflow completion: {workflow_times_json}\n- Repeated action patterns (same 3+ step sequence done 5+ times): {repeated_patterns_json}\n- Errors or retries: {error_events_json}\n\nPower user benchmark (top 10% users):\n- Their most-used features: {power_user_features_json}\n- Their average workflow time: {power_user_times_json}\n- Features they use that this user does not: {feature_gap_json}\n\nGenerate exactly 3 ranked suggestions. For each:\n1. What to change (specific action: 'Use keyboard shortcut Cmd+K instead of clicking Menu > Search', not 'try using shortcuts')\n2. Why this helps (quantified: 'saves ~15 seconds per search, you search 20x/day = 5 min/day saved')\n3. How to do it (step-by-step, 3 steps max)\n4. Confidence (high/medium/low based on data strength)\n5. Category (efficiency, discovery, automation, collaboration)\n\nRespond in JSON: {\"suggestions\": [{\"action\": \"\", \"benefit\": \"\", \"steps\": [\"\"], \"confidence\": \"\", \"category\": \"\"}]}"
  }]
}

Input Requirements

• product_description: Brief description of the product and its core workflows
• undiscovered_features_json: Features available but not yet used by this user (from PostHog feature tracking)
• event_sequence_json: Last 50 user actions from PostHog (use posthog-user-path-analysis)
• feature_usage_json: Feature usage counts from PostHog custom events
• workflow_times_json: Time between workflow start and completion events
• repeated_patterns_json: Repeated action sequences detected from path analysis
• error_events_json: Error events, retries, or undo actions from PostHog
• power_user_features_json: Aggregate feature usage from the Power Users cohort (use posthog-cohorts)
• power_user_times_json: Benchmark workflow times from the Power Users cohort
• feature_gap_json: Features used by power users but not by this user

Output

JSON with 3 ranked suggestions. Each suggestion must be:

• Specific: references actual features and UI elements by name
• Quantified: includes estimated time or effort savings based on the user's actual usage frequency
• Actionable: the user can implement the suggestion in under 2 minutes

Store generated suggestions in PostHog as a person property last_workflow_suggestions (JSON string) and in Attio as a note on the contact record.

Guardrails

• Never suggest features the user's plan does not include — check plan-level feature access first
• Never suggest more than 3 suggestions at once (cognitive overload reduces adoption)
• Rate limit: max 1 suggestion generation per user per week (prevent notification fatigue)
• If the user has fewer than 20 actions in the last 30 days, return insufficient_data — not enough behavior to analyze
• If the user is already in the top 20% by efficiency metrics, return already_optimized with a single "advanced tip" instead of 3 suggestions
• Cache suggestions: do not regenerate if the user's behavior pattern has not changed significantly (>20% shift in feature usage distribution)

Cost Estimate

At ~1,500 tokens output per user per week using Claude Sonnet 4.6:

• 100 users/week: ~$0.50/week ($2/month)
• 1,000 users/week: ~$5/week ($20/month)
• 10,000 users/week: ~$50/week ($200/month)

Use prompt caching for the system prompt and product context (90% savings on input tokens).