Add Analytics Instrumentation

Confirm: "I'll instrument the [checkout flow / PR #234 / feat/new-onboarding] — is that right?"

Step 1: Gather Code (Invoke Diff-Intake)

For PR or branch input:

• Fetch the diff or run git commands to get changed files
• Invoke the diff-intake skill logic to produce the structured YAML brief
• Output: YAML brief with new_surfaces, removed_surfaces, changed_behaviors

For file/directory input:

• Read the files directly
• Identify analytics-relevant code (UI components, handlers, API calls)
• Produce an equivalent YAML brief inline

For feature description input:

• Search git history for recent commits matching the feature
• Find the relevant files in the codebase
• Produce the YAML brief from code inspection

Validation: Confirm the brief captures what the user expected. Show a 3-bullet summary:

• "N new UI surfaces found (screens, forms, buttons)"
• "N behaviors changed in existing flows"
• "Estimated N new events needed"

Step 2: Discover Event Surfaces (Invoke Discover-Event-Surfaces)

Pass the YAML brief to the discover-event-surfaces skill logic:

• Check Amplitude for existing events (no duplicates)
• Map each new surface to candidate events
• Apply naming convention (verb_noun snake_case)
• Identify required properties per event
• Flag funnel-critical events
• Prioritize by business value (critical / high / medium / low)

Show results grouped by priority:

CRITICAL (funnel steps — must track):
  - checkout_confirmation_viewed
  - order_placed

HIGH (key engagement — should track):
  - shipping_address_submitted
  - promo_code_applied

MEDIUM (informational — nice to have):
  - order_details_expanded

Ask: "Does this list look complete? Are there events you'd add or remove before I write the specs?"

Step 3: Generate Tracking Plan (Invoke Instrument-Events)

For each approved event candidate, invoke instrument-events logic to produce the full spec:

For each event, present:

• What: Event name and trigger condition
• Why: Business question this event answers
• Where: Exact file and code location
• Properties: Full type-annotated property table
• Code: Ready-to-paste tracking call in the project's SDK style
• Test: Test case definition

Format as sequential steps an engineer can follow:

## Event 1: checkout_confirmation_viewed

WHY: Measures how many users successfully reach order confirmation
     (critical for purchase funnel conversion rate).

WHERE: src/pages/checkout/ConfirmationPage.tsx — in useEffect after order loads

WHAT TO ADD:
  useEffect(() => {
    if (!order?.id) return;
    track(Events.CHECKOUT_CONFIRMATION_VIEWED, {
      order_id: order.id,
      order_total: order.totalAmount,
      payment_method: order.paymentMethod,
      items_count: order.items.length,
    });
  }, [order]);

PROPERTIES:
  order_id (string, required): "ord_abc123"
  order_total (number, required, USD): 49.99
  payment_method (string, required): credit_card | paypal | apple_pay
  items_count (number, required): 3

Final Confirmation

After presenting all event specs:

1. Ask: "Should I implement these tracking calls directly in the codebase, or is this a spec for your team?"
2. If implementing: write the tracking calls file by file
3. If spec only: offer to export as markdown table or YAML tracking plan

Also offer: "Would you like me to add these events to your Amplitude tracking plan taxonomy?"

MCP Tools

• mcp__Amplitude__get_event_properties — check existing events before proposing new ones; match property naming to existing schema
• mcp__Amplitude__get_project_context — confirm the Amplitude project and get project ID for any schema lookups

Key Concepts

• Orchestration skill: This skill calls the logic of three sub-skills in sequence — users never need to invoke the sub-skills manually
• Tracking plan: The full output of this workflow — a document specifying every event, its trigger, and its properties
• Funnel completeness: A tracking plan is only valuable if it covers every step in the user's critical path — gaps break conversion analysis
• Event-by-event review: Presenting results event-by-event (not all at once) helps users catch mistakes before implementation
• Priority triage: Not all events are equal — critical funnel events ship first; lower-priority events can be a follow-up

Output Format

Three-phase output:

Phase 1 — Brief Summary (after diff-intake):

Found N changes across X files.
New surfaces: <list>
Changed behaviors: <list>
Estimated events needed: N

Phase 2 — Event Candidate Review (after discover-event-surfaces):

Grouped by priority (CRITICAL / HIGH / MEDIUM / LOW)
Await user confirmation before proceeding

Phase 3 — Implementation Specs (after instrument-events):

One section per event with: WHY / WHERE / WHAT TO ADD / PROPERTIES
Followed by: offer to implement or export