discover-analytics-patterns

Your goal is to find out how this codebase sends analytics events — not which events exist, but the specific code patterns engineers use to fire a tracking call. This output helps engineers add new events that look consistent with the rest of the codebase. It should also tell downstream skills how event names and property names are typically written in code here.

When determining naming conventions in this skill, use the following sources in strict order of preference:

1. Customer directives in .amplitude/instrumentation-agent-context.md (and any files it references), if present — explicit conventions the customer wants followed, so they override everything below.
2. Events and properties observed from the Amplitude MCP server
3. Real tracking call sites in the codebase
4. The taxonomy skill at ../taxonomy/SKILL.md

Step 0: Read repo instrumentation context

Before anything else, check for .amplitude/instrumentation-agent-context.md (repo root, or the subdirectory you're instrumenting). If it exists, read it and any repo-relative files it references. Any naming conventions, property standards, or SDK/wrapper patterns it states are customer directives — they take precedence over everything you infer below; record them and skip inference for whatever they cover. If it's absent, just continue — the instrument-events skill owns prompting the user to add one.

Step 1: Find tracking calls

Use two approaches based on what's available.

If the Amplitude MCP is connected

Inspect the connected catalog and use its current taxonomy reader to fetch a sample of event names from the project. Follow only the reader's advertised schema. When it supports caller attribution, identify this skill with the name from its YAML frontmatter. Use those results to choose a few representative non-system product events, then use its event-property read capability to inspect real property names. This is your primary naming reference.

Do not infer naming conventions from bracket-prefixed Amplitude system names such as [Amplitude], [Guides-Surveys], [Assistant], [Experiment] for either events or properties. Exclude those from pattern detection. If the MCP sample is dominated by Amplitude system names or otherwise does not provide enough evidence, fall back to codebase inference for naming.

Then search the codebase for the sampled non-system event names using Grep to locate the actual tracking call sites.

If the Amplitude MCP is not available (fallback)

Search the codebase for these signals using Grep. Cast a wide net — you can narrow down after:

Also actively look for custom analytics wrappers — a codebase often wraps the raw SDK in a utility like trackEvent(), track(), or a React hook like useAnalytics() or useTracking(). Search for these by looking for functions that call into Amplitude internally. Treat each wrapper as its own pattern, separate from the underlying SDK call, even if it ultimately calls amplitude.track() underneath. Engineers who encounter the wrapper will use it, not the raw SDK — so it's the more important pattern to document.

To find wrappers: search for files that import the Amplitude SDK, then check whether any of those files export a function or hook that other parts of the codebase import and use for tracking.

Exclude test files (.test., .spec., __tests__) and mock files unless they are the only place a pattern appears.

Step 2: Group by pattern

Two call sites use the same pattern if they share the same:

• Library/SDK/function being called
• Method name
• Argument structure (even if the event name or properties differ)

For example, these are the same pattern:

amplitude.track('Page Viewed', { page: '/home' })
amplitude.track('Button Clicked', { label: 'signup' })

But these are different patterns — always keep them separate:

amplitude.track('Page Viewed', { page: '/home' })   // direct SDK — one pattern
ampli.pageViewed({ page: '/home' })                  // Ampli typed method — different pattern
trackEvent('Page Viewed', { page: '/home' })         // custom wrapper — also a separate pattern

A custom wrapper is always its own pattern, even if it delegates to the SDK underneath. When documenting a wrapper pattern, note what it wraps (e.g., "Custom hook wrapping amplitude.track()") so engineers understand the layering.

Step 3: Resolve naming conventions

Resolve two conventions separately:

• event_naming_convention — casing, separators, word order, prefixes, and tense used for event names in instrumentation code. Examples: Title Case, snake_case, [Prefix] Action, object-first vs action-first.
• property_naming_convention — casing, separators, and common suffix/prefix patterns used for event properties. Examples: snake_case, camelCase, *_id, is_*, flat keys vs nested objects.

Use this precedence order:

1. Repo instrumentation context first. If .amplitude/instrumentation-agent-context.md (from Step 0) states an explicit event or property naming convention, it wins outright — record it and skip inference for whatever it specifies. Only fall through when the file is absent or silent on naming.
2. Amplitude MCP second. If the observed event names and property names returned by the active taxonomy reader for a few representative non-system events show a clear dominant convention, use that. Do not use bracket-prefixed Amplitude system names as naming evidence.
3. Codebase third. If the MCP evidence is unavailable, sparse, or inconsistent, infer the dominant convention from nearby, real tracking call sites in the repository. If the codebase shows multiple conventions, call out the dominant one and note meaningful local exceptions.
4. Taxonomy fallback last. If none of the above is clear enough, fall back to the taxonomy skill at ../taxonomy/SKILL.md.

Do not guess. If one or both conventions remain unclear even after checking those sources, say so explicitly.

Step 4: Output

Start with a short conventions section, then list each unique pattern.

event_naming_convention: "<from repo context file if it specifies one, else MCP if clear, else codebase, else taxonomy skill, or 'insufficient evidence'>"
property_naming_convention: "<from repo context file if it specifies one, else MCP if clear, else codebase, else taxonomy skill, or 'insufficient evidence'>"

Then, for each unique pattern, output a section in this format:

Pattern: <short descriptive name>

Description: What this pattern does and when it's typically used in this codebase (e.g., "Used throughout the React frontend for user action tracking").

Example (generalized):

// show the import(s) needed
import { amplitude } from '@/lib/analytics'

// show a representative tracking call with placeholder names
amplitude.track('Event Name', {
  propertyOne: value,
  propertyTwo: value,
})

Relevant paths:

• src/path/to/file.ts
• src/another/file.tsx

List patterns from most common (most file paths) to least common.

If two patterns are always used together (e.g., an import + a call), show them together in one example.

Step 5: Handle no results

If no tracking calls are found with any search strategy, say so clearly. Suggest that the user check whether Amplitude (or another analytics library) has been set up in the project, and offer to search for other analytics libraries (Segment, Mixpanel, PostHog, etc.) if relevant.