Agent skills · Claude Code · Codex · Python stdlib · MIT

Your coding agent, working as a game analyst

Install the kit into a game project and ask in plain words. The agent pulls the raw AppMetrica events into a local SQLite database, learns what this game's events mean, checks its own numbers and publishes a report you can forward to anyone — the same seven blocks every time.

Open a live example report Star on GitHub

The first screen of a full review: the request as the title, seven numbered blocks, key metrics with build versions as columns
A full review on made-up data: the title is the request, key metrics by build side by side, estimates from a device sample marked ≈.

How it works

  1. Installpy install.py <project> copies seven skills for Claude Code and Codex and a small engine into the game repository. Nothing to pip install.
  2. Connect“Connect AppMetrica.” The agent finds the SDK key in the code, uses your read-only token and pulls the data in the background — sampling devices when the game is big.
  3. Model“Build the data model.” It reads the catalog of events and the game's analytics code, types the parameters, filters developer devices, defines sessions and new players.
  4. Ask“Make a full review of the last two weeks.” “Where do new players leave?” “Did 1.2 help?” Every answer is a versioned report, and the findings stay in the project for the next session.

One structure for every report

A reader and the next agent find the same thing in the same place. The builder refuses a report whose structure or links are broken.

1

Summary

Key metrics with build versions as columns — D1/D3/D7, playtime, sessions, ads per player, ARPU, revenue — and short answers to what you asked.

2

Product

Retention, drop-off by loading and tutorial steps and by every 30 s of play, core actions by their number in the session, leaving after negative events, economy.

3

Revenue

Ads by format, placement and network; eCPM; purchases.

4

Technical

Technical problems and analytics defects: duplicates, empty fields, sources that disagree — and which metric each one biases.

5

Countries

The country split in one place.

6

Hypotheses

H1, H2… each resting on linked numbers, with a test and a threshold, and a confidence.

7

Recommendations

Each linked to a hypothesis, with the expected effect and how to measure it.

Hypotheses with confidence marks and recommendations, each linked to its hypothesis
Facts are kept apart from interpretation; every number in the answers, hypotheses and recommendations appears in a table or chart above.

Live examples on made-up numbers: full review · question report · report index with versions

What's inside

Seven skills

  • analytics-kit — entry point and command sheet
  • analytics-connect-appmetrica — key, token, first pull
  • analytics-data-model — typed events, devices, sessions
  • analytics-project-audit — what the code tracks vs what arrives
  • analytics-import-external — Google Ads, AdMob, GA4, Play Console exports
  • analytics-research — question → data → hypotheses → independent check
  • analytics-report — the fixed structure, charts, versions

An engine you can read

  • Python standard library only, Windows first class
  • AppMetrica Logs API by date ranges, detached background pulls
  • Device sampling and a disk guard for games with 500k+ events a day
  • Standard tables: key metrics, drop-off, ordinals, leaving
  • Self-contained HTML reports, light and dark, readable on a phone
  • Offline selftest on synthetic data; tested end to end with real agents

Quick start

Python 3.11+, git, Claude Code or Codex, and a game that already sends events to AppMetrica.

git clone https://github.com/Trafalgardi/game-analytics-kit.git <kit>
py <kit>/install.py <project>           # python3 on macOS/Linux

# a read-only AppMetrica token, once for all projects:
#   ~/.config/game-analytics-kit/.env  →  APPMETRICA_TOKEN=...

cd <project> && claude                   # or: codex
> Connect AppMetrica and pull the last two weeks.
> Build the data model.
> Make a full review of the game for the last two weeks.

The full guide — token steps, big games, updating, security, troubleshooting — is in the README.