Unjank — Firebase Performance Skills for AI Coding Agents

Agent skills for tracking and fixing screen rendering performance in Android, iOS, and Flutter apps. Works with any AI coding agent (Claude Code, Codex, etc.).

What This Does

4 skills that give any developer a complete performance workflow:

Skill	Command	What it does
perf-setup	/perf-setup	Discovers Firebase/GCP config, validates BigQuery access, runs smoke test
perf-query	/perf-query	Runs cost-optimized BigQuery queries for 30-day screen metrics
perf-dashboard	/perf-dashboard	Generates interactive HTML dashboard (worst screens, highest volume, trends)
perf-fix	/perf-fix ScreenName	Finds and fixes rendering anti-patterns in your code, safe changes first

Quick Start

1. Install the skills

Run this from your project root:

npx unjank-perf

Alternative methods:

# curl one-liner
curl -fsSL https://raw.githubusercontent.com/no-hup/unjank/main/install.sh | bash

# manual
git clone https://github.com/no-hup/unjank.git /tmp/unjank
cp -r /tmp/unjank/skills/perf-{setup,query,dashboard,fix} .claude/skills/
rm -rf /tmp/unjank

2. Restart Claude Code

Skills register as slash commands on startup. Start a new conversation after installing.

3. Run the workflow

/perf-setup              # one-time: auto-installs gcloud if needed, discovers config, validates BigQuery
/perf-query              # fetches 30-day performance data
/perf-dashboard          # opens interactive dashboard in browser
/perf-fix HomeFragment   # finds and fixes rendering issues for a screen

Prerequisites

• Firebase Performance SDK integrated in your app
• BigQuery export enabled in Firebase Console (Settings → Integrations → BigQuery)
• gcloud CLI — perf-setup will install and configure it automatically if missing
• BigQuery export enabled in Firebase Console (one-time, data takes ~48h)

Authentication (simplest path)

gcloud auth login
gcloud auth application-default login

That's it. No service accounts or key files needed for local development.

Dashboard

The dashboard shows three views:

• Worst Performing — screens ranked by rendering severity (frozen frames + slow rendering composite score), color-coded red/yellow/green
• Highest Volume — screens ranked by sample count (most-used screens)
• 30-Day Trends — app-wide frozen frame and slow rendering trends over time, with per-screen sparklines

Performance Fixer

/perf-fix uses a tiered approach to protect legacy codebases:

Tier	Agent behavior	Example
T1	Fixes immediately (safe, no behavior change)	Cache DateFormatter, add DiffUtil, hoist allocations out of onDraw
T2	Shows diff + explains trade-off, waits for approval	Async image loading, flatten layouts, add Compose remember
T3	Suggests only, developer decides	Defer SDK init, move startup work off main thread
T4	Notes for backlog, never attempts	Lock contention fixes, SDK replacement, architecture rewrites

The agent exhausts T1 before suggesting T2, and T2 before surfacing T3.

Project Structure

skills/
├── README.md
├── REQUIREMENTS.txt
├── perf-setup/
│   ├── SKILL.md
│   └── scripts/
│       ├── detect_config.sh
│       └── validate_gcloud.sh
├── perf-query/
│   ├── SKILL.md
│   ├── queries/
│   │   ├── smoke_test.sql
│   │   ├── screen_summary.sql
│   │   ├── screen_daily.sql
│   │   └── app_daily_trend.sql
│   └── scripts/
│       └── run_bq_query.sh
├── perf-dashboard/
│   ├── SKILL.md
│   ├── dashboard-template.html
│   └── scripts/
│       └── assemble_dashboard.sh
└── perf-fix/
    ├── SKILL.md
    └── references/
        ├── detection-patterns.md
        ├── fix-templates.md
        └── knowledge-base.md

How It Works Under the Hood

• No Gradle plugin, no build system dependency — pure skill files that any AI agent can follow
• BigQuery queries via bq CLI — no client libraries or SDK dependencies
• Cost-optimized — partition pruning + dry-run estimation before every query (~$0.05 for 30 days)
• Helper scripts for deterministic operations (config parsing, SQL templating, dashboard assembly)
• Agent-agnostic — scripts use dirname "$0" instead of Claude-specific variables

Analytics

The install script pings an anonymous hit counter (hits.sh) so we know how many people are using Unjank. No personal data is collected — it just increments a number.

To opt out: curl -fsSL ... | UNJANK_NO_ANALYTICS=1 bash