How Scoring Works

Every action in Claude Code earns base points, multiplied by active bonuses. Here is the full formula.

Privacy Guarantee

hackers.baby is designed with privacy as a core principle. We collect only the minimum anonymous metrics needed to compute your score. Your code, prompts, and file system are never accessed.

🚫No prompt text or message content is ever collected
🚫No file paths, file names, or file content is transmitted
🚫No command text or arguments are ever sent
🚫No code, diffs, or repository information is collected
✅Only anonymous counts and numeric metadata (e.g. character length, token counts)
✅Only tool names (e.g. "Bash", "Edit") — never tool input/output
✅Only agent type labels (e.g. "Explore", "Plan") — never task descriptions
✅All hooks are open source — inspect them at any time

Base Points

ActionPointsPer
Tool Call
5each
File Created
15each
File Edited
10each
Session Completed
Full coding session — big reward for deep work
100each
Prompt Submitted
Each message you send to Claude (only count + length tracked)
15each
Subagent Spawned
Spawning sub-agents for parallel work (only agent type tracked)
40each
Context Compacted
Pushing Claude to its limits until context compacts
30each
Response Completed
Each completed Claude response turn
10each

Multipliers

Streak (3+ days)1.5x
Streak (7+ days)2.0x
Streak (30+ days)3.0x
Quality Commits1.5x
Language Diversity (3+)1.2x

Anti-Gaming Caps

Session Base Point Cap20,000 pts
Tool Calls Per Hour Cap4,000
Min Session Tokens2,000

Formula

final_score = min(base_points, 20000) * streak_multiplier * quality_multiplier * diversity_multiplier

What Data We Collect

Tool Usage

Tool name only (e.g. "Bash", "Edit", "Read"). We never see what the tool did, which files it touched, or what commands it ran.

Prompts

Count of prompts and character length only. Your actual prompt text is never transmitted — it stays entirely on your machine.

Subagents

Agent type label only (e.g. "Explore", "Plan"). No task descriptions, prompts, or results.

Sessions

Session start/end and context compaction events. No transcript content.