singularity-claude

Skills that evolve themselves.
A self-evolving skill engine for Claude Code — create, score, repair, and crystallize skills through autonomous recursive improvement loops.

The Problem

Claude Code skills are static. You write them once, and they stay exactly as they are — even when they fail, produce subpar output, or encounter new edge cases. There's no feedback loop, no way to know which skills are working well, and no mechanism to improve them over time.

You end up with a growing pile of skills where some work great, some are mediocre, and some are silently broken. The only way to fix them is manual review.

The Solution

singularity-claude adds a recursive evolution loop to your skills:

graph TD
    Goal["🎯 Goal:<br/>Set up API proxy endpoints"]
    Goal --> HasSkill{Skill exists?}

    HasSkill -->|No| Gap["🔍 Gap detected:<br/>recurring pattern"]
    Gap --> Create["⚙️ /singularity-create\nGather requirements\nGenerate SKILL.md\nRegister + init scoring"]
    Create --> Execute

    HasSkill -->|Yes| Execute["⚡ Execute skill"]
    Execute --> Score["📊 /singularity-score\nDispatch assessor agent\nRate 5 dimensions · 0-100"]
    Score --> Maturity["🔄 Auto-update maturity\ndraft → tested → hardened"]

    Maturity --> Low{Avg < 50\n2+ runs?}
    Low -->|Yes| Repair["🔧 /singularity-repair\nAnalyze telemetry\nRewrite weak dimensions\nBump version · re-test"]
    Repair --> Execute

    Low -->|No| High{Avg ≥ 90 · 5+ runs\nhardened · edge cases?}
    High -->|Yes| Crystal["💎 /singularity-crystallize\nGit tag + lock version"]
    Crystal --> Done["✅ Production-grade skill\nImmutable · rollback-safe"]

    High -->|No| Loop["🔁 Keep using + scoring"]
    Loop --> Execute

    style Goal fill:#1d3557,color:#fff,stroke:#0d1b2a
    style Gap fill:#e9ecef,color:#000,stroke:#adb5bd
    style Create fill:#457b9d,color:#fff,stroke:#1d3557
    style Execute fill:#2a9d8f,color:#fff,stroke:#1a7a6e
    style Score fill:#e76f51,color:#fff,stroke:#c1440e
    style Maturity fill:#264653,color:#fff,stroke:#1d3540
    style Low fill:#fff3cd,color:#000,stroke:#ffc107
    style High fill:#d4edda,color:#000,stroke:#28a745
    style Repair fill:#f4a261,color:#000,stroke:#e76f51
    style Crystal fill:#2d6a4f,color:#fff,stroke:#1b4332
    style Done fill:#1b4332,color:#fff,stroke:#0d2818
    style Loop fill:#a8dadc,color:#000,stroke:#457b9d

Tip: Click the expand button (↔) on the diagram to see the full interactive flow.

Skills are scored after every execution. Low scores trigger automatic repair. High scores lead to crystallization — a locked, battle-tested version you can trust. Every step is logged for full auditability.

No external dependencies. No SmythOS, no OpenTelemetry collector, no Docker. Pure Claude Code.

What's Inside

Skills (7)

Agents (2)

Scripts (2)

Installation

From the official Claude Code plugin registry

Requires Claude Code CLI

Inside a Claude Code session, run:

/plugin marketplace add Shmayro/singularity-claude
/plugin install singularity-claude

From npm

npm install -g singularity-claude

Then inside a Claude Code session:

/plugin marketplace add $(npm root -g)/singularity-claude
/plugin install singularity-claude

From source

git clone https://github.com/shmayro/singularity-claude.git
cd singularity-claude

Then inside a Claude Code session:

/plugin marketplace add .
/plugin install singularity-claude

Quick Start

# 1. Start a new Claude Code session — singularity loads automatically

# 2. Create your first skill
/singularity-create

# 3. Use the skill, then score it
/singularity-score

# 4. Check the dashboard
/singularity-dashboard

That's it. The evolution loop starts running from the first score.

How It Works

The Evolution Loop

1. Create — /singularity-create walks you through building a new skill: requirements gathering, duplicate checking, SKILL.md generation, registry registration, and initial scoring.

2. Score — After every skill execution, /singularity-score dispatches a Haiku assessor agent that rates the output on 5 dimensions:

3. Repair — When a skill's average drops below 50 (configurable), /singularity-repair kicks in: reads score history + telemetry, identifies the weakest rubric dimensions, rewrites the skill to fix them, bumps the version, and re-tests.

4. Harden — Each edge case encountered is recorded. Skills that handle more edge cases with higher scores progress through maturity levels.

5. Crystallize — Once a skill hits 90+ average over 5+ runs with edge cases handled, /singularity-crystallize locks it with a git tag. Crystallized skills are immutable — a production-grade snapshot you can always roll back to.

Maturity Progression

graph TD
    Draft["📝 Draft<br/><i>&lt; 3 runs</i>"] --> Tested["✅ Tested<br/><i>3+ runs, avg ≥ 60</i>"]
    Tested --> Hardened["🛡️ Hardened<br/><i>5+ runs, avg ≥ 80<br/>edge cases handled</i>"]
    Hardened --> Crystallized["🔒 Crystallized<br/><i>git-tagged, immutable</i>"]

    style Draft fill:#e9ecef,color:#000,stroke:#adb5bd
    style Tested fill:#a8dadc,color:#000,stroke:#457b9d
    style Hardened fill:#264653,color:#fff,stroke:#1d3540
    style Crystallized fill:#2d6a4f,color:#fff,stroke:#1b4332

Capability Gap Detection

The using-singularity skill (injected at every session start) teaches Claude to recognize when a new skill is needed:

• Repetition — Doing the same multi-step procedure across sessions
• Failure without coverage — No existing skill addresses the task
• Generalizable pattern — The procedure applies beyond the current task

When a gap is detected, Claude suggests creating a new skill automatically.

Data Storage

All data is local. Nothing leaves your machine.

~/.claude/singularity/
├── scores/          # Score history per skill (JSON)
├── telemetry/       # Execution logs per skill (JSON)
├── registry.json    # All managed skills
└── config.json      # Thresholds and preferences

Configuration

Edit ~/.claude/singularity/config.json:

{
  "autoRepairThreshold": 50,
  "crystallizationThreshold": 90,
  "crystallizationMinExecutions": 5,
  "scoringMode": "auto"
}

Integration

• Works alongside superpowers and other Claude Code plugins
• Skills namespaced as singularity-claude:* — no conflicts
• Created skills go to ~/.claude/skills/ — the standard Claude Code location, usable by any plugin

Requirements

• Claude Code CLI
• jq (recommended) or Node.js (fallback) for JSON manipulation
• Git (optional, for crystallization via tags)

Inspired By

Moltron — a self-evolving skill engine built on SmythOS. singularity-claude takes the same core ideas (scoring, auto-repair, crystallization, hardening) and rebuilds them natively for Claude Code with zero external dependencies.

Contributing

Contributions welcome! See CONTRIBUTING.md for guidelines.

Ideas for contributions:

• New scoring rubric dimensions
• Additional maturity metrics
• Platform support (Cursor, Codex, Gemini CLI)
• Skill templates for common patterns

License

MIT — use it however you want.

_{Built with Claude Code. Skills that improve themselves.}