MetaSkill

Zero-token task router for AI coding agents. Classifies any request locally (without burning LLM tokens) and decides which archetype and complexity tier should handle it, with embedded work instructions.

user request ──▶ in-context tokenization ──▶ match against index.json
                                               ├─ archetype (e.g. mobile, devops, quick_fix)
                                               ├─ complexity tier (1-5)
                                               ├─ model recommendation
                                               ├─ fallback chain
                                               ├─ suggested tools
                                               └─ work instructions for the agent

Why it exists

Choosing a model by gut feeling burns tokens: trivial tasks end up on premium models and critical tasks on weak ones. MetaSkill makes that decision a zero-cost deterministic classification based on index.json (v9.0.0, 16 archetypes).

Install (Qwen Code / Claude Code)

# Windows
Copy-Item -Recurse . "$env:USERPROFILE\.qwen\skills\metaskill"
Copy-Item -Recurse . "$env:USERPROFILE\.claude\skills\metaskill"

# macOS / Linux
cp -r . ~/.qwen/skills/metaskill
cp -r . ~/.claude/skills/metaskill

The agent will pick it up automatically on the next session via SKILL.md.

How it works

The agent reads index.json and classifies the request by keyword matching — no Python subprocess, no shell calls, zero additional tokens. The classification happens in the agent's own reasoning.

4 tiers

Tier	Models	Use for
local_zero_token	ollama (if available)	typos, trivial patches
budget_fast	qwen-flash, deepseek-flash, glm-flash	docs, scripts, simple tasks
standard_coding	deepseek-pro, qwen-plus	product development
premium_reasoning	qwen-max, deepseek-pro-max	architecture, migrations, research

16 archetypes

Security audit, Frontend/UI, Backend/API, DevOps/IaC, Data science, System architecture, Documentation, Quick fix, Testing/QA, Database, Mobile, Migration/Refactor, CLI/Automation, Game dev, Research.

CLI (optional, for external integrations)

python metaskill.py "build a mobile app with flutter for android"
python metaskill.py "fix a bug in the log" --json

Returns archetype, tier, fallbacks, tools, and instructions. Use --json for programmatic consumption.

Adding or editing archetypes

Edit index.json → tasks[]. Each archetype needs: id, archetype, label, keywords (include common typos/variants), complexity (1-5), routing (tier, fallback, tools) and instructions. The router picks it up automatically.

Structure

metaskill/
├── SKILL.md      # skill definition (Qwen Code + Claude Code compatible)
├── metaskill.py  # optional CLI router (Python stdlib only)
├── index.json    # archetype index, tiers, and fallback policy
└── README.md

Model aliases are the only thing that changes

When a model dies, update ONLY model_aliases in index.json. The routing logic never changes.

License

MIT — use, modify, and distribute with attribution.