Skill Forge

What is a skill?

A folder containing a SKILL.md file with YAML frontmatter — a structured set of instructions an AI agent reads on demand. Think of it as a compressed expert brain you can hand to any agent. The agent only loads the skill when its description matches what the user is asking for.

Required frontmatter (per agentskills.io):

Optional frontmatter:

Directory structure (everything beyond SKILL.md is optional):

skill-name/
├── SKILL.md          required — entry point loaded at activation
├── scripts/          optional — executables the skill can run
│   └── *.py *.sh *.ts
├── references/       optional — on-demand docs the body links to
│   └── *.md *.json
├── assets/           optional — templates, sample data, images
│   └── *.docx *.pptx *.json *.png ...
└── evals/            optional — A/B test cases (see agentskills.io/evaluating-skills)
    └── evals.json

Example minimal SKILL.md frontmatter:

---
name: csv-cleanup
description: Clean a CSV file deterministically. Use when the user says "clean this CSV", "fix the data", or attaches a .csv that needs validation. Detects encoding, normalizes headers to snake_case, type-coerces, dedupes, emits a profile report.
license: MIT
metadata:
  version: "1.2.0"
  author: community
  tags: [data, etl, csv]
---

How to read the rubric

Each skill is scored on 8 evidence-based axes (0–100), weighted by importance. ≥90 = production-ready. Sourced from the agentskills.io spec + Anthropic's Building Skills guide + patterns observed in high-performing skills.

How evals work — and why this is the right way to measure a skill

The eval spec format comes from agentskills.io/skill-creation/evaluating-skills — Anthropic's published methodology, also used by their skill-creator meta-skill. Each case is a realistic user prompt plus assertions the output must satisfy (substring, regex, length, or LLM-judged).

The methodology is paired A/B runs:

• with_skill — the same agent runs the case with your skill loaded.
• without_skill — the agent runs the identical case without the skill.

Why this is the right way: it controls for the prompt (same input both sides), controls for the model (we use Qwen 3.6 Plus on both runs), and produces a measurable delta. Vague claims of "my skill makes the agent better" become quantified lift — or get falsified. We measure three things:

A good skill improves at least one without degrading the others. The benchmark scorecard in the Evals tab surfaces all three deltas after each run.