Agent Skill Forge

Local-first tool that turns rough AI-agent ideas into safer, testable skill workflows.

Comment

Inspiration

Most AI-agent ideas begin as a vague prompt: "make me an agent for studying", "triage support tickets", or "coordinate field reports". Agent Skill Forge turns that messy idea into a structured skill spec that another agent or human builder can actually run and test.

What it does

Agent Skill Forge is a local-first web tool for creating SKILL.md-style agent workflows. A user writes a mission brief, chooses risk tolerance and output depth, then the app produces:

  • the best-fit agent archetype
  • a step-by-step workflow
  • safety flags for privacy, sources, and high-impact decisions
  • acceptance checks
  • a portable Markdown skill and JSON agent card

How I built it

The project is a static web app built with HTML, CSS, and JavaScript. It runs immediately with a deterministic rule planner so the demo works offline. When the user clicks "Enable local AI", the app loads Transformers.js with Xenova/all-MiniLM-L6-v2 in the browser and uses embeddings to semantically rerank agent archetypes.

Clear use of AI

AI is used in the optional local embedding layer: the browser compares the user's brief against agent archetype descriptions and chooses the closest workflow. The app also keeps a rule-based fallback so it remains usable without an API key, server, or network model download.

Challenges

The hardest part was keeping the tool practical instead of turning it into a generic prompt generator. I focused on output contracts, guardrails, and acceptance checks so the result is closer to an implementation-ready agent skill.

What I learned

I learned how to design a small AI workflow where the deterministic layer and AI layer support each other. The fallback gives reliability; embeddings add better matching when available.

What's next

Next I would add import/export for real Codex skill folders, more archetypes, and a test runner that checks generated skills against sample user scenarios.

Links

Built With

Share this project:

Updates

Submission history