SkillForge

MANUAL — DOCUMENTATION

From zero to skilled agent in ten minutes

What a skill is, how to install one on four platforms, how to author and publish your own, and how to drive everything as an agent over MCP or REST.

Step 1 — Understand the unit: SKILL.md

A skill is a markdown file with three sections: triggers (when to fire), workflow (numbered steps the agent follows), and guardrails (lines it must not cross). No SDK, no build step, no vendor lock-in — any agent that reads markdown can run it. Every SkillForge record embeds this file and scores it for completeness.

# My Skill

## Triggers
When the user asks for <thing> or the agent sees <signal>.

## Workflow
1. Clarify the exact outcome first.
2. Do the smallest safe step.
3. Show the result with evidence.

## Guardrails
- Never invent sources or IDs.
- Ask before destructive actions.

Step 2 — Install one anywhere in 60 seconds

Open any skill page and pick your platform tab. The pattern is always the same: create a folder, curl the SKILL.md into it, point your agent at the folder. OpenClaw uses skills/, Claude Code uses .claude/skills/, Cursor uses .cursor/skills/, and any MCP agent can skip files entirely and call the registry over HTTP.

Step 3 — Author your own in 5 moves

  1. 01 Name one job the skill does. One job, one skill.
  2. 02 Write triggers as observable signals, not vibes.
  3. 03 Write the workflow as numbered steps with a visible evidence rule.
  4. 04 Write guardrails first-class: secrets, destruction, autonomy limits.
  5. 05 Declare risk flags honestly and POST it to /api/score to see your grade.

Step 4 — Publish it to the registry

Fork the repo, add your skill to src/lib/fallback.ts following the existing shape, run npm run lint and npm run build, then open a PR. Curated skills ship with the offline catalog so they work even with zero network. See CONTRIBUTING.md for the checklist.

Step 5 — Drive it as an agent (MCP)

Add this block to mcp.json in OpenClaw, Claude Code, Cursor or any MCP-capable agent. Five tools appear: search_skills, get_skill, score_skill, compose_bundle, verify_seal.

{
  "mcpServers": {
    "skillforge": {
      "url": "https://skillforge-aniruddha-adaks-projects.vercel.app/api/mcp",
      "transport": "http-json-rpc"
    }
  }
}
Live MCP consolePOST /api/mcp

Request

Press a preset to call the live MCP endpoint.

Response

Response appears here.

Step 6 — Use the REST API directly

No keys, cached hourly, offline-sealed, paginated across 10,000+ records. Health, catalog search with source, license, stars and exclusion filters plus category facets, single skill with seal, raw SKILL.md for install scripts, deterministic scoring, and full-chain verification (10k hashes in ~0.5s).

curl -s "https://skillforge-aniruddha-adaks-projects.vercel.app/api/skills?q=pdf&sort=score&limit=3"
curl -s https://skillforge-aniruddha-adaks-projects.vercel.app/api/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call",
       "params":{"name":"search_skills","arguments":{"query":"pdf","limit":3}}}'
curl -s https://skillforge-aniruddha-adaks-projects.vercel.app/api/seals/verify \
  -H "Content-Type: application/json" \
  -d '{"ids":["pr-reviewer","pdf-miner"]}'

Step 7 — Understand what you are building toward

Prompts become skills, skills become agents, agents become economies of agents — and the same ladder climbs toward AGI and superintelligence. The safety briefing explains why the registry layer is where alignment gets practical.