Skip to content

schizo16/opencode-skillforge

Repository files navigation

SkillForge

Create better OpenCode skills from vague user requests.

opencode-skillforge is an OpenCode skill that creates, adapts, and audits agent skills (.opencode/skills/SKILL.md). It turns vague prompts into structured specs, generates audited skill files, and supports community discovery of existing skills and MCP servers.


Most "write me a skill" prompts produce vague, assumptive skills. SkillForge makes the agent interview, spec, generate, and audit the skill before writing a single file — turning a fuzzy request into a reviewed, read-only skill you can trust.


Before vs After

Before (simple prompt) After (SkillForge)
"Make a skill that reviews my frontend code." Same prompt — but the agent first analyzes intent, checks existing skills, asks clarifying questions, writes a spec for approval, generates files, then audits the result.
AI immediately writes a vague SKILL.md with assumptions about your framework, tools, and workflow. The agent waits for your spec approval before touching any file.
No spec. No audit. No safety boundaries. 9-criterion quality audit. Explicit safety boundaries. Read-only by default.

How It Works

Vague request
  ↓
Intent analysis
  ↓
Composition analysis
  ↓
Community discovery
  ↓
Questions / defaults
  ↓
Skill spec
  ↓
Quality audit
  ↓
User approval
  ↓
Generate skill files

SkillForge is an OpenCode skill. Install it, ask for a skill, and the agent walks through all 9 steps before writing anything.


Quick Install

# macOS / Linux
curl -sSL https://raw.githubusercontent.com/schizo16/opencode-skillforge/main/install.sh | bash
# Windows PowerShell
iex (Invoke-RestMethod -Uri 'https://raw.githubusercontent.com/schizo16/opencode-skillforge/main/install.ps1')

Manual install instructions →


Quick Verification

Run this in OpenCode:

"Use the SkillForge skill. Make a skill that reviews frontend code. Do not create files yet."

Expected: Skill Intent Analysis, Existing Skill Check, and Configuration Questions — not a generated file.

Full verification guide →


Compatibility

OpenCode (primary target)

Copy the skill folder into your project:

cp -r .opencode/skills/skillforge <your-project>/.opencode/skills/skillforge
Copy-Item -Recurse -Path .opencode/skills/skillforge -Destination <your-project>/.opencode/skills/skillforge

The skill activates automatically when OpenCode detects SKILL.md under .opencode/skills/.

Claude Code

Claude Code supports Agent Skills using the same SKILL.md format. Copy the skill folder into your project's skill directory (typically .claude/skills/ or project root depending on your setup):

cp -r .opencode/skills/skillforge .claude/skills/skillforge

Then activate by asking: "Use the SkillForge skill. Make a skill that reviews frontend code."

Codex

Codex uses AGENTS.md or CODEX.md for project instructions. Port the workflow manually:

  1. Create or edit AGENTS.md in your project root.
  2. Copy the Core Rule, Workflow steps, Do-Not Rules, and Safety / Boundaries sections from SKILL.md into AGENTS.md.
  3. Save and trigger with: "Suggest a skill for this project."

Cursor

Cursor supports Cursor Rules (.cursor/rules/) and AGENTS.md. To adapt:

  1. Create .cursor/rules/skillforge.mdc or add to AGENTS.md.
  2. Include the workflow steps, do-not rules, and decision rubrics from SKILL.md.
  3. Trigger with: "Create a skill for reviewing my frontend code."

Other agents

The 7-step workflow can be used as reusable project instructions for any coding agent that supports custom rules or instructions. See docs/compatibility.md for detailed porting guidance.

Note: SkillForge is OpenCode-first. Support for other agents has not been fully tested. Installation steps above are based on each tool's documented conventions.


File Structure

skillforge/
├── README.md                         # This file
├── INSTALL.md                        # Install and verification guide
├── demo.md                           # Browse generated skill examples
├── install.sh                        # macOS/Linux install script
├── install.ps1                       # Windows install script
├── LICENSE
└── .opencode/
    └── skills/
        └── skillforge/
            ├── SKILL.md              # SkillForge's own workflow definition
            ├── README.md             # Skill-level documentation
            ├── templates/
            │   ├── skill-spec.md     # Spec template for new skills
            │   ├── skill-audit.md    # Audit scoring template
            │   └── skill-readme.md   # README template for new skills
            └── examples/
                ├── create-frontend-review-skill.input.md
                ├── create-frontend-review-skill.expected.md
                ├── create-responsive-a11y-review-skill.expected.md
                ├── create-ai-claim-audit-skill.input.md
                └── create-ai-claim-audit-skill.expected.md

See demo.md and .opencode/skills/skillforge/examples/ for complete workflow walkthroughs.


What Makes It Different

Aspect Simple "create-skill" prompt SkillForge
Intent analysis Assumes it knows what you want Explicitly restates and confirms
Existing skills Ignores them Checks before creating
Questions Anything that comes to mind Only blocking questions (max 3)
Spec None — generates immediately Written contract, user-approved
Quality check None 9-criterion audit, fails until clean
Safety Implicit Explicit boundaries documented

Verification

No-code manual checks for v0.1.1:

  1. Skill file exists — confirm .opencode/skills/skillforge/SKILL.md is present in your project.
  2. Agent does not jump to generation — ask: "Make a skill that reviews frontend code." The first output should be Skill Intent Analysis, Existing Skill Check, and Blocking/Configuration Questions — not a generated SKILL.md file.
  3. Browse examples → — complete workflow walkthroughs.
  4. Browse generated skills → — skills produced by SkillForge: responsive-a11y-review, ai-claim-audit.

Roadmap

  1. OpenCode skill — SkillForge as a self-contained skill (current)
  2. MCP server — expose SkillForge workflows via the Model Context Protocol for editor integration
  3. OpenCode plugin — package as a plugin with auto-discovery and versioning
  4. CLI / GitHub Action — standalone CLI for CI pipelines and non-OpenCode environments

Limitations

  • The existing-skill check only searches locally available skills — there is no global registry search
  • SkillForge creates skill files but does not install or activate them; the user must place them in the correct directory
  • Templates are Markdown-only; no programmatic validation of generated skills
  • No version management or dependency resolution between skills

License

MIT — see LICENSE for details.

About

OpenCode skill that creates, adapts, and audits agent skills from vague user requests. Generates SKILL.md files with spec, audit, and safety boundaries.

Topics

Resources

License

Stars

Watchers

Forks

Packages

 
 
 

Contributors