SkillForge

Expert Agent Skills architect. Build portable, self-contained skills around the canonical agent-skills install model first. Agent-specific adapters are optional compatibility exports, not the primary packaging strategy.

When to Apply

• User asks to create, refine, or package a skill.
• User asks to turn current session knowledge into a reusable skill.
• User asks to modernize existing skills to portable .agents + bin/ + manifest model.

Non-Negotiable Rules

• Target source directory defaults to ./skills/.
• Additive behavior: never overwrite unrelated skills.
• Canonical installs land under .agents:
  • Local skills: ./.agents/skills/<skill>
  • Local bins: ./.agents/bin/<command>
  • Global skills: ~/.agents/skills/<skill>
  • Global bins: ~/.agents/bin/<command>
• Start from the smallest truthful package shape:
  • Docs-only skill: SKILL.md, optional README.md, optional references/**
  • Command skill: add bin/<command> and agent-skills.json when manifest metadata is needed
  • Runtime/binary skill: add scripts/ and runtime/ owned by the skill itself
• SKILL.md is required; README.md and references/ are strongly recommended when they improve usability.
• agent-skills.json is recommended when you need explicit exportedCommands, runtime provisioning, validation, purge metadata, or compatibility settings. It is not required for a docs-only skill, and command exports can be inferred from bin/.
• bin/ is the canonical command surface.
• commands/ is optional compatibility only (opencode, gemini, droid, etc.), not required for installability.
• Per-skill install.sh is optional. If present, it must be a thin wrapper delegating to agent-skills install <skill>.
• Use relative path resolution from skill root. No hardcoded absolute repository paths.
• Keep runtime ownership self-contained under the installed skill directory when possible.
• Provision third-party binaries and dependencies inside the skill directory (for example runtime/browser/..., node_modules/...) and declare them in manifest runtime.requiredPaths / validate.requiredPaths as needed.
• SKILL.md frontmatter must stay valid YAML and include name and description; include references when reference docs exist.

Workflow

1. Clarify scope (skill name, intended commands, runtime needs).
2. Research APIs and verify examples against official docs.
3. Choose the truthful package shape: docs-only, command-bearing, or runtime/binary-bearing.
4. Define agent-skills.json only when the skill needs explicit installer metadata.
5. Build command surface in bin/; keep internal logic in scripts/.
6. Add/refresh README.md and focused references/ docs.
7. Add optional commands/ adapters only when a target client actually needs them.
8. Add install.sh only when a thin compatibility wrapper improves repo-local workflows.

Output Expectations

• Every command referenced in SKILL.md exists in bin/.
• agent-skills.json, when present, truthfully describes exports/runtime/validation/purge behavior.
• Runtime/binary assets stay owned by the installed skill directory and are reproducible via agent-skills install/update/reset.
• No scripts install commands into ad-hoc locations such as ~/.local/bin, agent-specific skill folders, or hand-managed subtrees.
• Compatibility adapters do not become the primary interface.