Skill Forge
A skill that builds skills. Skill Forge is an agent-agnostic skill for authoring other agent skills that are discoverable, reliably followed, and actually change agent behavior — not just well-formatted. It runs in any agent harness — Claude Code, Cursor, Codex, Gemini CLI, Copilot, or a custom loop — and the skills it builds are harness-neutral too.
Most hand-written skills fail in one of two ways: an agent never loads them (bad description), or it loads them and ignores them (weak instructions). Skill Forge encodes the practices that fix both, and gates every skill it produces against them.
What makes it different
- One home for every rule. Word limits, required sections, the quality gate, and bulletproofing requirements live in a single rule registry (
references/rules.md). Nothing drifts.
- Behavioral-force levers, enforced. Five levers that decide whether an agent obeys a skill — imperative force, positive specification, load-bearing examples, concrete anchors, and position — are rules it checks, not advice it mentions.
- Clarify intent before building. It interviews you (with a multiple-choice UI) until the scope, triggers, and success criteria are confirmed — so it builds what you meant, not what it guessed.
- A real validation gate. Every generated skill is run against a seven-point quality gate before it ships.
- Bulletproofing for discipline skills. Rules that must hold under pressure get a rationalization table, red flags, and an escalation path — with anti-bloat and "deliver, don't lecture" guidance baked in.
Install
Option A — skills CLI / skills.sh (recommended, works across agents)
From your project root:
npx skills add adriancodes/skillforge
Installs directly from the GitHub repo — no prior registration — and your agent (Claude Code, Cursor, Codex, Gemini, Copilot, …) picks up SKILL.md on the next session. Start a new session and ask to "create a skill", or invoke /skill-forge where supported.
Option B — Claude Code plugin
In Claude Code:
/plugin marketplace add adriancodes/skillforge
/plugin install skill-forge@skillforge
Then start a new session. Skill Forge will trigger on requests like "create a skill" or you can invoke it with /skill-forge.
Option C — manual (personal skill)
git clone https://github.com/adriancodes/skillforge.git
cp -r skillforge/skills/skill-forge ~/.claude/skills/skill-forge
Start a new session and it's available.
Option D — load directly in any other harness
Skillforge is harness-neutral. Point your agent at AGENTS.md (the cross-tool entry point), or load skills/skill-forge/SKILL.md directly — into a Cursor rule, GEMINI.md, the Copilot skill tool, or your system prompt. Tool names map to your harness via skills/skill-forge/references/harness-tools.md, which also gives a fallback for any tool your harness lacks. In a harness with no skill system, it's a markdown instruction doc you include deliberately — every step still works; only auto-triggering is manual.
Usage
Trigger it naturally or explicitly:
/skill-forge
Build me a skill that enforces conventional commit messages.
Skill Forge will:
- Clarify intent — ask targeted (multiple-choice) questions about scope, triggers, and what "done" looks like, then restate it back for confirmation. (Say "just make reasonable choices" to skip the questions; it will state its assumptions instead.)
- Scope and classify — pick the skill type (technique, discipline, reference, or workflow) and its build recipe.
- Draft against the canonical template, applying the five behavioral-force levers.
- Bulletproof if it's a discipline skill.
- Validate against the seven-point quality gate.
What's inside
The skill lives at skills/skill-forge/:
| File | Purpose |
|---|
SKILL.md | The workflow: clarify → scope → draft → extract → bulletproof → validate |
references/rules.md | The rule registry — single source of truth for every checked invariant |
references/behavioral-force.md | The five levers that make an agent obey a skill, with before/after rewrites |
references/body-template.md | The canonical SKILL.md template + per-type adaptations |
references/description-guide.md | How to write a description that triggers (and the anti-patterns that don't) |
references/bulletproofing-guide.md | Hardening discipline skills against rationalization |
references/validation-checklist.md | The pre-flight gate, run before any skill ships |
references/harness-tools.md | Maps Claude Code tool names to other harnesses, with a fallback for each |
CONTEXT.md documents the project's own vocabulary (rule registry, behavioral force, drift, harness-neutral).
xu-xiang
affaan-m
Jeffallan