skill-forge — The meta-skill that forges skills

The problem

Three gaps in the skill lifecycle

01 Don't know when to create a skill Auto-detect complex tasks, proactively ask

02 Don't know if a skill is well-written 5-dimension evaluator, threshold before save

03 Don't know if it will actually trigger Eval-driven description optimization

The forge

Ore /scan Discover opportunities in your codebase

Heat /create Research patterns, draft the skill

Strike eval 5-dim scoring, reject below 6/8

Temper /improve Patch content, optimize triggers

Blade SKILL.md Production skill, registry tracked

Commands

claude-code

$ /scan [prompt]

Scan project for skill opportunities.

$ /create <prompt>

Create skill from prompt. Name auto-derived.

$ /improve <prompt>

Iterate skill from prompt. Target matched from registry.

$ /list

Print skill registry — versions, scores, trigger rates.

Under the hood

Dual-file isolation

Injection-proof by design

External content lands in low-trust skill_insights.md. Only validated patterns promote to the high-trust draft that hooks inject.

Attention anchoring

Hooks keep the goal in focus

UserPromptSubmit and PreToolUse hooks re-inject the draft header every turn. Context overflows can't erase the mission.

Session catchup

Nothing slips through the cracks

Missed a complex workflow yesterday? SessionStart parses previous JSONL, surfaces uncaptured 5+ tool-call tasks.

Eval-driven triggers

Description is a separate problem

20-case trigger evaluations, 60/40 train/test split, up to 5 optimization rounds. Your skill fires when it should.

Then run /scan to discover your first opportunities.