Skill Tracker

Status tiers (read thresholds from /home/hubt/.claude/skills/skill-tracker/config.json → thresholds; defaults: candidate=5, priority=10):

• 0 to (candidate-1) uses → Learning
• candidate to (priority-1) uses → Candidate (eligible for codification)
• priority+ uses → Priority (should be codified)

Also sort by estimated_cost_usd descending so the most expensive skills surface first.

Codifying a skill

When the user picks a skill to codify:

1. Read its SKILL.md and understand exactly what it does end-to-end.
2. Identify the parts that are deterministic (file reads/writes, git commands, API calls with known shapes) vs. the parts that require LLM reasoning (drafting prose, inferring intent, open-ended decisions).
3. Write a Go program (preferred) or bash script in the skill's directory that handles the deterministic parts. If LLM reasoning is unavoidable, have the script emit a structured prompt and pipe it to claude -p as a subprocess.
4. Compile any Go code: go build -o <name> ./<name>.go
5. Update the skill's SKILL.md instructions to simply invoke the compiled binary or script, passing $args through.
6. Tell the user what was automated vs. what still uses LLM reasoning, and why.