A Self-Improving Supervisor for AI Skills

What it does

Task Observer is a SKILL.md bundle that sits alongside your AI sessions and watches for patterns. When you correct the AI, adjust its output, or do something manually that ought to be systematised, it logs structured observations. These feed back into your existing skills or spawn new ones, so your playbook evolves instead of rotting.

The interesting bit

The observer watches everything, including itself. If its own logging format is clunky or it triggers where it shouldn’t, it captures that too — a kind of ouroboros for prompt engineering. The author claims it logged over 900 improvements across 50 skills in six months, though your mileage will depend heavily on how many skills you actually maintain.

Key highlights

• Produces structured observation logs and cross-cutting principles that apply across your whole skill library
• Runs in Claude Cowork, Claude Code, and web/mobile via a degraded “handoff doc” mode
• Community ports exist for Codex, and users report success with Hermes and Openclaw
• Keeps the always-loaded footprint small by loading reference files on demand
• Licensed CC BY 4.0, so you can remix it freely with attribution

Caveats

• Built for Claude’s skill architecture; other platforms need to interpret Claude-centric concepts like <available_skills>
• The README admits this is overkill for small setups — if you only have a handful of skills, direct editing is faster than maintaining a review cycle
• Requires active review: it suggests improvements but does not apply them automatically

Verdict

Worth adopting if you maintain a growing library of AI skills and want systematic maintenance without manual audits. Skip it if you just want a few static skills that never change.