Let a Meta-Skill Watch Your Sessions and Improve Your Skills

Skills are frozen the day you install them: nothing about your skills learns from how you actually use them. task-observer, the skill in rebelytics/one-skill-to-rule-them-all, watches your work sessions and turns recurring friction into concrete skill updates — with a human review between the observation and the change.

Why This Skill Matters

A meta-skill runs alongside your work and looks for three things, per the README: corrections you make to the AI's output, gaps no existing skill covers, and the observer's own blind spots. Each session produces a structured observation log — what it noticed, which skills are affected, suggested improvements. The author reports over 1,800 observations across 81 skills, most converted into improvements. Patterns that cut across skills get captured as cross-cutting principles in a separate log, and new skills are checked against them whenever they are created or updated. Because the observer watches every active skill, including itself, its own methodology improves with use.

Installation

npx skills add rebelytics/one-skill-to-rule-them-all --skill task-observer

Add -g to install it for all your projects. In the Claude apps (web, desktop, mobile, Cowork), upload the .skill bundle from the latest release via Settings → Customize instead. Installing alone is not enough: the README warns that description matching under-triggers, so add the activation instruction from references/environments.md to your CLAUDE.md or your platform's equivalent.

Then verify it actually runs — in a new session, since the one you install in cannot prove it. The README names the external tell: if skill-observations/observation-log/ doesn't exist after a few sessions of real work, activation never happened.

Real Workflow: Turn Corrections Into Skill Updates

1. Work normally. During each session the observer writes observation entries: what it noticed, which skills are affected, what to change.
2. At the end of a session, ask "Any observations logged?" — the README's own quick-start tip; the skill then sometimes surfaces additional improvement potential it hadn't logged yet.
3. Proposed updates land in skill-updates/ as staged copies. You review and approve each change in an interactive review, then install the updated bundle. Nothing goes live before that.

For recurring upkeep, the README describes a scheduled review that applies only the observations needing no decision from you — and even those go to staged copies first.

Tips

• The README positions this against project memory: project memory records a pitfall for one project and applies itself automatically; task-observer turns recurring friction into a reusable-skill change that carries into every project loading that skill, with a human review in between.
• A pitfall you store in project memory more than once, or in more than one project, is a task-observer signal — it has stopped being about the project and started being about the skill.
• The skills.sh listing reports over 10,000 installs — a lower bound, the README notes, because the CLI's install telemetry is opt-out — and carries independent security audits.
• Setting up? The README suggests handing the skill, README, and user guide to your agent and letting it configure your specific environment.

When Not to Use This

The README's own honest boundary: the formal log-and-review cycle pays off most as your library and usage grow. With a handful of skills, your AI system's built-in memory may cover much of the same ground with less overhead, and editing a skill directly is quick. Also mind the platform fit: the skill assumes Claude-centric concepts, so other platforms need adaptation — a Codex version exists (maintained by AllstarGER, based on an older version), and Claude web or mobile runs in a handoff-doc mode without filesystem access.

See the leaderboard for more skills.