task-observer: the meta-skill that improves your skills, itself included
Skills tend to stay frozen even when you correct the agent in the same spot every day. task-observer runs alongside your work and logs corrections, gaps no skill covers yet and its own blind spots. A weekly review turns those observations into staged skill updates, and nothing goes live until you install them.
Saves each observation as its own Markdown file with YAML frontmatter under skill-observations/observation-log/ in a stable workspace folder.
Spots manual work you repeat and flags it as a candidate for a new skill, then helps you build it.
Collects lessons that apply across skills as cross-cutting principles and checks every new or updated skill against them.
Offers a review in one line when the last one is 7 or more days old and observations are open, without blocking your task.
Stages updated skill copies in skill-updates/ and leaves your installed skills untouched until you install the new bundle.
When to use it
You maintain dozens of custom skills and can no longer audit each one by hand.
You keep correcting the agent the same way and want the fix to land in a reusable skill rather than one project's memory.
You have no skills yet and want help finding which recurring tasks are worth turning into one.
Ask your agent
Any observations logged this session?
Run the weekly observation review, starting with the 10 oldest
Check whether task-observer is activated in my CLAUDE.md and runs at session start
Questions
How do I install task-observer?
Run npx skills add rebelytics/one-skill-to-rule-them-all --skill task-observer, or upload the release's .skill bundle in the Claude apps via Settings → Customize. Then add the activation instruction to CLAUDE.md.
How is task-observer different from project memory or /insights?
Project memory stores facts for one project automatically; task-observer turns recurring friction into reviewed skill changes. /insights gives a one-off report and keeps no running log.
Does task-observer work in Codex?
The author tests it in Claude Cowork, Claude Code and Claude.ai. Community members published Codex and OpenCode adaptations based on an older version.
