What is the task-observer skill for Claude Code?

Most of what you teach an agent evaporates when the session ends. You correct it four times, it gets the task right, and the next session starts from zero. task-observer is a small skill aimed squarely at that leak: it runs alongside the work and records what should have been a skill. It sits in the AI and agent building category.

Key takeaways

• The task-observer skill has 6,070 installs on rebelytics/one-skill-to-rule-them-all, a repo with 2,392 GitHub stars last pushed 4 September 2026 (skills.sh registry, GitHub).
• It is designed to be invoked at the start of a session, not at the end, so corrections are captured as they happen rather than reconstructed afterwards.
• Its stated outputs are workflow patterns, user corrections, skill-improvement opportunities, and taxonomy refinements, feeding an observation log.
• The name in the repo is "One Skill to Rule Them All", which is why the head query and the repo name do not match; the installable skill is task-observer.
• Do not confuse it with observer-pattern by PatternsDev (534 installs, 244 stars), which teaches the publish and subscribe design pattern for JavaScript and TypeScript (skills.sh registry, GitHub).

What does task-observer actually do?

It monitors task execution during multi-step agentic workflows to capture patterns, user corrections, and methodology worth preserving as reusable skills (skills.sh registry). In practice that means the session produces two artifacts: the deliverable you asked for, and a record of how you had to steer to get it. The second one is the input to your next skill.

What is the best skill for capturing agent workflow patterns?

task-observer is the leading option in this narrow niche at 6,070 installs (skills.sh registry). Install it with:

npx skills add https://github.com/rebelytics/one-skill-to-rule-them-all --skill task-observer

The honest comparison is that this niche is thin. The nearest catalog match by name, observer-pattern, has 534 installs and is about software architecture rather than agent observation. If what you want is durable project context instead of session capture, the memory skills are a different and larger family, covered in our persistent memory guide. The ranked general list is on the AI agents pillar.

Why invoke it at session start?

Because corrections are only legible in context. The skill triggers at session start for deliverable work by design (skills.sh registry): once the agent has already finished, the reasoning behind your third correction is gone and only the final state survives. Starting the observer first is what turns a messy session into an input for skill authoring.

How do you turn an observation log into a real skill?

By writing the captured pattern into a SKILL.md with a description precise enough to trigger on the right tasks. Anthropic documents the file format and the frontmatter contract in the Claude Code skills documentation, and our own walkthrough of the authoring loop is in the guide to building custom Claude Code skills. The observation log is the raw material; the skill file is the artifact.

This pairs naturally with review skills. If the corrections you keep repeating are about code quality, the patterns are worth encoding once, and the ranked options are in the code review skills guide.

A complete starter stack

1. task-observer (6,070 installs) invoked at the start of any multi-step deliverable session.
2. A skill-authoring workflow to convert the log into SKILL.md files; see the custom skills guide.
3. A memory or context skill so the encoded knowledge is actually loaded next time, from the persistent memory guide.

Common pitfalls

• Installing it and never reading the log. Capture without a weekly pass over the observations produces a file nobody uses.
• Turning every correction into a skill. A one-off fix is not a pattern; wait until the same correction appears in three separate sessions before encoding it.
• Searching the catalog for "observer" and installing observer-pattern instead. That skill has 534 installs and teaches publish and subscribe architecture, which is a different subject entirely (skills.sh registry).

Run it across one week of real sessions, then read the log before you write a single new skill; the repeated corrections will pick the topic for you.

Common questions