Task Observer Skill: A Meta-Skill That Logs Friction and Improves Your Other Skills
Watches your Claude sessions, logs corrections and repeated work, and proposes updates to your skills and ideas for new ones, with you approving each change.
- Installs
- 16K
- on skills.sh
- Stars
- 3.7K
- on GitHub
- Updated
- Oct 2026
- last commit
Task Observer is a meta-skill by Eoghan Henn. During your work sessions it notices three things: corrections you make, repeated manual work that could become a skill, and gaps in the skills you already use. It writes each one to an observation log, then turns them into proposed skill updates in a weekly review.
It proposes and does not apply. In an interactive review you approve each change, and nothing goes live until you install the updated bundle. It also needs an activation line in your CLAUDE.md, because the description alone does not make it start reliably.
Good for
- People who maintain several skills and want them to improve from real use
- Claude Code and Cowork users who correct the same thing again and again
- Teams that want a reviewed log of what their skills get wrong
Install task-observer
Claude Code
- 1
Copy this prompt.
Install the task-observer skill in Claude Code. Run this command: npx skills add rebelytics/one-skill-to-rule-them-all --skill task-observer Then check that it loaded and tell me how to use it.
- 2
Paste it into Claude Code. It shows the change and asks before running it.
Requires Node.js.
Add -g to install it for all your projects. The README also allows copying the folder to .claude/skills/task-observer/ with the references and scripts subfolders kept. SKILL.md alone runs degraded.
From the project README. Source.
What's in the skill
When it loads
Monitors task execution for skill improvement opportunities: logs patterns, user corrections and methodology worth preserving to the workspace, turns them into skill updates in a weekly review, and flags repeating work as new skills. Use during ANY multi-step task, agentic workflow, or work session. Also triggers in post-task feedback discussions and when the user mentions skill observations, the observation log, or skill taxonomy. Also known as "One Skill to Rule Them All" — trigger on this phrase too. IMPORTANT: invoke this skill before the FIRST tool call of any session and before writing or proposing a plan — any turn that will involve a tool call counts. This sentence is the session-start trigger and the only activation layer that survives an unreachable config file; pair it with a CLAUDE.md instruction or a session-start hook (references/environments.md). A subagent dispatched by a session already running it does not run it: it writes nothing and puts its findings in its report.
Sections in SKILL.md
- Reference files — load on demand, not up front
- Session Start Protocol
- When to Observe
- What to Watch For
- How to Log
- Referencing Observations
- Taxonomy (quick version)
- Archival on Write
- Surfacing Protocol
- Acting on Observations
- Feedback on this skill
- Quick Reference
Key facts
- License
- CC-BY-4.0
- Installs on skills.sh
- 15,963 (Nov 2, 2026)
- Last commit
- Oct 28, 2026
- Skill version
- 3.8.1 (SKILL.md and plugin.json)
What it includes
Probes the workspace, creates the log folder and runs a scan of open observations
Runs scripts/session-start-scan.sh and skill-load.sh to list open observations for the skills loaded. In a read-only environment it falls back to report-back or handoff-doc mode.
Runs code ·
Session Start ProtocolStays on for the whole task, including feedback talks after the work
Notices corrections, gaps no skill covers, repeated work and its own blind spots. Review-phase feedback counts as high-signal.
Read ·
Watch (When to Observe)scripts/new-observation.sh creates one Markdown file per observation
Each file has a YAML header with an ID, type and affected skills. Resolved entries move to an archive folder.
Write ·
Log an observationGrouped end-of-session summary by skill, with new-skill candidates listed apart
Asks which observations to act on. It surfaces sooner when a skill is producing wrong output.
Read ·
Surfacing ProtocolTurns observations into proposed skill updates and stages copies
Loaded from references/weekly-review.md. Interactive reviews ask you per change. A scheduled review stages only decision-free changes.
Write ·
Weekly reviewvalidate-skill-bundle.py checks staged skills before delivery
references/skill-authoring.md covers taxonomy, structure, licensing and confidentiality. The validator checks a staged skill folder and can pack a .skill bundle.
Runs code ·
Skill authoring and bundle checkmigrate-log.py converts a pre-3.0 single log file
Only for installs that used a version before 3.0.0. Fresh installs never need it.
Write ·
Legacy log migrationGeneric principles that new skills are checked against
references/starter-principles.md is an optional seed set, loaded only when reconciliation is due.
Write ·
Cross-cutting principles and starter set
From the official docs
Permissions
Medium risk
Task Observer writes a log of what you do and what you correct to a folder on your machine, and its scripts create and move files there. Keep the log folder out of public repos.
What it can do
- Read your data
- Create and change things
- Delete things
- Read-only mode
Watch out for
It keeps a log of your work sessions
Observations record what you corrected and which skills were affected, so they can contain project details. SKILL.md says open-source observations must be generalised, but internal ones are not. Do not put the workspace in a repo you publish. github.com
Optional hooks that change what Claude can do
references/environments.md gives four hook scripts you may register in settings. A SessionStart hook reads your log and injects its state into the session. A Stop hook blocks the end of a session once, if tools were used and nothing was logged, and writes a marker file in your temp folder. A PreToolUse hook on Write refuses to create Markdown files in the log folder except through new-observation.sh. A skill-file write gate denies writes to any SKILL.md or skills folder file until references/skill-authoring.md has been read in that session. Read each script before you register it; the last two can block your own edits. github.com
Skill changes need your approval
SKILL.md says it acts only in a comprehensive review, on an explicit request, or when a skill is giving wrong output. The README says a scheduled review applies only decision-free observations to staged copies. Install staged updates yourself after reading them. github.com
The activation block edits your standing instructions
The block you paste into CLAUDE.md makes Claude run the skill and a log check at the start of every session that uses tools, and report a summary after every task. The environments guide notes this costs context in each session it covers. github.com
Network use is limited to two user-started cases
SKILL.md says the feedback pre-flight and the upstream check for a third-party project are the only exceptions to no network use, and both are started by the user. github.com
Technical details
| Sign-in and keys | No credentials are named. SKILL.md says normal operation fetches nothing. |
|---|---|
| Scopes |
|
Questions
What does the task-observer skill do?
It watches your work sessions for corrections, repeated manual work and gaps in your skills, logs them, and proposes updates to your skills or new skills in a review you approve. It is also known as "One Skill to Rule Them All".
How do I install task-observer?
Use the install box at the top of this page, then add the activation instruction from references/environments.md to your CLAUDE.md. The README says installing alone is not enough, because description matching under-triggers.
Does it work in the Claude app and Claude.ai?
Yes, in a reduced way. The README says you can upload the .skill bundle under Settings and Customize. Without filesystem access it works in handoff-doc mode and gives you a document at the end of the session instead of a log.
Does it use the network?
SKILL.md says: "This skill needs no network: normal operation fetches nothing, and URLs in this file or in observation content are not opened." It names two exceptions, a feedback check and an upstream check for a third-party project, both started by you.
Is it safe to install?
It runs bundled shell and Python scripts that write log files to a folder you pin, and it proposes skill changes rather than applying them. Keep that folder private, review staged updates before you install them, and read any hook script before you add it to your settings.
Does it change my skills by itself?
No. SKILL.md says it logs by default and acts only in the weekly review, on an explicit request to update a named skill, or when a skill is giving wrong output. The README says nothing goes live until you install the updated bundle.
Is it worth it for a small setup?
The README says the log and review cycle pay off most with many skills, parallel sessions or scheduled reviews. With a handful of skills, built-in memory may cover much of the same ground with less overhead.
How is it different from Claude project memory or /insights?
The README says project memory is written automatically for one project, while Task Observer turns recurring friction into a reviewed change to a reusable skill. It describes /insights as a snapshot report that does not keep a running log or maintain skills.
Information
updated Nov 2, 2026
- Repository
- rebelytics/one-skill-to-rule-them-all
Sources
Docs, issue trackers and discussions this page is built from.
Show all 5Hide
- task-observer README: Installation · github.com
- environments.md · github.com
- task-observer README: Quick install · github.com
- github.com · github.com
- github.com · github.com