Anthropic Agent Skills
anthropics/skillsAnthropic's reference skills repo: 20 skills including document tools (docx, pptx, xlsx, pdf), frontend-design, skill-creator and mcp-builder.
Recommended
Check first Depth 4/5
task-observer (One Skill to Rule Them All) Claude skill
A meta-skill that watches your sessions, logs corrections and repeated work, and proposes improvements to your other skills for you to approve.
Our verdict Recommended
A thoughtful, actively maintained feedback loop for skill authors; the cost is a long skill file and a manual activation step.
What task-observer (One Skill to Rule Them All) does
task-observer runs alongside normal work. During any multi-step task it records observations to a workspace folder: corrections you made, preferences you stated, judgment calls, and work that keeps repeating. A weekly review then turns those logs into specific edits to your existing skills and into candidates for new skills. Nothing is installed automatically; you choose which changes to accept. It also applies the same process to itself.
The SKILL.md is about 45 KB, with 14 reference files covering the observation log, signals, weekly review, skill authoring, environments and migration, plus five helper scripts (a session-start scan, observation and skill-load shell scripts, a log migration script and a bundle validator). Because description matching tends to under-trigger, the README says installing is not enough: you must add an activation instruction to CLAUDE.md or use a session-start hook, and then check that it runs. It can be installed with npx skills add or uploaded as a .skill bundle in the Claude apps. The skill states that normal operation fetches nothing from the network. Its benefit grows only if you already maintain a library of your own skills.
Good for
- People who maintain their own skill library and want it to improve from real use
- Spotting repeated work that deserves a skill
- Teams that review skill changes on a weekly cadence
Not for
- Users with only a few off-the-shelf skills
- Sessions where extra logging overhead is unwelcome
What's inside depth 4/5
A 45 KB skill with 14 reference files, 5 scripts and CI, though no eval suite for the skill itself.
- Skills
- 1
- Reference docs
- 14
- Scripts
- 5
- Tests / evals
- 0
- Templates
- 0
- Avg SKILL.md
- 44 KB
Counted from the repository file tree. How we score depth.
Dependencies and costs
- A persistent workspace folder for the observation log
- Shell and Python for the helper scripts
- An activation instruction in CLAUDE.md or a session-start hook
- Telemetry: none found; the skill states normal operation fetches nothing and opens no URLs.
Security review
Review before installingA meta-skill that logs friction and corrections from your work sessions to a local workspace folder and proposes skill improvements in a weekly review. Its helper scripts only read and write that folder and use no network; there is no allowed-tools pre-approval. The log can contain session content, and the skill asks to be invoked before the first tool call of every session, so check where the workspace is anchored before sharing it.
Worth knowing
- Changes files outside your projectThe skill keeps its observation log in a persistent workspace, recommended at a user-scope path such as ~/.claude/skill-observations/, and appends to checkpoints.log there; env docs also offer (with user consent) to add an activation line to CLAUDE.md or a session-start hook.SKILL.md:33
What we checked
The repository (29 files) on Oct 29, 2026. We had every file and read every script.
NVIDIA SkillSpector static scan: 37 pattern hits (analysis-evasion 21, Dangerous Code Execution 5, Rogue Agent 2, Agent Snooping 1, Tool Misuse 1, Memory Poisoning 1). The 21 'incomplete referenced artifact analysis' hits are scanner artifacts on files that exist as plain markdown. The 'backdoor_persistence' YARA hit is a docs line telling the user to run crontab -l / launchctl list to see whether a scheduler exists. The subprocess and os.environ hits are in validate-skill-bundle.py's --selftest, which runs the repo's own scripts against temporary fixtures; 'git reset --hard', 'find ~/.claude/' and 'update skill' are documentation text.
A review, not a certification. Skills change after we check them — read the files before installing.
License
CC BY 4.0 — commercial use allowed with attribution.
Can it run unattended?
Not really. It observes sessions and produces a weekly review for a human to approve; it does not run tasks unattended.
Alternatives to task-observer (One Skill to Rule Them All)
Harness
revfactory/harnessA Claude Code meta-skill that designs a project-specific agent team, writes the agent definitions and skills for it, and has an evolve skill to refine them from feedback.
With caveats
No issues Depth 3/5
FAQ
What is task-observer?
It is a meta-skill, also called One Skill to Rule Them All, that logs patterns and corrections during your work and suggests updates to your other skills and new skill candidates.
How do I install task-observer?
Run npx skills add rebelytics/one-skill-to-rule-them-all --skill task-observer, then add the activation instruction from references/environments.md to your CLAUDE.md and check that it runs.
Does task-observer change my skills automatically?
No. It proposes edits in a weekly review and you decide which ones to apply.
Sources: GitHub repository (stars, license, file tree, last push Oct 28, 2026). Review based on the README and SKILL.md files as published.