Task Observer
Watches your work sessions, turning corrections, preferences and repeating patterns into skill improvements and new-skill candidates — so your skill library keeps getting better with use.
Writes are scoped to skill-observations/ and skill-updates/ subdirectories of the user's chosen shared folder; skill updates are staged for user review, never auto-installed; the skill declares no runtime external URL fetches; attribution and license are complete. Deductions: the skill demands activation before the first tool call of every session and observes all session content — a broad behavioral footprint; and its 'retry once, then try a second tool' guidance on denied writes pushes against user permission settings, weakening the confirmation boundary.
High self-consistency: the scan script self-checks parse failure (file count vs parsed count mismatch halts), frontmatter-only scans keep cost bounded, legacy logs have a migration script with rollback, incomplete installs self-report missing files. Deductions: static review cannot reproduce key paths; no committed test suite covering the skill's key paths is visible; the embedded bash snippets are unverified in execution.
Audience and scenarios are clear (knowledge workers with growing skill libraries); the README honestly declares the non-fit boundary (small setups may be better served by built-in memory), and multiple environments including a handoff-doc mode are covered. Deductions: trigger conditions are extremely broad (any multi-step task, any tool-calling session) creating clear false-trigger risk; content is deeply Claude-architecture-bound with no declared Chinese support; reachability from mainland China for the core experience is undiscussed.
Well-layered docs (SKILL.md plus on-demand references/), complete README, user guide, contributing guide and license; versioning history (pre-3.0 migration, v3.1.0), maintenance responsibility, update path and contribution-credit rules are all explicit. Deductions: no formal CHANGELOG file; some critical details (migration, activation config) live in reference files not provided as evidence, leaving hidden assumptions in a static read.
Core outputs (observation log, staged skill-update proposals, cross-cutting principles) are structurally clear and directly readable; the README compares honestly against alternatives (direct editing, built-in memory). Deductions: all effect evidence is author self-report (1300 observations, 75 skills), static review cannot verify actual output usability, and value depends heavily on user-side ritual (weekly reviews, CLAUDE.md activation), so marginal benefit is independently unproven.
The repo includes CI workflows (release bundling runs scripts/validate-skill-bundle.py), full license text, a traceable contribution-credit process, a third-party security-audit badge, and many third-party recommendation links. Deductions: no committed test suite over the skill's key paths; the author's core effectiveness claims cannot be independently checked; the audit badge content is unverifiable from static files.
- The skill is designed to observe and log content in nearly every session; confirm your session content (especially client or sensitive data) is appropriate to persist into an observation log.
- The 'retry with a second tool when a write is denied' guidance can fight your permission-approval settings; consider disabling or adjusting this behavior in strict-permission environments.
- Triggers are extremely broad (any multi-step task or tool-calling session activates it), so token cost and false triggering are real; users with small skill libraries may prefer direct editing or built-in platform memory.
- Content is deeply Claude-bound; other environments are 'compatible with caveats'. No Chinese-language support is declared; mainland-China users must assess reachability of the shared folder and platform themselves.
- Effectiveness claims (1300+ observations, 75 skills) are author self-report and independently unverified; the security-audit badge content was not verifiable in static review.
What does this skill do, and when should you use it?
Task Observer is a meta-skill ("One Skill to Rule Them All") created by Eoghan Henn (rebelytics.com) as a practical application of the Augmented Expertise methodology. It runs alongside your work, observing task execution to capture user corrections, workflow insights, and methodology worth preserving into a structured observation log. It never modifies skills directly — it produces recommendations that you review and approve. Because it observes itself too, it is self-improving. The author reports logging over 1,300 observations across 75 skills in seven months of use.
At session start it probes the workspace, sets up and scans an observation-log directory (skill-observations/observation-log/, one Markdown file with YAML frontmatter per observation); it monitors task execution, post-task feedback and meta-discussion about skills, capturing three signal types: recurring workflows worth a new skill, corrections and gaps in existing skills, and flaws in its own methodology. It runs Bash snippets on every tool call to derive observation IDs and archive resolved entries, writes observations silently to disk (with a mandatory flush after every 3rd completed todo item and at deliverable completion), and surfaces a grouped end-of-session summary asking which items to act on. Approved skill updates are staged into skill-updates/ for installation. It also maintains a cross-cutting principles file that new skills are automatically checked against.
- Users with a growing library of Claude skills who want systematic maintenance instead of manual audits of each one
- Users with no skills yet — the observer starts flagging skill candidates and helps build them
- Practitioners who repeatedly correct the AI on the same mistakes and want those corrections to become permanent skill rules
- Builders running parallel autonomous sessions in Claude Cowork or Claude Code who need skill learning that persists across sessions
- Developers integrating the methodology into existing agent setups, including user-reported Hermes and Openclaw environments
What are this skill's strengths and limitations?
- Self-observing design: improves its own methodology while improving all your skills
- You stay in control — it only recommends, never edits skills directly
- Proven at real scale: author reports 1,300+ observations across 75 skills in 7 months
- Active community: incorporates improvements from 44 contributors via 82 issues and 23 PRs
- Honestly states the boundary where small setups may not need it, rather than overselling
- No independent third-party verification of the author's usage claims
- Content includes Claude-specific concepts (CLAUDE.md, Cowork directory picker) requiring adaptation elsewhere; full experience only on Claude Cowork/Dispatch/Claude Code desktop
- SKILL.md path is truncated — referenced files in references/ are not provided, so not all rules can be fully audited
- Overhead may outweigh value in small setups with few skills
- Requires disciplined recurring review sessions to deliver value — otherwise the observation log just accumulates
How do you install this skill?
Download the .skill bundle attached to the latest release, or ZIP-download/clone the repo keeping SKILL.md, references/ and scripts/ together (installing only SKILL.md runs degraded and isn't recommended). Claude web/desktop/mobile and Cowork: upload the .skill bundle via Settings → Customize. Claude Code: place the folder at .claude/skills/task-observer/ (project-level) or in your user-level skills directory, preserving subfolders. Critical step: installing files isn't activating — add the activation instruction from references/environments.md to your CLAUDE.md (or equivalent) or install the session-start hook, then verify in a fresh session that the skill fires before the first tool call.
How do you use this skill?
Once installed and activated, the skill should invoke itself before the first tool call of every session and log observations silently as you work. At session end, ask "Any observations logged?" — the skill sometimes then finds additional improvements it hadn't logged. Schedule recurring review sessions to apply open observations (the author runs Monday, Wednesday and Friday mornings; adapt to your needs). Upgrading from a version before 3.0 triggers automatic conversion of the old single-file log in the first session. Filesystem-less environments (Claude.ai web/mobile) use handoff-doc mode, producing a structured handoff document at session end.