| name | task-observer |
| description | Meta-skill that watches your work sessions, captures corrections and judgment calls, and turns them into skill improvements automatically. Use during any task-oriented session to build and improve your skill library over time. Activates automatically when you begin multi-step work. |
Task Observer
A meta-skill that runs quietly alongside your work, notices what's working and what isn't, and produces structured improvement recommendations for your entire skill library — including itself.
What It Does
During any work session, the observer watches for three signals:
- Corrections — when you redirect, rephrase, or override Claude's output, that's a signal a skill could be clearer
- Gaps — when you do something manually that could be systematised, that's a candidate for a new skill
- Blind spots — when the observer itself misses something, it logs that too
At the end of each session, it produces a structured observation log with specific, actionable improvement suggestions.
How to Use
No setup required. Simply work normally. At the end of a session, say:
"Summarise your observations from this session."
The observer will produce a log like:
OBSERVATION LOG — [date]
Session type: [e.g., content writing, code review, research]
CORRECTIONS NOTICED
- [Skill affected]: [What was corrected] → [Suggested improvement]
GAPS IDENTIFIED
- [Task pattern]: [Description] → [Candidate skill name + draft]
SELF-IMPROVEMENTS
- [Observer methodology note]
Modes
Cowork / Claude Code: Observer writes logs to the filesystem so improvements persist between sessions.
Claude.ai web: Observer produces a handoff document at session end. Copy it and use it to update your skills manually.
Review Process
The observer never modifies skills directly. Review each suggestion and apply only the ones you agree with. You stay in control.
Examples
Correction detected:
User asked for a formal email, corrected Claude's casual opener → Observer notes: internal-comms skill needs a tone selector.
Gap detected:
User manually formatted a table three times → Observer drafts a table-formatter skill candidate.
Self-improvement:
Observer triggered too eagerly on a simple Q&A → Adds note to only activate during task-oriented sessions with 3+ steps.
Tips
- Run the observer during complex, multi-step tasks for the richest observations
- Review the log at the end of every session, not once a week — recency matters
- When the observer drafts a new skill candidate, use
skill-creator to polish it before uploading
Credits
Inspired by rebelytics/one-skill-to-rule-them-all. Licensed CC BY 4.0.
