Task Observer runs alongside your work and watches for friction: a correction you make, a preference you state, a rule the agent breaks, or a workflow you repeat. Each finding is written to a file in an observation log with its source, then surfaced as a grouped summary at the end of the session.
A periodic review turns open observations into staged edits of your skills and flags repeating work as candidates for new skills. Nothing is installed without you: changes are made on a staged copy and handed over for approval.
What You Get
- A skill-observations folder in a stable workspace with one Markdown file per observation, plus archive, checkpoint and session-start logs
- A session-start scan that reports open observations, a due review and staged updates
- A review procedure that turns observations into staged skill edits and new skill candidates
- Reference files for activation setup, signals worth logging, skill families, authoring and migration
- Ten helper scripts that create, edit, scan and validate observation files
Common Use Cases
- Capture the corrections you give an agent so they reach the skill that should have prevented them
- Spot repeated workflows that deserve their own skill
- Keep a team’s skills improving from real sessions with a reviewable history
- Fix a rule the agent keeps breaking with a hook or check on the second violation
What Makes It Unique
It treats improvement as a pipeline with evidence at each step: observe, log, review, stage, install. Edits are never made to live skills, only to staged copies, and a second violation of the same rule requires a hook or check as the fix. The README reports more than 2,500 observations logged across 87 skills by its author and credits over 100 contributors.
Under The Hood
The skill is a core SKILL.md plus reference files loaded on demand and ten helper scripts, mostly bash with two Python tools. They create the workspace folders, scan observation frontmatter at session start, assign observation ids, write and edit observation files, stage patches, diff staged updates against installed skills and validate skill bundles. It reads and writes files in a persistent workspace folder you pin, and expects the agent to run these scripts through the shell. Normal operation fetches nothing from the network; a few opt-in features, such as a version check of a foreign skill or a feedback pre-flight, are enabled by one line in your instruction file. Activation is best set up with a session-start hook and an instruction block in CLAUDE.md, since description matching alone tends to under-trigger.