AI agent workflow skill
Work Skill
A compact standalone skill for improving a real target without treating delegation as unlimited permission.
Make the working contract portable
The repository is named work-skill and its installable artifact is the work directory. Its README presents Improve as the working method: understand the target’s purpose, identify what limits it most, make a justified change, verify it and preserve the reasoning. The separation between repository identity and installable skill identity is deliberate; users should follow the source’s current installation contract.
Keep the human boundary explicit
The default workflow pauses for intent confirmation and implementation approval. An operator can delegate routine pauses for a bounded run while reserving consequential actions such as publishing, deleting data or changing the destination. That is a more precise contract than assuming that a request for autonomous work authorizes every action that could make the task easier.
Retain the prediction and result
A useful improvement record contains what was expected before the change, what was actually checked and what remains uncertain afterwards. The skill writes that semantic trail beside the target. Durable direction is optional for a one-off task, but when supplied it remains operator-owned. The record supports a later challenge to the reasoning rather than merely celebrating that an edit occurred.
What the standalone skill does not claim
Markdown instructions do not themselves create an executor sandbox, independent surveillance or structural enforcement of every boundary. The source explicitly distinguishes an agent-authored trail from optional external capture. It also identifies harness discovery and correlation as integration work rather than silently claiming those capabilities. The skill can be used without installing a harness, and its installation is not permission to start routing private model traffic through another process.