- Author
- Addy Osmani (Member of Technical Staff at Anthropic, per the page)
- Published
- May 3, 2026
- Kind
- Personal blog post introducing the author's own open-source project
- Original
- addyosmani.com/blog/agent-skills/
- Subject
- Harnesses and agent loops
- Themes
- Handing over control
What it is
Addy Osmani wrote this post about Agent Skills. Agent Skills is a set of markdown workflows for coding agents. The post explains the design choices and how they map to the software lifecycle. It also lists patterns worth copying without installing anything.
What it says
- Osmani says AI coding agents skip specs, tests and reviews by default.
- A skill is a workflow with checkpoints and an exit criterion, not reference prose.
- Each skill has a table of excuses for skipping steps, with written rebuttals.
- Skills load by lifecycle phase through a router, not all twenty at session start.
- The skills encode Google practices such as Hyrum's Law, DAMP tests and Chesterton's Fence.
- The repo has twenty skills in six lifecycle phases, with seven slash commands on top.
- The post offers five rules for any AGENTS.md, starting with surfacing assumptions before building.
What evidence it gives
The post offers argument and a description of the repo. It gives the repo star count as a sign of interest. It reports no tests of whether the skills improve agent results.
Interests and limits
Osmani works at Anthropic on Claude Code. The post promotes the author's own repo and book. The post tests none of its claims about agent behavior. The claims about agents skipping steps are stated as argument, with no data.