Stop Prompting, Start Composing: Agent Skills in Practice
summary
Claims that skills turn capability into competence with concrete gains: modular skill.md folders, progressive disclosure of references, and moving repeatable work into deterministic scripts. Implementation uses a LangGraph harness around an Archimedes agent with an expert search skill and evaluation via DeepEval G Evals. Results include fewer orchestration nodes, leaner context windows, faster feedback to change skill YAML and references, and measurable quality improvement when running A/B skill evaluations before promotion.
AI-written summary · about
chapters (14)
- 0:00Agentic AI experiments intro
- 1:09Need for agent skills
- 4:59Agent skills explained
- 7:47Archimedes enterprise architect case
- 10:35Challenges without agent skills
- 11:40Benefits after implementation
- 14:30Invocation and structural challenges
- 20:07Expert search and sub-agents
- 21:52Early exits and constraints
- 22:54Surprises and new roles
- 24:18From prompting to composing
- 25:22Evaluating agent skills
- 26:46Python example with DeepEval
- 28:51Stop prompting start composing
related talks
31mSpecialize Your Agents with Skills
01h 19mSkill Issue: How We Used AI to Make Agents Actually Good at Supabase
21mBuilding Great Agent Skills: The Missing Manual
21mBest practices are trapped in your head and your AI has no idea
01mSpecialize Your Agents with Skills
16mDon't Build Agents, Build Skills Instead
01h 21mFull Walkthrough: Writing & Using Skills
27mSkills are the New SDKs