Skills are the New SDKs
summary
The team claims that enterprise agent teachability requires versioned, task-specific packages that encode workflow ordering, failure modes and security hygiene; they package ML prediction, model training, deployment, monitoring and CI/CD as installable skills that run across Claude Code Cursor and Codex for a Fortune 50 platform. Skills use progressive disclosure to cut tool-definition token cost by roughly tenfold versus dumping MCPs into context, address front matter/body/scripts as three runtime layers, and prove operable though authoring and verification remain costly.
AI-written summary · about
chapters (15)
- 0:00Skills as the new SDKs
- 0:46Three hidden layers of AI apps
- 1:48Infinite context windows myth
- 3:11Documentation and coding agents
- 4:15Friction to fluency modes
- 6:44Skills commoditize experience
- 7:48Teachability in enterprise AI
- 9:15What is context engineering
- 12:27Progressive disclosure
- 15:59Skills vs MCP
- 18:24When to use skills over MCP
- 20:50Three levels of skill management
- 22:56Growing ecosystem of skills
- 24:42Risks of LM-generated skills
- 25:45Closing thoughts
related talks
31mSpecialize Your Agents with Skills
16mDon't Build Agents, Build Skills Instead
01h 21mFull Walkthrough: Writing & Using Skills
01mSpecialize Your Agents with Skills
29mStop Prompting, Start Composing: Agent Skills in Practice
43mManage Skills Like Dependencies
01h 19mSkill Issue: How We Used AI to Make Agents Actually Good at Supabase
38mThe New Skill Stack for SDETs in the Agentic Era