How can coding agents the exact context they need to build a project without repeating yourself on every single prompt?
In this video, Annie Wang and Tilde Thurium break down the markdown-based specification files that define how your AI agents should think, design, and execute tasks across your codebase:
* *DESIGN.md:* Visual guardrails, UI design systems, and styling tokens.
* *AGENTS.md:* Global project guidance, coding standards, and build/test instructions.
* *SKILL.md:* Reusable, domain-specific tasks and workflows.
Learn how agent specifications differ from standard prompt and context engineering, see benchmark data on token savings and performance, and discover best practices for scoping shared team rules versus personal developer preferences.
Chapters:
00:00 – Introduction & why we need agent specifications
01:00 – Specifications vs. prompt & context engineering
02:50 – DESIGN.md: Visual guardrails & design systems
04:56 – AGENTS.md: Project guidance & global behavior
06:12 – Benchmark study: How AGENTS.md impacts performance
07:28 – SKILL.md: Domain-specific tasks & reusable prompts
08:46 – Best practices: Separation of concerns & consistency
10:17 – Scoping preferences: Project layer vs. user layer
11:43 – Real world examples: Google ADK & DeepMind
12:12 – Conclusion & wrap up
Resources & Links Mentioned:
* Context engineering deep dive with Smitha → https://g.dev/ai/context-engineering-explained
* Skills vs. MCP deep dive → https://g.dev/ai/skills-vs-mcp
* Google Labs DESIGN.md specification (GitHub) → https://g.dev/ai/design-skils
* Google Agent Development Kit (ADK) → https://g.dev/ai/adk-python
* DeepMind Simply repository → https://g.dev/ai/simply-repo
* Agentic AI Foundation benchmark study → https://g.dev/ai/benchmark-study
🔔 Subscribe to Google Cloud Tech → https://goo.gle/GoogleCloudTech
Let us know in the comments how you’re using AGENTS.md, DESIGN.md, or SKILL.md in your development workflows!
Speakers: Tilde Thurium, Annie Wang
