Context Is the New Code — Patrick Debois, Tessl

Debois's central thesis (0:55) is that since AI agents now generate most code, the developer's primary artifact has shifted from code to context — prompts, instructions, skills, and documentation that guide the agent. The quality of output depends entirely on the quality of context provided.

Modeled as an infinity loop (2:45), the CDLC encompasses five phases — generate, test, distribute, observe, and adapt — mirroring the Software Development Life Cycle but applied to context artifacts rather than code.

Debois describes converting large code helpers into declarative "skills" (1:15) — natural language workflows that instruct the agent step-by-step (e.g., detect package manager, detect ecosystem, onboard user). These skills solve problems that would be impractical to hard-code across all permutations.

Just as code has linters, unit tests, and end-to-end tests, context needs analogous testing layers (4:55) — format validation (linting), comprehension checks (Grammarly-like feedback from LLMs), LLM-as-judge evaluations checking generated code against criteria, and sandbox-based end-to-end tests where an agent judge can actually execute and curl endpoints.