A Dev.to post describes “Context Architecture,” arguing that the repository itself functions as the effective context for AI coding agents. The author recounts working in Skyward’s monorepo with AI agents in 2025, where repeated prompt instructions to avoid certain patterns or reuse existing components were frequently ignored. The author attributes this to the agent following what it finds in the codebase and existing documentation rather than what was stated in the prompt, especially for subagents that start with less conversational context.

The post says the initial fix—adding more rules and example text to prompts—only helped temporarily, because the underlying repository still contained contradictions such as stale AGENTS.md/CLAUDE.md files, outdated comments, and modules marked deprecated but still imported. The author then proposes shifting from readable structure alone to “verifiable context,” where claims about the repository must be enforced by mechanisms that fail if the claims become false.

The author frames this as extending “Screaming Architecture” (organizing by domain intent) with four pillars: domain-driven structure, embedded boundary context docs, turning intent into mechanisms via tests/types/lint, and making capabilities discoverable through predictable paths. Enforcement is described as layered checks including type checking, linting, tests, and human review for semantic truth.