project-ai-docs

The quiet killer in AI-agent-assisted development is not low quality in a single response, but context drift across sessions. A new window opens and the agent has forgotten last round of architecture decisions, cannot tell which file paths were actually verified versus merely assumed, and all hard-won knowledge evaporates the moment the chat log scrolls away.

project-ai-docs gives every project a persistent docs/.ai/ knowledge base. index.md is the single entry point and a priority routing table; documents are organized into four layers (core, modules, config, best-practices) with explicit token budgets. Hard constraints sit at the very top of the index, ahead of every technical detail. The agent reads the index before acting, point-reads only the relevant slice for the current task, and must write back updates plus a changelog entry when the task ends, so knowledge lives in files rather than in session memory. A human/agent separation keeps the user-facing README and the agent-facing docs/.ai namespace cleanly apart, and every recorded fact is required to be a tested path, version, or command that can be verified.

Built for individual developers and teams that rely on Claude Code, Codex, DSH, or similar agent hosts day to day, and for multi-module, long-cycle projects where cross-session project awareness is non-negotiable.