A developer going by Sukitly has released Agentic Docs Templates, an open-source GitHub template repository designed to impose structured documentation workflows on AI coding agents. The project addresses a recurring frustration among developers using tools like Claude Code, Cursor, OpenAI Codex, and Gemini CLI: agents that lack <a href="/news/2026-03-14-contextual-commits-open-standard-ai-agent-decision-context-git-history">persistent context</a>, ignore architectural conventions, and produce code that gradually drifts from a project's intended design. The template is language- and framework-agnostic, centering on two core files — AGENTS.md for agent instructions and workflow rules, and ARCHITECTURE.md for project structure and layering conventions — supported by a broader docs/ directory covering application state, decision logs, testing strategy, and product specifications.
The system enforces five mandatory rules on any AI coding agent working within the project: read documentation before touching code, create a Design Doc or Execution Plan when specific criteria are met, present planned changes and await user approval before executing, self-review and keep docs in sync after every change, and apply test-driven development for core business logic. The template distinguishes between two document types — Design Docs for new modules or multi-approach decisions, and Execution Plans for changes spanning three or more modules, database migrations, or ordered dependencies — each backed by Markdown templates with required frontmatter fields and sections. A Python documentation integrity checker script (check-docs.py) is included to verify that links, index files, and referenced paths remain consistent.
A practical feature aimed at lowering the adoption barrier is the bootstrap prompt: a single command that instructs any AI coding agent to fetch a bootstrap.md file from the repository, analyze an existing codebase's tech stack, architecture, and conventions, then <a href="/news/2026-03-15-wikigen-go-cli-github-wiki-claude-code">auto-generate all documentation files</a> pre-filled with real project content rather than empty placeholders. Compatibility across the major agent tools is handled through a simple naming convention — Claude Code users rename or symlink AGENTS.md to CLAUDE.md, while Cursor, Codex, and Gemini CLI consume AGENTS.md natively. The project is MIT-licensed and cites OpenAI's Harness Engineering approach to AI-assisted development as a primary inspiration.
The explicit citation of OpenAI's Harness Engineering post is worth following up on: that piece, published by OpenAI's internal engineering team, argued for using structured context files and documented conventions as the primary lever for improving agent output quality — ahead of prompt tuning or model selection. Sukitly's implementation takes that thesis and packages it as a ready-to-fork template, with the AGENTS.md file serving the same role that OpenAI's internal context documents serve in the Harness approach. Whether the five-rule constraint set is tight enough to hold across complex multi-agent workflows remains untested in public, but the repo's structure gives teams a concrete starting point to fork and adapt rather than design from scratch.