Developer joebew42 has released aido, a new plugin for Claude Code that enforces a rigid, spec-driven engineering workflow. The tool, which stands for AI do, is designed to prevent autonomous drift by requiring explicit human approval at every decision point. It structures feature development into five distinct stages: planning, implementation, commit organization, pull request creation, and review handling. By integrating directly with Claude Code’s plugin marketplace, aido aims to bridge the gap between raw AI capability and accountable software engineering.

The Five-Stage Gatekeeping Process

The core innovation of aido is its refusal to let the AI proceed without consent. The workflow begins with /aido:1-plan, which drafts a specification in SPEC.md.draft and refines it with the user. Only after the user approves SPEC.md does the agent move to /aido:2-implement, where it builds features using test-driven development. Subsequent skills handle commit cleanup, PR creation, and review feedback. Each stage ends with a hard stop, ensuring that the engineer retains final judgment over architecture, scope, and code quality.

Specification as the Single Source of Truth

Unlike many AI coding tools that treat prompts as ephemeral instructions, aido centers every action around a persistent SPEC.md file. This document contains the Context, Intent, Out of Scope, and numbered Acceptance Criteria (AC1, AC2, etc.). The acceptance criteria are not just descriptive; they are executable goals. During implementation, the agent writes a failing test for each criterion, keeping it red until the code passes. This ensures that the definition of done is verifiable rather than subject to the AI’s subjective interpretation of the prompt.

Verification and Repository Alignment

The plugin includes a /aido:verify skill that checks the current branch against the agreed-upon specification from a fresh context. This separation of duties means the agent that wrote the code is not the one judging its correctness. Furthermore, aido actively checks for conflicts with the repository’s existing CLAUDE.md or AGENTS.md files. If a proposed change violates architectural boundaries or existing conventions, the agent cites the specific files and recommends revision before proceeding. This mechanism prevents the AI from introducing technical debt under the guise of feature completion.

Key Takeaways

  • aido enforces human-in-the-loop approval for all major engineering decisions, preventing autonomous overreach.
  • The SPEC.md file serves as the single source of truth, linking high-level intent to executable acceptance criteria.
  • Implementation strictly follows test-driven development, with acceptance tests written before code.
  • The plugin includes safeguards to check for architectural conflicts with existing repository conventions.
  • Users must manually configure ~/.claude/settings.json to disable auto-memory and AI attribution for optimal results.

The Bottom Line

aido is a necessary corrective to the trend of unstructured AI coding, prioritizing accountability over speed. It proves that AI agents are most valuable when they are constrained by rigorous engineering practices rather than allowed to freewheel.