AndThen: A Lightweight Spec-Driven Development Toolkit for AI Programming Agents

AndThen is an open-source toolkit for AI programming agents. Its core concept is Spec-Driven Development, which shifts AI programming from direct coding to a model of writing specifications first before execution. It provides flexible layered workflows, a complete set of skills, agent agnosticism, and Agent Teams coordination capabilities. It aims to promote AI programming from improvisation to engineering implementation, solving problems like requirement drift and missing acceptance criteria in complex feature development.

AI programming assistants are rapidly popularizing, but direct code generation often leads to requirement drift, missing acceptance criteria, and hard-to-verify outputs when handling complex features. The AndThen project redesigns the working model suitable for AI agents through the concept of Spec-Driven Development, improving the engineering level of AI programming.

AndThen's core proposition is 'Write specifications before coding, agents execute autonomously', producing a structured Feature Implementation Specification (FIS). It provides four flexible workflows:

1. Quick Path: Suitable for minor fixes, skipping the specification phase;
2. Feature Workflow: Suitable for complex single features, including specification generation and execution;
3. Manual Planning Workflow: Suitable for multi-feature MVPs, with fine-grained control per story;
4. Automated Planning Workflow: Agents automatically orchestrate execution plans, supporting gradual adoption.

AndThen has a skill set covering the entire development cycle:

• Requirement phase: clarify (requirement clarification), wireframes, etc.;
• Planning phase: spec (generate FIS), plan (generate PRD), etc.;
• Implementation phase: exec-spec (execute according to specification), exec-plan-team (Agent Teams parallel execution), etc.;
• Review phase: review (comprehensive review), remediate-findings (automatic repair), etc. In addition, the intelligent document index (CLAUDE.md defines output locations) adapts to project structures; agent agnosticism supports tools like Claude Code, Codex CLI, and Aider; the Agent Teams feature enables parallel multi-agent coordination (experimental flag required).

AndThen has a built-in implementation→verification→evaluation cycle, with automatic retries on verification failure. Verification includes code review, test execution, etc. The practical usage process:

1. Requirement clarification (generate requirements-clarification.md);
2. Generate specification (feature.md includes requirements, acceptance criteria, etc.);
3. Execute specification;
4. Review (optional);
5. Fix issues (if needed).

AndThen is an experimental project under rapid iteration; APIs and skill interfaces may change. Installation methods:

• Claude Code plugin: /plugin marketplace add IT-HUSET/andthen → /plugin install andthen;
• Other agents: Run ./scripts/install-skills.sh to export skills; Initialization command: /andthen:init generates CLAUDE.md; optionally run map-codebase to analyze the codebase.

AndThen explores the evolution of software development processes in the AI era, providing a structured framework for AI to participate in the entire lifecycle. It is recommended for developers to try, but caution should be exercised when evaluating risks in production environments. The project code is hosted on GitHub: https://github.com/IT-HUSET/andthen.