AI Agent-Driven Development: Persistent Slice Workflow Template - Systematic Practice from Idea to Release

This article introduces bkm1291's open-source repository-agnostic persistent slice construction workflow template, designed to provide a standardized practice solution for AI Agent-assisted development. The template includes core modules such as roadmap planning, slice package management, validators, testing, and Claude/Codex handover mechanisms, addressing issues like quality control, context switching, and tool handover in AI collaborative development, suitable for individual developers, small teams, and AI-native projects.

With the improvement of AI programming assistants like Claude and Codex, software development models have shifted from traditional linear processes to a human-machine collaboration model of "humans define direction, AI executes implementation, humans supervise verification". However, the new model faces challenges such as code quality controllability, complex context switching, and seamless handover between multiple AI tools. This workflow template is designed to solve these practical problems.

Core Concept: Persistent Slice Decompose large projects into independent, verifiable, and traceable work units. Each slice includes clear scope boundaries, a complete context package, verifiable completion criteria, and persistent state records, drawing on agile ideas and optimized to adapt to AI Agent characteristics. Workflow Architecture:

1. Roadmap-driven planning layer: Structured evolution path (long-term vision, priorities, technical debt, dependency tracking) to maintain human control over direction.
2. Slice package execution layer: Self-contained work units (goal description, technical specifications, reference implementation, testing strategy) to support independent AI work.
3. Validator quality gate: Multi-layer verification (syntax, style, functionality, security) to ensure code meets standards.
4. AI tool handover mechanism: Standardized protocols (context packaging, intent transmission, constraint declaration, fallback mechanism) to support flexible switching between tools like Claude and Codex.

The repository structure reflects engineering thinking:

• .claude/skills: Claude-specific skill configurations, integrating AI capabilities into version control to ensure team collaboration consistency.
• .github/workflows: Preconfigured GitHub Actions pipelines covering the entire lifecycle of code submission, verification, and release, linked with slice states to implement automated quality gates.
• configs/: Centralized management of environment, tool, and strategy configurations, supporting multi-environment deployment and configuration overrides.
• contracts/: Defines interaction protocols between modules, serving as both design documents and constraints for AI implementation.
• exports/: Unified management of slice outputs (documents, reports, artifacts) for easy tracking and archiving.

This template is suitable for three types of users:

1. Individual developers: Provides a structured project management method to avoid "analysis paralysis" before complex features and maintain a clear development rhythm.
2. Small teams: Reduces coordination cognitive load through standardized documents and automated checks, focusing on value creation.
3. AI-native projects: Offers a directly implementable best practice starter kit, adapting to deep AI-assisted development needs.

Implementation Path Recommendations:

1. Familiarize with core concepts (slices, roadmaps, validators, etc.);
2. Conduct small-scale pilots on non-critical projects to accumulate experience;
3. Adjust slice granularity and verification rules according to team conditions;
4. Integrate existing IDE, CI/CD, and project management tools;
5. Train the team to understand the design intent of the workflow. Limitations:

• Learning curve: The new workflow requires an adaptation period, which may reduce efficiency in the short term;
• Over-engineering: Using the full process for simple projects may be redundant;
• AI capability boundaries: Effectiveness depends on the capabilities of specific AI tools;
• Maintenance cost: The workflow itself needs maintenance, so the input-output ratio should be evaluated.

This open-source project represents an important evolution of software development methodologies in the AI era, providing a "common language" and "collaboration protocol" for human-machine collaboration, helping developers maintain code quality and project controllability while enjoying the efficiency dividends of AI. As AI programming assistants improve their capabilities, such workflow templates will become increasingly important, serving as a high-quality reference implementation for teams exploring best practices in AI-assisted development.