AgenticWorkflow: A Complete AI-Driven Development Workflow from Idea to Code

AgenticWorkflow is an end-to-end AI-assisted development framework that turns rough ideas into mergeable code through a combination of an interactive planning phase and an autonomous execution phase. Its core is a two-stage architecture: first, interactive planning using three skills (/grill-me, /to-prd, /to-issues), then autonomous execution via the Ralph loop. The original author is kimlundjohansen, and the project is open-sourced on GitHub (link: https://github.com/kimlundjohansen/AgenticWorkflow).

With the popularity of AI-assisted programming tools today, developers face the challenge of effectively integrating AI agents into actual development workflows: traditional workflows require extensive manual intervention and context switching, while pure AI automatic coding lacks planning and quality control. AgenticWorkflow aims to resolve this contradiction by proposing a hybrid workflow where humans focus on thinking, planning, and decision-making, and AI agents handle repetitive tasks—improving development efficiency while maintaining code quality.

AgenticWorkflow adopts a two-stage architecture:

1. Interactive Planning Phase: Developers engage in in-depth conversations with AI assistants (such as Claude Code or GitHub Copilot) to refine plans using three skills:
   • /grill-me: AI uncovers requirements through relentless questioning, explores the codebase to reduce cognitive load, and forms a definitive project plan;
   • /to-prd: Generates a structured PRD from the conversation content and publishes it as a GitHub issue;
   • /to-issues: Breaks down the PRD into ordered vertical-slice sub-issues, marked as "afk" for use in the execution phase.
2. Autonomous Execution Phase: After the developer steps away, the Ralph loop automatically processes sub-issues marked as "afk" in a Docker sandbox, completing coding, testing, committing, pushing, and other operations until the queue is empty.

Technical Implementation: The project is a set of tools written in Shell, including three skill definitions under .claude/skills/ and loop scripts in the ralph/ directory, which can be integrated with Claude Code and GitHub Copilot. Environment requirements: Docker Desktop (sandbox), GitHub CLI (gh) with authentication; note to set GH_TOKEN instead of the expired GITHUB_TOKEN. Tag Model:

• afk: Triggers the local Ralph loop;
• hitl: Human-in-the-loop, Ralph skips;
• agent:implement: Optional parallel path, executed in the cloud via GitHub Actions.

Applicable Scenarios: Scenarios where ideas need to be quickly transformed into runnable code. Value includes:

1. Structured planning reduces later rework and scope creep;
2. Automated execution frees developers to focus on creative work;
3. Docker sandbox provides isolation and security;
4. Structured PRD and sub-issues make project status transparent, reducing code quality inconsistencies in team collaboration.

Limitations:

1. Currently mainly oriented towards the .NET ecosystem (depends on dotnet test/build), migration to other languages requires adjustments;
2. The Ralph loop processes only one task at a time, so completing the queue for large projects takes a long time;
3. Automatically generated commit messages may lack context and narrative, requiring manual review and supplementation.

AgenticWorkflow represents a new approach to the deep integration of AI agents into development workflows—it is not just a simple code completion tool, but a partner that undertakes the full task execution cycle. Through the two-stage workflow, it retains the advantages of human planning and decision-making while leveraging AI's execution efficiency and consistency. As AI coding assistants improve their capabilities, this hybrid workflow model will become more common, providing actionable implementation references for AI-assisted development.