How to make coding agents complete the development process from planning to testing
Instead of immediately generating code, the agent asks what to build and presents a design proposal. Once the user approves, a set of rules and skills keeps the agent on track while it works independently for hours. It is a framework that helps perform the entire software development process systematically.
Design and planning before coding
When a user wants to build something, the agent does not start coding right away. Instead, it uses conversation to understand the actual intent and presents the design in short units. If the user agrees with the design, it creates a detailed implementation plan that even a junior engineer can follow. This plan includes test-first principles and rules for removing unnecessary features. When the user says 'start', the agent automatically proceeds with executing and reviewing each task through sub-agents.
Automatic execution skills for each development stage
Skills that activate automatically at each development stage guide the workflow.
- Brainstorming Refines ideas and explores alternatives before writing code. It presents the design divided into verifiable units.
- Planning Breaks work down into small tasks of 2 to 5 minutes. Each task includes exact file paths and verification steps.
- Test-driven development Writes failing tests first, then writes only the minimal code to pass them. Code written before the tests is deleted.
- Code review Performs a review against the plan between tasks. It blocks progress if critical issues are found.
Supported agents and installation
It can be used with various coding agents such as Claude Code, Cursor, Codex, and Gemini CLI. Installation commands are provided for each agent, and once installed, it activates automatically at the start of a session. If using multiple agents together, it must be installed separately for each. Updates are also handled differently for each agent.
Things to know before using
This repository is released under the MIT license. Since the time the agent works autonomously can be long, it is necessary to develop the habit of checking if the plan is well-structured midway. Some features collect only the usage version through telemetry, which can be disabled via an environment variable. New skill contributions are generally not accepted, and when modifying existing skills, they must work in all supported agents.
