Motivation

There is still ongoing hype on AI aided programming and the hype seems to have no end. Indeed, I also use AI support to enhance my performance as an software engineer. In my current job position I am allowed to use Github Copilot. Recently I wanted to improve my performance even further and I was searching for the way to go full into agentic development. Inspired by the following blog post by Harper Reed where he described his agentic development process in Claude Code I wanted to setup something similar with use of Github Copilot. In this post you can find my setup and results achieved so far.

Git Hub Copilot - agentic options

Github Copilot offers 3 standard modes of work:

• Ask mode. You can use Copilot to ask for help or clarification based on highlighted code by you.
• Edit mode. In this mode you provide context files to Copilot and he adjusts files based on your request.
• Agent mode. In this mode Copilot can work autonomously based on your high level request. Copilot will make-up the solution, find the relevant files, right tests and so on.

Beside that, the recent version of Github Copilot gives as possibility to automate some repetitive activities. Here we have basically 3 options:

• custom instruction file. You can create .github/copilot-instruction.md file where you will provide some general rules for AI behaviour: “generating code, performing code reviews, or generating commit messages.”
• prompt files. Using prompt files you can define reusable prompts for common repeatable tasks: code review, angular component creation. You place them in the folder .github/prompts
• chat modes. Chat modes give you possibility to define AI persona for specific roles or tasks. you place them in the folder .github/chatmodes

You can also conntect MCP (Model-Context-Protocol) based tools: Link

More information you can find in the official documentation: Link

My setup

With these all options it is clear that you can create agentic mode that will resemble Claude Code. I really liked idea presented in blog to use LLM for creation of specification and prompt-plan for AI. In my recent project I tried to mimic this setup in Github Copilot. Finally I created the following .github structure:

Agentic programming usage

I created this setup specifically for the recent small project that I wanted to develop. The goal was to give whole control over the process to the AI.

Creation of spec.md

Together with AI we brainstormed nice and tailored specification using chat mode idea-spec-honing.chatmode.md. So far so good.

Creation of prompt-plan.md

In the next step AI generated nice prompt-plan.md. In this step I used tdd-prompt-plan-create.chatmode.md. The biggest test of AI agentic mode will come however in the next step.

Agentic work on application

In this step I used tdd-prompt-plan-work-on-it.chatmode.md to ask Copilot to develop the application in an iterative way. AI managed to go through all 16 prompts from prompt-plan.md with small help from my side. During the development AI regurarly was creating unit tests and verified its correctness. I did barely none review of the functionality after each prompt. I reviewed the whole application after all 16 prompts.

The end result was not ideal:

• some functionality was overlooked
• some feature where not working correctly

Overall it was the good starting point for further improvements, but for sure it was not ready solution.

Conclusion

AI Agentic programming clearly have some potential, but it isn’t ideal. There is still a need for software engineer who will:

• resolve errors that are to difficult for AI
• supervise AI so that the context is not lost
• verify if everything is implemented according to the specification.

In the future agentic vibe coding sessions for sure I will change:

• verifification of correctness will be done after each prompt.
• I will run each prompt in a new chat window so that the context window will not be overloaded with data from the past prompts