Unlocking New Superpowers: AI-Assisted Coding Workflow with Cloud Code and GitHub

In recent weeks, I’ve been experimenting with a powerful AI-assisted coding workflow using Cloud Code and GitHub to build a new web application. This workflow has truly unlocked new superpowers for me as a developer, streamlining how I plan, create, test, and deploy software. In this blog post, I’ll walk you through the workflow, explain why it’s effective, and share practical tips on how you can implement it in your own projects.

The High-Level Workflow: Plan, Create, Test, Deploy

The workflow is elegantly simple yet powerful, revolving around the four classic phases of the software development life cycle:

1. Plan: Create and refine GitHub issues to clearly define atomic, manageable tasks.
2. Create: Use Cloud Code’s custom slash commands to generate code that addresses the issues.
3. Test: Run automated tests, including UI tests powered by Puppeteer, to ensure quality.
4. Deploy: Commit and push changes to GitHub, open pull requests (PRs), and merge after review to deploy via platforms like Render.

By leveraging GitHub flow — a tried-and-true workflow designed for small teams — and integrating AI-powered coding assistants, this process makes it feasible for a “team” of one human and one AI to build complex applications efficiently.

Creating and Refining GitHub Issues: Your Project’s Backbone

The first step is to capture all work as GitHub issues. I started by dictating initial requirements and then worked with Claude Code to translate them into issues. However, I quickly learned the importance of granularity and specificity in these issues. The more atomic and well-defined the issues, the better Claude Code could handle them.

This phase reminded me of my managerial days, as I found myself writing detailed specs, reviewing code, and leaving feedback for improvements — essentially playing the role of an engineering manager. This approach ensures that the AI-generated code is aligned with your vision and standards.

Setting Up a Solid Foundation: Testing and Continuous Integration

Before diving into rapid development, it’s crucial to establish:

• A robust test suite to verify that new changes don’t break existing functionality.
• Continuous Integration (CI) using GitHub Actions to run tests and linters automatically on every commit.
• Puppeteer integration to simulate user interactions and test UI changes in a real browser environment.

Using frameworks like Rails (with its MVC architecture and integrated testing) makes it easier for AI coding agents to work on modular code sections rather than sprawling, monolithic files.

Custom Slash Commands: Automating the Plan-Create-Test-Deploy Cycle

Cloud Code slash commands are prompt templates with command-line arguments that instruct the AI on how to handle each issue. My main /process-issue command breaks down into:

• Plan: The AI reviews the GitHub issue, searches previous related work and pull requests, and creates a detailed plan with atomic tasks using “scratchpads” (dedicated planning files).
• Create: The AI writes code addressing the plan.
• Test: The AI runs tests to verify its work.
• Deploy: The AI commits changes, opens a PR, and optionally requests or performs code reviews.

This structured approach ensures clarity and accountability throughout the development cycle.

The Human-AI Partnership: Code Review and Responsibility

One common concern about AI-assisted coding is trust — how do you know what the AI wrote is correct? The answer remains the same as with any developer: you must review the code.

I’ve found it helpful to:

• Read through pull requests carefully.
• Optionally have Claude Code perform a PR review using a separate slash command, emulating expert styles like Sandy Metz’s principles for maintainable code.
• Rely heavily on tests to catch regressions and unexpected issues.

While I sometimes let Claude commit code directly, I make sure tests pass and the changes look good before merging.

Managing Context: The Importance of /clear

After completing and merging an issue, I always run /clear in Cloud Code to wipe the AI’s context window. This forces Claude to start fresh on the next issue, relying solely on the issue description, scratchpads, and repository history — no leftover “working memory.”

This practice helps:

• Maintain focus on the current issue.
• Reduce token usage.
• Improve AI performance and accuracy.

Using Claude in GitHub Actions vs. Cloud Code Console

Anthropic recently launched Claude integration via GitHub Actions, allowing you to tag Claude directly on GitHub. While this is convenient for small tweaks and copy changes, I prefer using Claude Code in the console for more significant development work because:

• GitHub Actions usage incurs metered billing, even on premium plans.
• The console provides better insight and control.
• For large code changes, the console-based approach is more efficient and manageable.

Running Parallel Agents with Work Trees

Work trees let you run multiple instances of Claude on different branches simultaneously, similar to multitabling poker. However, I encountered some challenges:

• Permission approvals need to be repeated for each new Claude session.
• Managing multiple work trees can feel clunky and increase babysitting overhead.
• For my project, sequential work on a single instance sufficed.

Still, as projects grow or teams scale, work trees offer a way to increase parallelism in AI-assisted development.

Final Thoughts

This AI-assisted workflow combining Cloud Code, GitHub, and Puppeteer has revolutionized how I build software. It marries the power of classic software development principles with cutting-edge AI coding assistance to create a cycle of continuous, manageable progress.

If you want to get started, focus on:

• Writing clear, atomic GitHub issues.
• Setting up a solid test suite and continuous integration.
• Creating custom slash commands to automate planning, coding, testing, and deployment.
• Embracing your role as the reviewer and planner to guide the AI effectively.

For more insights, I recommend checking out my related video on Claude Code pro tips and reading Thomas Tacic’s excellent post on AI-assisted coding skepticism.

References and Resources

• Thomas Tacic’s post: All of My AI Skeptic Friends Are Nuts
• Anthropic Cloud Code Documentation
• GitHub Flow
• Sandy Metz’s Principles for Writing Maintainable Code

Harness the power of AI in your development process — it might just unlock new superpowers for you too!