Building software with AI

Work ethic based on experience

The below post is 100% human-generated. Proof or it didn’t happen.

1. It's time to share my insights after a few months of working with LLM for software development...
   
   I agree with starting with the planning mode. What does it look like in practice?
   
   - If it's a project for a client, I gather requirements during the conversation(s), prepare a document based on that (mostly on my own, sometimes with the help of a summary/transcription) - Google Docs. Based on the comments on this document, we get to the specifics.
   
   - If the project is based on my ideas, thoughts, or findings, I want to practice new solutions, etc. I conduct a brainstorm with the agent I prepared for these purposes - mvp-e which was created based on suggestions from one of the posts.
   
   After this stage, the PLAN.md file is created.
   

2. For some time now, I have been an ardent fan of the entire Cloudflare ecosystem. Based on its capabilities and tech stack, and drawing on Mat's experience, I refined the template from which I start, available here. This template at this stage has 3 agents and useful skills and rules.
   

3. I use Claude Code and the Opus model (currently the latest version), and I use Visual Studio Code. I often rely on Google Gemini 3 Pro (I think it has more imagination than Claude, who is perfectly dry and technical).
   

4. I am a fan of deep planning and short programming sessions. Of course, you can't always plan everything, but usually, you can plan a lot, especially when you know what needs to be done (see point 1).
   
   - I am launching the dd-w agent (design doc writer), which is based on the prepared PLAN.md. It creates a multi-step implementation plan for me. PLAN.md is a document that is too general and too extensive to build a project based on it (which is not just another fitness-todo-calendar-tracker-newsletter-email-parser-generator) - fugazii fugazii.
   
   (here an interesting solution is to instruct the agent to prepare a simulation/prototype, clickable—if the project also includes UI.)
   
   - The project documents (design docs) created this way, I keep in the /docs directory; example,
   
   - thanks to the suggestions read from /insights Claude Code, I prepared a project document consistency verifier, which checks the prepared documents (whether something is missing, whether there are ambiguities and inconsistencies, etc.)
   

5. When I already have PLAN.md divided into verifiable stages (in the form of project documents), checked for consistency, sometimes there is a need to divide them into smaller ones. For this purpose, I use the dd-w agent again to break it down into smaller steps (ideally linked to deployment), narrowing the context window, usually around 50k tokens, sometimes some stages reach 80/90k tokens.
   

6. The final step is task division (if the project is large) and their implementation based on the dd-i agent(s), based on the dedicated command.