1000 Cursor Prompts Failed, Desktop Commander Succeeded: AI Coding Tips from a Real Startup Founder

Key Concepts

• AI-assisted coding: Using AI models like Claude for code refactoring, debugging, and development.
• Desktop Commander: A tool that allows AI models to interact with the local file system without API costs.
• Context window: The amount of information an AI model can consider at once.
• Null safety: A feature in Dart and Flutter that helps prevent errors caused by null values.
• Microclimate control: Balancing temperature and airflow in different rooms of a house.
• MCP (Managed Code Project): A project that uses AI to manage and modify code.
• Token: A unit of text used by AI models to process information.
• Sub-agents: Delegated tasks to smaller AI agents within a larger AI system.
• D3.js: A JavaScript library for creating data visualizations.
• Next.js: A React framework for building web applications.
• API costs: The cost of using an AI model through an API.
• Vibe coding/Grind coding: Iterative coding process with AI assistance.
• Knowledge files: Files containing information that an AI model can use to improve its performance.
• End-to-end tests: Tests that verify the functionality of an application from start to finish.
• Continuous Integration and Continuous Delivery (CI/CD): A set of practices that automate the software release process.
• Hallucination: When an AI model generates information that is not based on reality.
• Compaction: A technique used to reduce the size of a context window by summarizing information.
• Awareness: A state of being present and aware of one's thoughts, feelings, and surroundings.

Aircore and AirPro: A Home Climate Control System

• Michael Ajamian is the founder of Aircore, a tech startup developing AirPro, a hardware and software system for home climate control.
• AirPro uses sensors and smart vents to automatically balance temperature and airflow in different rooms, improving comfort and energy efficiency.
• The system consists of hardware installed in air vents, room sensors, and a mobile app for user interaction.
• The app allows users to view temperature data, configure settings, and control the system.
• COVID-19 significantly impacted the project, leading to funding issues and a halt in development.

AI as a Catalyst for Restarting Development

• Michael, with a semi-technical background, decided to leverage AI to restart development on a lower budget.
• He initially experimented with Claude 3.7 and other AI tools to refactor the buggy Android app and build an iOS version.
• He emphasizes the importance of research and planning before using AI for complex coding tasks.
• Example: Implementing databases for the website required a specific framework to prevent Claude from going off on tangents.
• He learned to ask exploratory questions in new chats to avoid reusing old context and getting stuck on previous solutions.

The Desktop Commander Revelation

• Michael faced significant challenges using Cursor, burning through two 500-prompt subscriptions without success in refactoring the app.
• He discovered Desktop Commander and was intrigued by its ability to interact with the file system without API costs.
• Using Desktop Commander with Claude, he successfully migrated the app to null safety, enabling it to build and run.
• This experience was a "night and day revelation" and a "game changer" for him.
• He left his first-ever YouTube comment on the video showcasing Desktop Commander, expressing his gratitude.

Cost Analysis and Context Window Limitations

• Cursor's standard plan burns two requests for Claude models and has a limited context window (120,000 tokens).
• Desktop Commander, by allowing direct access to Claude's full capabilities, provided a broader reasoning and thinking beyond just coding tasks.
• Michael believes coding agents like Cursor are tuned for clean coding but lack the broader reasoning capabilities of Claude used with Desktop Commander.
• He acts as a customer to a software company, providing feedback and observations to guide the AI's work.

App Demo and Data Visualization

• Michael demonstrated the AirPro app running on iOS (emulated on a Mac), showing room-specific sensor data and configuration options.
• He also showcased a D3 graphing project developed using Desktop Commander and Claude, visualizing temperature data from his home.
• The D3 graph queries a Superbase database and displays data for different devices over various time periods.
• He plans to integrate the D3 graph into the Flutter app and the website's admin panel.

The Power of Planning and Iterative Refinement

• Michael used Desktop Commander to analyze the codebase of his website (Next.js 14) and the D3 graphing project (Next.js 15).
• Claude recommended downgrading the D3 project to Next.js 14 to avoid redesigning the entire website.
• He describes the coding process as "vibe coding" or "grind coding," similar to sculpting, where each prompt refines the code.
• He emphasizes the importance of not over-refining and backing up if necessary.

Managing Packages and Build Environments

• Michael advises against auto-allowing command executions in Desktop Commander to avoid unwanted package installations.
• He suggests manually reviewing each command to understand its impact on the system.
• He acknowledges the learning curve involved in managing packages and build environments.
• He expresses a desire for a "time travel" feature to revert to previous states of the system.

Testing Strategies and End-to-End Tests

• Michael used existing tests from the original app project to validate the null migration.
• He lacks the knowledge to create tests from scratch but is interested in learning how to implement them properly.
• The interviewer suggests using end-to-end tests that inspect the UI and simulate user interactions.
• He recommends creating tests that fail when the functionality is broken and succeed when it is working.

Knowledge Management and Context Window Optimization

• Michael initially loaded a large amount of data into the project knowledge, but found it consumed the context window quickly.
• The interviewer suggests using a single file with a table of contents pointing to other knowledge files.
• This allows the AI to selectively read relevant files based on the user's query.
• It also enables the AI to update its own knowledge by modifying the files directly.

The Importance of Staying Updated and Avoiding Bad Information

• Michael encountered issues with outdated file paths in his Ozero authentication documentation.
• Claude initially used the correct path but reverted to the old one due to conflicting information in the project knowledge.
• He emphasizes the need to ensure that all documentation is updated to avoid inconsistencies.

The Evolving Landscape of AI and the Potential for Skill Loss

• Michael expresses concern that people may become overly reliant on AI and lose their skills.
• The interviewer argues that this has happened throughout history with other technologies.
• Michael acknowledges that AI has significantly expanded his knowledge and understanding of code.
• He worries that complacency could hinder people's motivation to learn and retain knowledge.

The Dopamine-Driven Nature of Learning and Achievement

• The interviewer discusses the role of dopamine in motivation and achievement.
• He suggests that successful humans are dopamine addicts who have found meaningful ways to use it.
• Michael relates this to his experience with video games and his newfound passion for coding with AI.
• He finds the iterative process of coding with AI to be highly rewarding and addictive.

The Importance of Awareness and Stepping Back

• The interviewer emphasizes the importance of awareness and regularly stepping back to assess one's goals.
• He suggests that this helps to ensure that one is pursuing meaningful goals and not simply chasing dopamine.
• Michael applies this concept to his use of Desktop Commander and Claude, stopping periodically to evaluate progress and direction.

UX Tips and Troubleshooting

• Michael shares a tip for dealing with context window limitations: if a chat crashes, it may be possible to get a few more prompts out of it.
• The interviewer suggests using the "edit" button to branch off a previous prompt in a filled-up chat.
• He also notes that Claude sometimes gets ahead of itself, asking about things it hasn't yet shown.
• He mentions that there are two types of mistakes that are happening with Claude when it's trained: losing connection and capacity issues.

Desktop Commander Updates and Versioning

• Desktop Commander updates automatically on restart if installed through npx or symmetry.
• Users can check their version by asking "What is my Desktop Commander version?" in the chat.
• Future updates may include a link to the release log.

Conclusion and Call to Action

• Michael expresses his gratitude for Desktop Commander and its impact on his project.
• He is now actively seeking investment to bring AirPro to market.
• He encourages HVAC contractors, general contractors, custom integrators, homeowners, and investors to get in touch.
• The interviewer thanks Michael for the discussion and encourages viewers to leave feedback, questions, and comments.