%22%20%2F%3E%0A%20%20%3C%2Fsvg%3E)
Planning the AI-Assisted Coding Session
Thariq Shihipar, a creator known for ambitious AI projects, has detailed a highly efficient workflow for tackling complex coding tasks with Claude, an AI assistant. His method, which he claims can be executed in under 40 minutes, emphasizes meticulous planning and structured prompting to achieve significant progress on ambitious projects. The core idea is to treat the AI not as a magic bullet, but as a highly capable, albeit junior, developer that requires clear direction and context.
Shihipar’s approach begins with defining the project’s scope and desired outcome. Before even interacting with Claude, he spends time sketching out the architecture, identifying key components, and listing out specific functionalities. This pre-computation phase is crucial. It prevents aimless prompting and ensures that the AI’s output is directed towards a concrete goal. He likens this initial planning to a human developer creating a detailed specification document before writing a single line of code. The goal is to provide the AI with a blueprint, not just a vague request.
The planning stage involves breaking down the ambitious project into smaller, manageable sub-tasks. For each sub-task, Shihipar identifies the required inputs, the expected outputs, and any specific constraints or dependencies. This granular approach ensures that the AI can focus on one piece at a time, reducing the cognitive load on both the human and the AI. He stresses the importance of anticipating potential issues and thinking about how different code modules will interact. This foresight allows for more targeted prompts, asking Claude to consider edge cases or to generate code with specific error-handling mechanisms.
Structuring Prompts for Maximum AI Output
Once the plan is solidified, Shihipar moves to the prompting phase. He advocates for a “role-playing” prompt, where Claude is instructed to act as a senior developer tasked with a specific problem. This sets the tone and encourages the AI to adopt a more analytical and problem-solving mindset. The prompt itself is structured to include all the necessary context gathered during the planning phase: the project’s overall goal, the specific sub-task at hand, input requirements, desired output format, and any relevant constraints.
A key technique Shihipar employs is the use of “example-driven” prompting. Instead of just describing what he wants, he provides concrete examples of input and expected output. This is particularly effective for tasks involving data transformation, API integrations, or specific code patterns. By showing Claude exactly what a correct solution looks like, the AI can more accurately replicate the desired behavior. This is akin to providing a junior developer with a few solved examples to learn from before assigning them a new problem.
He also emphasizes the iterative nature of AI interaction. It’s rare to get perfect code on the first try. Shihipar expects to refine his prompts based on Claude’s initial responses. If the code is not quite right, he doesn’t just say “fix it.” Instead, he provides specific feedback, pointing out the exact errors or deviations from the desired outcome. This feedback loop is critical for guiding the AI towards the correct solution. He might ask Claude to refactor a section, improve efficiency, add comments, or ensure adherence to specific coding standards. This detailed, constructive criticism helps Claude learn and adapt within the session.
Running and Verifying the Generated Code
The final stage involves running and verifying the code generated by Claude. Shihipar doesn’t blindly trust the AI’s output. He treats the generated code as a draft that needs thorough testing. This involves setting up a local development environment, compiling the code, and running it with various test cases, including edge cases identified during the planning phase. If the code fails, he uses the error messages to further refine his prompts or to manually correct the code himself.
This verification process is not just about catching bugs; it's also about ensuring the code meets the project’s architectural requirements and performance expectations. He might ask Claude to generate unit tests alongside the main code, which can significantly speed up the verification process. If the generated code is complex, he might ask Claude to explain specific sections or to provide documentation, further aiding his understanding and review.
Shihipar’s 40-minute framework is not about completing an entire complex application. Instead, it’s about achieving a substantial, demonstrable step forward on an ambitious project. It’s about planning, prompting, and verifying efficiently to make significant progress. The speed comes from the structured planning and the precise, context-rich prompts that guide Claude’s capabilities effectively. This method transforms AI from a novelty into a powerful tool for accelerating complex software development, enabling creators and developers to build more ambitious projects faster.
The Unanswered Question: Scalability and Long-Term Maintenance
While Shihipar’s 40-minute workflow offers a compelling method for rapid AI-assisted development of specific features or modules, a significant question remains: how does this approach scale for larger, more complex projects that require long-term maintenance and evolution? The current method seems optimized for rapid prototyping and feature iteration. However, managing a large codebase generated with AI, ensuring consistency across modules developed at different times or with different prompt iterations, and maintaining that code over months or years presents a new set of challenges. Will AI-generated code be as maintainable and adaptable as human-written code in the long run? What are the best practices for version control, documentation, and team collaboration when a significant portion of the codebase originates from an AI assistant? These are critical questions for anyone looking to integrate AI-assisted development into their core product lifecycle.