923: Getting the Most Out of AI Coding

Episode Summary

Executive Summary: Wes and Scott share practical ways to get better results from AI coding tools: scaffold projects manually, be explicit about versions and constraints, use rules files and structured prompts, break work into small tasks, supply docs/logs, and restart chats when context degrades. The episode emphasizes that AI works best as a focused assistant, not a magic one-shot app builder.

Main Topics: Manual scaffolding before prompting AI: The hosts argue it is faster and more reliable to set up the project skeleton, dependencies, and initial example code yourself before asking AI to generate features. Prompt clarity and version specificity: They stress using explicit instructions like 'do not use X, use Y' and naming exact frameworks, syntax, and versions to reduce AI ambiguity and outdated choices. Rules files, markdown, and structured prompt inputs: The discussion covers cursor rules, Copilot rules, LLMs.txt, XML-like tags, and MDC files as ways to persist style and behavior instructions across chats and projects. Task decomposition and incremental implementation: They recommend breaking large requests into clear, small, testable tasks so the model can execute and verify each step more effectively. Using documentation and logs as context: The hosts highlight feeding the model library docs, LLMs.txt files, MCP-based documentation retrieval, and runtime logs to improve code accuracy and debugging. Managing context limits and chat hygiene: They warn that long-running chats can loop, forget earlier instructions, or degrade in quality, so starting fresh chats for new features is often better. AI as an assistant, not an autonomous builder: Across the episode, they frame AI coding tools as junior collaborators that need guidance, constraints, and feedback rather than broad open-ended requests.

Key Arguments: Scaffolding the first 2-5% of a project yourself is usually faster than having AI guess and then correcting it repeatedly. Being extremely explicit about what to use and what not to use improves results more than clever phrasing. Rules files and structured formats reduce repeated prompting and help maintain consistency across sessions and projects. Breaking work into small, concise tasks leads to higher-quality output and makes review easier for the developer. Supplying logs, docs, and concrete code references gives the model the missing context it needs to diagnose issues and generate correct changes. Long chats eventually become less reliable; starting a new chat can reset bad assumptions and looping behavior. AI can generate standards documents and rules from an existing codebase, which can then be reused across projects.

Data Points: Episode length: Quick episode - The hosts describe this as a short, tip-focused episode rather than a full deep dive. Prompt breakdown: First 2-5% - Wes recommends scaffolding the first small portion of the app manually before involving AI. Troubleshooting loop: Six or seven times - Scott describes repeatedly prompting and correcting AI when it fails to fix an issue. Chat degradation threshold: Two, 300 times - Scott says very long back-and-forths can make a chat messy and less useful. Temperature in office: 25.9°C / 78.62°F - Wes mentions the room temperature while discussing air conditioning and joking about overheating during the recording. Context window growth: Much larger - The hosts note that larger context windows are helping, but degradation still happens in extended chats.

Pivotal Quotes: "scaffold it out yourself. At least install all of the dependencies that you want." — Wes Boss: Advice to manually set up the initial project rather than having AI generate the entire starting structure. "be as explicitly clear as you want and not trying to be cute or whatever" — Scott Tolinski: Scott explains that direct, unambiguous prompts work better than implied or clever wording. "There is so much value in what went wrong." — Wes Boss: Used while discussing logging, debugging, and the usefulness of feeding failures back into AI tools.

Implications: AI coding works best when treated like a constrained collaborator: give it structure, context, and small tasks. Teams that standardize rules, docs, and debugging feedback loops will get more reliable output and less frustration.