871: Coding Agents Cursor + Windsurf Tips

Episode Summary

Executive Summary: The episode explains how AI coding agents in tools like Cursor, Windsurf, and Replit Agent go beyond one-shot prompts by iteratively scanning code, editing multiple files, running terminal commands, and checking their own work. The hosts share practical tactics—rules files, screenshots, explicit technical instructions, and refactoring requests—to get higher-quality results, while noting the need for human oversight and domain knowledge.

Main Topics: What AI coding agents are and how they differ from chat: The hosts define agents as systems that can perform multi-step tasks: inspect a codebase, modify files, run commands, and use feedback loops to correct errors. This is contrasted with one-shot chat prompts that only produce a single response. Cursor workflow: chat, composer, and agent mode: Cursor’s newer agent mode is presented as the most powerful option because it can search the codebase, edit multiple files, and run validation. Composer is described as a staging area for multi-file, multi-step changes with previews and accept/reject control. Practical prompting strategies for better outputs: The hosts emphasize giving precise, technical instructions, including APIs to use, files to touch, and expected implementation details. Vague prompts lead to weaker code; specific prompts produce better, safer results. Rules files and project-specific guidance: Cursor rules files (and similar instructions files in other tools) help encode project conventions such as ESM, TypeScript, fetch over Axios, and framework-specific preferences. These are especially useful for newer or less common stacks. Using screenshots and docs as context: The episode highlights that screenshots of bugs or documentation can dramatically improve the agent’s understanding, including OCR-like extraction from PDFs and visual debugging of layout problems. Agents as a productivity multiplier for repetitive or messy tasks: The hosts repeatedly note that agents are best used for scaffolding components, refactors, experimentation, and bulk edits that are tedious to do manually. They also caution that some debugging still requires human expertise. Comparison across tools and feature parity: Cursor, Windsurf (Cascade), Replit Agent, and increasingly Copilot are discussed as similar tools in a fast-moving space. Listener choice may depend on current implementation quality and how often each product updates.

Key Arguments: AI coding agents are more effective than one-shot chat because they can loop through code search, edits, terminal checks, and error correction until a satisfactory result is reached. Cursor’s agent mode is currently the most useful feature because it can make multi-file changes and validate them, making it especially strong for larger or repetitive tasks. Composer is valuable because it stages changes, shows diffs, and lets developers accept or reject edits without losing control of the codebase. Precise technical guidance matters: telling the agent which APIs, conventions, or frameworks to use produces code closer to what a senior developer would ship. Project-specific rules files significantly improve output quality by reducing the need to repeat conventions like ESM, TypeScript, or framework choices. Screenshots and documentation images can be surprisingly effective inputs for fixing UI bugs or interpreting structured data. Despite strong results, agents still fail on certain debugging tasks, especially environment- or browser-specific issues, so human review remains necessary. These tools are especially useful for low-stakes, high-reward work such as scaffolding dashboards, refactoring, or exploratory experimentation.

Data Points: Number of files changed in a typical agent task: Multiple files - Agent mode can edit more than one file during a single task rather than only the active tab. Iterations on a Venn-diagram demo: About 15 back-and-forth changes - Wes described using Composer/agent mode to build and refine a visual set-operations demo. Iterations before first successful Venn diagram output: 6 or 7 revisions - The initial version needed several refinements before it matched the desired visualization. Time to build a set-operations visualization: About 20 minutes - A demo showing union, intersection, symmetric difference, and other set operations was built quickly with AI assistance. Time spent debugging one Safari issue: Longer than writing the code manually - The agent struggled with a touch animation issue in Safari, making human debugging slower overall. Number of small tasks completed in a weekend before injury interruption: Several / many - Scott mentioned multiple home and dev tasks that were interrupted by his finger injury. Agent feedback loops: 3 or 4 times typically - The agent often cycles through search, edits, and lint/error checking several times to reach a final result. Rule of thumb for output quality with specific prompts: 9 times out of 10 better - Wes argued that giving concrete technical instructions substantially improves results.

Pivotal Quotes: "What an agent is, is something that can do multiple tasks. Meaning, it can generate code, NPM install things, run things on terminals." — Wes Bos: Definition of AI coding agents early in the episode. "I think nine times out of ten, the results are so much better in terms of like giving you exactly what you want." — Wes Bos: Explanation of why precise, technical prompting improves agent output. "I had built, I don't know, three or four things in the last week or so with it, and I'm amazed at how quick and productive I can be." — Wes Bos: Wes describing his recent experience using Cursor agent mode.

Implications: AI agents can materially speed up coding, refactoring, and prototyping, but only when paired with clear technical direction, project rules, and careful review. Expect rapid tool churn as Cursor, Windsurf, and Copilot race toward better autonomous workflows.