Your agent is great at writing terrible code, and terrible at writing great code. I guarantee it. But with just a few guardrails (many of which we've already using for many years in industry), you can steer it from slop to splendor. We'll cover essential tips and tricks that I've been using for years in open source and this year as a frontend developer at Sentry, namely:
* Skills: teach your agent how to do the common tasks specific to your stack and repository.
* Static analysis: linting! type-checking! These will save you bugs, time, and tokens.
* Tests: know when to test, both in unit/integration/end-to-end and ad hoc at runtime with tools like MCP chrome-devtools.
* Yourself: sometimes, it's better to just do the task yourself. Keep those skills fresh!
We are at the "peak of inflated expectations" of a hype bubble, and it's easy to be overenthusiastic or overly pessimistic about LLM-based tools. But with the right hand-holding, they really can be magical for development.
This talk has been presented at React Summit US 2026, check out the latest edition of this React Conference.























