Here’s the crazy paradigm shift: we can use LLMs on themselves to improve applications, and we can be smart about it.

Evaluate its own output as a feedback loop

For example, we can compare the two different responses using a “Copy to LLM judge” button instead of having to try to wrap our brain slowly around which version might be better.

For example, we can add some level of reproducibility and replay-ability along the way to each step of some process, then simply tell a LLM to look for improvements.

For an even further developer experience improvement, we can add knobs and switches, or helpers like some “copy to clipboard” button so that we can one-click copy and then paste into a LLM of choice.

Can even copy the response of one LLM’s findings to another.

I like: Gemini 3.5 Flash, Opus 4.8 Max, Grok expert mode.

Vibecode a fix, and tell it to add to context.md

Often I run into (CSS) stylistic issues or bugs. When I see these, I tell the LLM to fix it, then update a centralized “design style” context file or doc.

Tell LLM to help debug by temporarily add detailed frontend/backend logs

Sometimes we can’t be sure of what is going on within a difficult process, this is when we can temporarily add rich debug logging, then paste it back to the LLM to give it more context to debug itself.

Update design docs / context rules continuously

Persist and update design docs when adding or modifying major feature changes. This is akin to human-written technical design documents in the past, but now its the LLMs reading and updating them.

Code comments aren’t good enough IMO, since they can be scattered across different files, so it takes longer to read from them or update them - docs are more centralized.

Heck, you can even surface docs in a Markdown renderer on the frontend to read easier (instad of in VS Code preview).

Vibecode a feature, tell the LLM to test it with an ad-hoc script or cURL request

I think a lot of SOTA coding agents already do this nowadays without additional prompting, but its quite ergonomic to tell the LLM to just test a feature it just built, itself. It can think of more test cases than a human anyways.