Better SINGLE-TURN Aider inspired features #14181
david-derooy
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
-
Background:
In November 2024 I started experimenting with 'ai-code-gen' (no vibe code yet). I played a lot with Claude Sonnet 3.5 and made an honest attempt at building something. I noticed a recurring trend: the longer the conversation went, the worse the LLM got. The problem of context rot was very noticeable in those early days and provided valuable lessons. Later in 2025 I came across Aider and started using it with Gemini 2.5 pro when it came out in March. My prompting strategy focused on a single well defined feature prompt, with minor follow up tuning prompts. If the output took a wrong direction I would clear everything and roll again with more details and clarifications in the starting prompt. Since then I've never looked back and had the pleasure of building incredible things using that setup.
Present day I've been experimenting with Gemini CLI and once again realize an old lesson from the early days: treating an LLM like msn messenger does not yield quality output.
Problem:
LLMs are bad at multi-turn long conversational workflows. This research paper (and my background experiences) drive home that point. "Our experiments confirm that all the top open- and closed-weight LLMs
we test exhibit significantly lower performance in multi-turn conversations than single-turn, with
an average drop of 39% across six generation tasks".
Gemini CLI has a UX that primarily promotes multi-turn prompt flows and can be improved by adopting single-turn features inspired by Aider. Basically take what makes Aider great and apply it to Gemini CLI.
Proposed ideas:
Summary:
Gemini CLI is a great tool and has a lot going for it. Quality output can further be improved by expanding support for various prompt engineering styles, specifically single-turn prompts which is research backed to achieve better results.
Beta Was this translation helpful? Give feedback.
All reactions