Ask HN: Which LLM are you using to evaluate your ideas?

I am mostly using Gemini, but that's a decision I made 2-3 months ago. I try my best not to hint at what answer I am hoping for/would benefit me, and I play devil's advocate as much as possible.

For now, I'd say Claude. I tried to work with other LLMs such as Gemini, Codex, GPT and Grok(high versions), but still Claude (4.8) gave me the best outcome.

I actually haven’t gotten to evaluating ideas yet I’ve been too busy asking each llm why they thought they would be best for the task. Just kidding

I use ChatGPT and Gemini and Claude in various ways in my workflow

[flagged]

I landed with Claude as my daily driver LLM. As it is still limited in some capacities I do also work with ChatGPT and Gemini where they give better results in specific circumstances.

I’ve started using different models for different stages.

GPT mostly for product direction and positioning discussions. Codex for implementation-heavy changes and large refactors. v0 more for landing page structure / UI iteration.

I still find product judgment and SEO decisions require a lot of manual thinking though.

Manus was great, used it for months for exploring ideas, making reports and such. I dropped it when I heard that Meta was buying it.

Then I used Gemini and it was great as well, but very limited when it comes to coding (in Antigravity) because instead of giving me credits for coding in the subscription they give you GB on Google Drive. So you burn through the coding very quickly.

Github Copilot is the worst. It usually gets lost while editing some file. And it is incredibly slow.

Then I swithced to Codex and ChatGPT. Great combo for everything. Good at coding, good at exploring, generous with limits (much more so the first month).

I haven't used Claude Code directly, but Claude Opus 4.x was probably the best model for coding used via Copilot and Antigravity.

I use Claude for initial idea validation.It is decent. For iteration and refinement, I route simpler clarifications through local models (Qwen via Ollama)

[dead]

I tipically use Sonnet 4.6 to have new ideas on how to develop my existing ideas, and iterate over them, I find it perfect for this.

I use Claude to brainstorm, then I ask Chatgpt and Gemini to 'stress test' the idea. This process helps me come up with better ideas and then I use Chatgpt or gemini to help me create the right prompts to feedback to Claude and iterate the idea further.

I use Claude for initial validation. Then move to deepseek my stack is langchain, complexity sucks, but once you setup the skills, claude.md, AI is ok dealing with it, langchain makes it easy to switch LLMs. some minor issues here and there, e.g. different models have different messaging format.

so far so good.

I don't use any LLM to evaluate idea or any kind of brainstorming, I go by google search and detailed reading - internet as well as real talk with people. the llm is going to pour out only whatever it has been trained on. talk to more and more people and you will be able to evaluate your idea/brainstorm better.

[flagged]