How do you use AI assistants?

1. web chat
2. standalone GUI
3. plugin to my code editor
4. official CLI
5. terminal (warp)
6. unofficial (self-written plugin/harness)
7. Other

Isn’t it “plugin to my code editor” though?

Edited, thanks!

Official CLI for code, web chat for brainstorm / plan

What’s wrong with official CLI for brainstorm/plan sessions?

For project-related tasks, I still use /plan, but for multiple or general topics, I like how the webchat memory system (ChatGPT Memory) works.

I am asking because I am building my own CLI harness and I wonder what people lack in other CLIs.

Could you please elaborate, what exactly in ChatGPT Memory you find attractive? How exactly is it better than say agy or claude memory?

IMO, ChatGPT web is much better at long-term memory (https://openai.com/index/chatgpt-memory-dreaming/). Most local coding CLIs still rely heavily on repo-local files like AGENTS.md; Codex does have a newer memory layer beyond that, but it still doesn’t feel as rich as ChatGPT’s long-term memory.

Hermes CLI is probably closer on the memory side, though it’s a general agent harness rather than a coding-focused one.

After billions of tokens spent on experimentation I am yet to see any model to produce a single 100 lines module that I would consider good, production ready code.

It is still very useful to me for debugging, verifying code and general exploring. In general I have shifted my workflow so it provides me with useful stuff instead of generating code with patterns that I have spend years scolding people not to use when doing code reviews.

I use pi coding agent because it’s easily customizable and minimal.

I have it running inside of a container (nspawn with regular arch linux inside it) that has almost everything my system has.
I clone all relevant repos into a folder that is bound to a container so it has access to all information locally.

My system prompt tells it is is not a coding agent but coding assistant and that it can modify only agent_playground folder. That it should provide answers in .md files with full reasoning behind pieces of code it provides. Never to have examples that are more than 50 lines long and that if asked it can generate full project with runnable code as a new project inside of agent_playground folder.

I mostly use it as chat interface that has access to the code. . Bigger answers and research answers in .md are there so I can easily open it in neovim and copy paste useful sections.

I’ve tried Web Chat, standalone GUI and even the code-editor plugin route.

But honestly, the official CLI is my weapon of choice. Whether it’s Claude Code, OpenCode or Codex, I somehow always end up back at the terminal.

At this point, I think I just have a toxic relationship with the command line. But hey, it works for me.