Feature idea: Automated prompt pushing.

[arthurwolf]

Browsers like Chrome (I think firefox might have support now too) can be started with a command line option that provides a websocket port that can be used to control the browser fully.

You can get a list of tabs open, select one of them (for example chatgpt.com or the gemini website), and fill forms you find on those pages.

This is very much like pupeteer, but instead of working in a background/abstract browser, it works in your real day to day browser, in which you're logged into chatgpt/anthropic/google AI etc.

We could use this to add a feature to prompt-tower where there's a "push" button at the bottom of the page, you generate a prompt, and instead of having to copy/paste it, you just click the "push" button, and it's automagically added to the textarea input in your AI page of choice.

It would need to be updated when the HTML page changes too much, but I think that would probably be manageable.

What do you think?

( in theory, we could also use the same technique to grab the answer from the LLM, and use that for tool use and back-and-forth debugging, but I looked at the TOS for OpenAI and Anthropic and that's unfortunately against their rules... but just pushing to the page, that doesn't seem to be a problem)

[backnotprop]

Yea, love this. I think there are only 2 possible ways:

1. Users have to launch chrome in debug mode (so I can communicate with the browser)

2. Companion browser extension

Browser extension I like for other reasons and future integrations/features.

[arthurwolf]

I didn't suggest a companion browser extension because I don't know how much an extension would be allowed to talk with a local process, and how much it'd be able to manipulate the OpenAI or Gemini tab, but if they are allowed to do both, then yes that's definitely better than asking users to start with a special flag.

On Fri, Mar 28, 2025 at 7:58 PM Michael Ramos @.***> wrote:

Yea, love this. I think there are only 2 possible ways:

Users have to launch chrome in debug mode (so I can communicate with the browser) 2.

Companion browser extension

Browser extension I like for other reasons and future integrations/features.

— Reply to this email directly, view it on GitHub https://github.com/backnotprop/prompt-tower/issues/19#issuecomment-2762190345, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAA2SFLLYUJSW3SQRXPZGB32WWLULAVCNFSM6AAAAABZ6LUG3CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDONRSGE4TAMZUGU . You are receiving this because you authored the thread.Message ID: @.***> [image: backnotprop]backnotprop left a comment (backnotprop/prompt-tower#19) https://github.com/backnotprop/prompt-tower/issues/19#issuecomment-2762190345

Yea, love this. I think there are only 2 possible ways:

Users have to launch chrome in debug mode (so I can communicate with the browser) 2.

Companion browser extension

Browser extension I like for other reasons and future integrations/features.

— Reply to this email directly, view it on GitHub https://github.com/backnotprop/prompt-tower/issues/19#issuecomment-2762190345, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAA2SFLLYUJSW3SQRXPZGB32WWLULAVCNFSM6AAAAABZ6LUG3CVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDONRSGE4TAMZUGU . You are receiving this because you authored the thread.Message ID: @.***>

--

勇気とユーモア

[backnotprop]

Figured this out.

Ideally I could send to cursor / code chat. but that is actually harder, private APIs.

[backnotprop]

automatically creating context and sending prompts - almost done for Cursor, and any web interface (gemini, claude, openai, aistudio) - full automation and execution of the prompt

[backnotprop]

next push will be to support windows, will need dedicated time on my pc.

feature/prompt-pushing

[backnotprop]

#38 rolled out to mac users in 1.4.0