-
@ Evan
2025-01-19 22:36:34
# Writing with Claude Watching My Every Move
As the result of a series of poor decisions in my life, I possess $500 of nonrefundable Anthropic API credits. This is an astounding amount of money, especially in the context of this extremely low cost API. I could use Claude Opus ($15/MTok input, $75/MTok output) for regular chatting and that would probably eat through these credits pretty quick. But I don't exactly want to WASTE this money and Opus is way more expensive than Sonnet 3.5 ($3/MTok input, $15/MTok output) with [worse benchmark performance](https://www.anthropic.com/news/claude-3-5-sonnet). I'm not sure if there is any use case where it makes sense to use Opus so using it would make me feel like I am lighting money on fire. So I am relegated to figuring out how I can inject Sonnet 3.5 into as much as my life as possible so I can use this ridiculous amount of credits and get some sort of benefit from it.
Thus, my [realtime writing feedback tool](https://github.com/edverma/llm-writing-feedback) is born. It's a simple program that sends the contents of any specified file to Claude and asks for feedback on each save. Originally, every version of the file was sent to Claude. This allowed for iterative feedback as the file was edited, but mostly importantly it used lots of input tokens to help burn through these credits. This actually ended up using too many tokens, as I was spending $2-5 per writing session. That was a little bit too much and made me realize that maybe it will be easier than I originally anticipated to burn all of these credits. I changed the program to only send the current revision and the most recent revision. This results in the same quality of iterative feedback while minimizing the input token count.
## What is it Like to Use?
[Here](https://youtube.com/shorts/oKXzZ7lG56g?feature=share) is a quick demo. I have my text editor open on one side of the screen and my terminal running the program on the other side. Every time I save my work in the text editor the program displays new feedback on the revision. The feedback is brief, usually between 3 and 5 sentences. When I first started using it I mostly ignored it, but now I check it after almost every save. I also have noticed that I have changed my patterns of saving my work to match up with the moments when I want to see the feedback.
Compared to other LLMs, Claude is particularly agreeable and nice. This makes the feedback tool feel like a supportive friend cheering me on and encouraging me as I write. Other LLMs are more monotone and matter-of-fact, and for this use-case I appreciate the friendliness and expressiveness of Claude. On the other hand, Claude has stronger opinions about ethics and morality than other LLMs. When I was writing an article exploring some ideas around hacking, Claude constantly was warning me about the legal and ethical implications of black-hat and grey-hat hacking. This can be helpful or tiresome, depending on your disposition.
In addition to supporting me, the tool often helps guide and unblock me when I'm stuck. It doesn't offer concrete edits or suggestions, but it gently suggests potential continuations or new ideas to explore. These can be ignored as much as I like, but when I'm stuck they typically are very helpful. Even when these suggestions are ignored, it's nice to have some feedback and alternative ideas for consideration. I haven't found that this harms anything about the writing process, but it makes it feel somewhat collaborative. I'm in the drivers seat, but Claude suggests places to go when I get lost. And otherwise when I'm cruising down the road, I can always check with Claude to see how he thinks things are going.
## Next Steps
If you write and would like to use this, let me know and I will help you get set up. It's a new process to get used to but so far for me it has made writing more productive and more fun. I want to build a MacOS app around this so that it can be more accessible to use for nontechnical people. I'll let people use it for free if they provide their own API key, and I'll add a low-cost subscription for anyone who does not know or does not want to set up and manage an API key.