Claude Spiel: A Week with Claude Pro

by Rian

It has been roughly a week, first week, on the paid subscription LLM service. I did hear good things about Claude, especially its versatility when it comes to answering general questions and coding. For the life of me, I was not satisfied with ChatGPT and Gemini when it came to getting helps on coding, — so much so vibe coding on those platforms would be risky — whereas Claude printed results like a bureaucrat.

Right off the bat, if Claude is your go-to choice as an enthusiast or a hobbyist, I fear “Pro” tier may not be enough. It’s the tier I am on and paying for. Claude doesn’t provide useful monitoring tools or gauges to keep track of how much token I can use per week. But if my anecdotal experience is to be believed, after working with 2k – 2.5k lines long project for around 5 days, I hit the weekly limit. That is, if I am being generous to the subscription model Claude is offering.

Claude also has a daily limit and a session limit. If my understanding on the topic is correct, daily limit is yet another meter running consecutively with weekly, but the session limit is the window of how long can a user work with Claude uninterrupted. Anthropic seems to suggest it is around 5 hours. I beg to differ. After handful number of prompts, (my Claude usage was mainly on Claude Code) I would hit the session limit. After 5 hours when it refreshes, I could try again only to hit the daily limit this time.

My workflow and many of my DIY scripts had to be adjusted to optimize with the way Claude works. One of which was to make all projects modular as possible and reusable as possible, just to avoid rewriting the same code again accidentally. Remember, you are ultimately paying for “usage”, i.e. tokens, which expires every week. Restructuring my codes was sitting on the back burner for some time, so I took this chance to do exactly that. I believe I had used my second day with Claude doing only that for one project.

I suppose the questions many readers would want the answer for is — whether Claude is suitable to learn how to code. My answer is no. Claude Code currently offers “planning mode”, which is to say Claude creates a blueprint (or a sales pitch) of what it understood from the human user, then the actual work is done after the planning stage. While I appreciate having more finer controls, ultimately I am the one who understands what I want, how I want, and why I want it. Not only the user still needs to write the algorithm, — blueprint is just paraphrasing what it is entered — there is no guarantee Claude understands the context of the well-written documents of other APIs.

AI hallucination is still is a big issue, much less so with Claude, only that it has bigger tendencies to print false negatives, instead of usual false positives. Allow me to elaborate because I don’t think this was a one-off freak accident. I was looking for a manga artist who had gone viral. I didn’t recall much of the artist’s work — some themes, some art styles, when the one shot went viral, and possibly her name or an alias. Gemini was adamant it found the manga I was looking for. It even gave me ISBN, the title of the one shot, the author’s name and other aliases, and the quick run down of the work. Claude was adamant it was not real, and it is a product of Gemini’s hallucination. Turns out, neither of them were correct. Gemini’s ISBN, manga title, the author’s full name were all hallucinations. But I had remembered the part of the name right, right summary, and the period when it became popular. Google Search gave me the results, not the LLMs.

I will continue to use Claude for my personal projects. I don’t see how it will work for professional lines, however. I suppose it requires some balancing act between the expenditure toward an LLM and the revenue from the final work. I can comfortably say Claude Pro is as Pro as iPhone Pro is to professional photographers. It is a new set of tools to have and it is versatile, but it won’t replace old-fashioned tools-tools. Personally, I like how both LLMs often suggest APIs, packages, open-source projects I didn’t think to look up. A lot of the projects that have been sitting in the corner can now finally come to fruitions with Claude. That is, if I am willing to do my part in this dance.