Back
Post
First experience with Claude Code — is this normal?
Hey, guys!
Today I’ve bought an LLM subscription for the first time (Claude Pro plan) and wanted to give it a go on a real project task.
I’ve been watching Claude Code videos for quite a while, read the docs (regarding Claude Code in desktop app though), and, well, went for it…
The results are not fun.
The task was pretty simple, on paper, at least, imo:
Fix/add new fields to the Baserow DB since it doesn’t support lookup and formula fields from Airtable.
Enabled a few plugins, used the plan mode in the beginning and during this task, so I thought I’m good to go.
But get this:
• This is my first time using Claude with an exception of a few small chats — after I bought the subscription, the first chat asking about Claude’s capabilities cost me 2–4% of the 5 hour usage.
• Everything discussed here happened in the desktop app, not CLI or web version.
• The model used is Sonnet 4.6, not Opus 4.6.
• I started this task (prompt below) about 7 hours ago (literally) — actually, woah, I didn’t expect this. I set down with Claude Code all this time and was, mostly, clicking “yes” when it asked for permissions. While it maybe isn’t 7 hours, but it’s definitely 5.
• The task still isn’t finished. Technically it should be, the new fields should be there (I didn’t check), but there are a few errors that need to be fixed.
• Most important part is that I managed to deplete whopping 27% of the weekly usage just for this task — and I consumed first 5 hour usage and now the second one did hit the limit.
Honestly speaking, I still want to be a believer, I do think I didn’t do more “optimization” and maybe I did a few prompting mistakes or something like that… but I also think that I was pretty efficient with the approach, since it was a single task with a focused goal, not some mishap of different “build be X and Y”.
My question to all of you using LLMs and Claude Code specifically: is this fucking normal?
Today I’ve bought an LLM subscription for the first time (Claude Pro plan) and wanted to give it a go on a real project task.
I’ve been watching Claude Code videos for quite a while, read the docs (regarding Claude Code in desktop app though), and, well, went for it…
The results are not fun.
The task was pretty simple, on paper, at least, imo:
Fix/add new fields to the Baserow DB since it doesn’t support lookup and formula fields from Airtable.
Enabled a few plugins, used the plan mode in the beginning and during this task, so I thought I’m good to go.
But get this:
• This is my first time using Claude with an exception of a few small chats — after I bought the subscription, the first chat asking about Claude’s capabilities cost me 2–4% of the 5 hour usage.
• Everything discussed here happened in the desktop app, not CLI or web version.
• The model used is Sonnet 4.6, not Opus 4.6.
• I started this task (prompt below) about 7 hours ago (literally) — actually, woah, I didn’t expect this. I set down with Claude Code all this time and was, mostly, clicking “yes” when it asked for permissions. While it maybe isn’t 7 hours, but it’s definitely 5.
• The task still isn’t finished. Technically it should be, the new fields should be there (I didn’t check), but there are a few errors that need to be fixed.
• Most important part is that I managed to deplete whopping 27% of the weekly usage just for this task — and I consumed first 5 hour usage and now the second one did hit the limit.
Honestly speaking, I still want to be a believer, I do think I didn’t do more “optimization” and maybe I did a few prompting mistakes or something like that… but I also think that I was pretty efficient with the approach, since it was a single task with a focused goal, not some mishap of different “build be X and Y”.
My question to all of you using LLMs and Claude Code specifically: is this fucking normal?
Here’s the initial prompt:
Hey, Claude. I have a self-hosted Baserow and an Airtable base, but I'm on a free plan. I can provide whatever you need: login info (create a new account or give my existing one), SSH to my VPS (if you think it is better this way), MCP (Baserow has an MCP, but I have no idea how it works and what it is), webhook, etc for Baserow (whatever you think is necessary) and access public link (visible data) of my Airtable base. What I want you to do is the following: I imported Airtable base into Baserow, but there was no support for Airtable’s formula fields in Baserow, so I want you to analyze Airtable base via the public link (or maybe there's another more efficient way on a free plan), compare the columns/fields from Airtable to my Baserow (data should be identical but still you can check it it), and create new fields in Baserow with necessary formulas based on how they work in Airtable. Also, there’s a database in Baserow that lists all incompatibilities that happened during the import, so you can check it to and fix the entire imported data so it matches AIrtable’s one to one or as close as possible. That being said, I do not want you to modify any data in Baserow like companies, tariffs, etc — I just want you to edit/add new fields, so new data appears (like added formula).
👋 Join WIP to participate
if you wen't around in circles, probably normal -- but definitely weird at first sight. Do you have extended thinking mode enabled? If so, that will consume way more credits
Also, the less clear your prompts are and less context it has on your architecture, it will spin in circles consuming credits in the background.
Make sure you use a CLAUDE.md file with as many details as you can about your project, and the way you want it to respond.