Inefficient Tool Invocation Causing Significant Development Productivity Loss
Bug Description
Subject: Feedback on Claude Code – A Loyal User’s Perspective
Dear Claude Team,
I am writing this letter as a loyal admirer of the Claude model. I have always been deeply impressed by its power and capabilities. When Claude Code was launched, I immediately subscribed to the Max plan to use it, expecting it to become my primary tool for coding. However, my experience over the past two months has been disappointing.
For the last two months, I have relied on Claude Code for my daily software development tasks. Yet, I have often felt an inexplicable sense of inefficiency, as though my productivity has dropped compared to before. I initially dismissed this as an illusion, believing that with the Opus model – one of the most powerful models available – such a decline was impossible. But after two months, I can no longer ignore the reality: my projects have faced significant delays, and the frustration has become overwhelming.
Out of curiosity, I recently tried using Cursor, a tool that costs just $20 per month and is powered by Claude Sonnet 4 Thinking. What I could barely achieve in an entire evening using Claude Code, I was able to accomplish in just 30 minutes with Cursor. This stark contrast left me both shocked and dissatisfied.
I want to emphasize that Claude is the most powerful model in the world, especially for programming tasks. Its reasoning and coding abilities are unmatched. But Claude Code, as a product, feels underwhelming and frustrating at times. Issues that should be resolved quickly often linger for a month or two without improvement. Many of these problems seem neither technically complex nor difficult to fix – they simply require more attention to detail.
For example, when invoking tools, Claude Code often “overthinks” trivial tasks, wasting time unnecessarily. Sometimes, it even fails to write code properly on the first attempt, leading to redundant interactions that drag down efficiency over time. These seemingly small inefficiencies compound and prevent the model from fully showcasing its strengths.
Claude Code is a groundbreaking product, and I firmly believe it represents the future of programming. But instead of focusing solely on flashy new features, I urge you to refine the core experience. Get the basics right – smooth, fast, and reliable tool use – before anything else.
Despite the frustrations, I have continued to use Claude Code because I believe in its potential. Yet I am now exhausted. I am shocked by what feels like a lack of user-centric thinking and attention to the essential details that truly matter.
Claude Code can be great – it should be great. I hope this feedback can serve as a constructive call for improvement.
Sincerely,
yiwen lemo
Environment Info
- Platform: darwin
- Terminal: cursor
- Version: 1.0.61
- Feedback ID: 17db96a8-3cb7-4e5e-99f9-a8b5501f9360
Errors
[{"error":"Error: Command failed: security find-generic-password -a $USER -w -s \"Claude Code\"\nsecurity: SecKeychainSearchCopyNext: The specified item could not be found in the keychain.\n\n at genericNodeError (node:internal/errors:983:15)\n at wrappedFn (node:internal/errors:537:14)\n at checkExecSyncError (node:child_process:882:11)\n at execSync (node:child_process:954:15)\n at i7 (file:///Users/haoyiwen/.nvm/versions/node/v22.16.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:865:5106)\n at file:///Users/haoyiwen/.nvm/versions/node/v22.16.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:777:49649\n at Q (file:///Users/haoyiwen/.nvm/versions/node/v22.16.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:674:14420)\n at kJ (file:///Users/haoyiwen/.nvm/versions/node/v22.16.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:777:47087)\n at XH (file:///Users/haoyiwen/.nvm/versions/node/v22.16.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:777:45954)\n at Y9 (file:///Users/haoyiwen/.nvm/versions/node/v22.16.0/lib/node_modules/@anthropic-ai/claude-code/cli.js:777:52275)","timestamp":"2025-07-26T09:12:58.905Z"}]
17 Comments
@Hao-yiwen I am in a similar situation. Literally bought the MAX subscription the first day that Claude code dropped. I was very happy with the results, and then around 10 days ago I started getting subpar results.
I haven't changed anything in my stack or introduced new tools, so it was quite weird. At first, I thought maybe it was my mistake and I'm doing sloppy prompts. But basically, this started when Anthropic began having problems with their models being overloaded.
The reasoning capabilities of the model degraded significantly. I am running everything in a test environment as I'm doing heavy development. Yesterday, Claude code pruned the system from everything Docker related because it believed it had a cache issue with Node.
Today, after being told to clear a port, it killed the node process it was running on. After that, it killed all next,js processes instead of a single PID on a port.
I can not comment on the code quality that it generates, but as I'm seeing it has no problem running very destructive actions in order to get something done. I really hope nobody is running Claude code anywhere near production unsupervised because this can be a hot mess. This can totally be related to the issues that the guy from SassStr had with Replit. https://x.com/jasonlk/status/1946069562723897802
I am not an ML expert, but I believe they are doing quantization to get the cost down. They introduce the most powerful model first with its full precision. After that, they lower the precision so they can save on computing power and electricity. I totally understand this since everybody in the industry is doing it. But this time it took Claude to a level where it just isn't the best in class anymore. Gemini is closing fast, especially with its huge context window and speed.
I started manually coding and running most of the processes again, as I cannot rely on Claude being correct and efficient anymore. I am still using it and will continue using it for some of the smaller, less dangerous tasks. Below are some of the screenshots that display the destructive actions.
<img width="1234" height="465" alt="Image" src="https://github.com/user-attachments/assets/e9b79ac0-78c4-4891-95ec-7fd3747c5164" />
<img width="1288" height="458" alt="Image" src="https://github.com/user-attachments/assets/ed8823b6-b22f-420a-91b2-2f06b2be87cf" />
I'm tagging the core of the Claude code team for visibility @bcherny @rboyce-ant @ashwin-ant @8enmann @sid374 ........
I’m very disappointed with them. I’ve reported many issues for almost two months, but they still haven’t been resolved or improved.
<img width="647" height="66" alt="Image" src="https://github.com/user-attachments/assets/b132660a-3def-4b20-99bf-2eb02c224a02" />This issue shouldn’t be happening. They should optimize the prompts and ensure successful execution on the first attempt. However, right now, the first attempt often fails. I believe this situation leads to deterioration of subsequent results. Many times, I need the model to write about 10k of content into a file, but it always fails on the first try. I think this is causing the accuracy of subsequent sessions to decline.
@Hao-yiwen 10k lines of content is a lot of work for the model. To be honest, I try to keep the lines of code per file around 200-300 lines max.
I am excessively using the planning mode today because otherwise the results are just not there.... If the pricing wasn't capped today would've been one very expensive day.
Cursor and Gemini CLI clearly don’t have this issue.
I will be trying Gemini CLI today, specifically for front-end work. If I do indeed get better/faster results, I will have to rethink things for the future.
<img width="337" height="53" alt="Image" src="https://github.com/user-attachments/assets/8a4f7c3a-90f5-4a87-8440-6cfe1ecfbe5f" />
The write update and replace tools often encounter errors, especially the replace tool.
Yes, it has been like this since the very beginning, even though I'm not seeing this as a big issue. The sub-agents are fixing it immediately as it appears. The only problem that I'm experiencing with this is wasted time. Which can definitely accumulate since we are using Claude non-stop.
<img width="508" height="38" alt="Image" src="https://github.com/user-attachments/assets/c1a0619f-74fc-4633-b394-38fdee2d2a60" />
For me, this is a huge issue because Claude just told me I'm going to destroy your system (nuclear), and it continued doing so because it believe to be caching issue 🫤
Yes, using Claude Code is a huge waste of time. For the same tasks, Cursor is several times faster.
I once thought the issue was due to slow model endpoints, but after testing several identical tasks, I found that the endpoint speed is actually the same.
Then I realized it’s a problem with the system design of Claude Code.
For me, time cost is a huge issue. My project was supposed to go live at the end of June, but it’s almost August now, and I still don’t know when it will be finished. I feel an overwhelming sense of helplessness, and now I have no choice but to switch to other tools.
I'm in the same boat as I can see you're solo building a project and doing that as well. Now imagine if this was a company that's paying developers by the hour + tokens. The costs will be sky-high. I guess we're not losing our jobs after all 😃
<img width="424" height="250" alt="Image" src="https://github.com/user-attachments/assets/309de28d-ed7b-4456-8c88-77207bd43751" />
This is such a simple task—just writing a basic management system into my test.html—yet Claude Code actually needs three attempts to succeed. The worst part is that each attempt consumes 3.2k tokens, meaning this single task requires around 9k tokens to complete. I’ve really had enough of this.
@Hao-yiwen
<img width="710" height="72" alt="Image" src="https://github.com/user-attachments/assets/a713e8d3-9637-4022-a581-4f875727daa2" />
Set the model to Opus 4 exclusively. It is fast and correct and loads the model from May. I believe this will solve your problem.
The above test example is from Opus.
This issue has been inactive for 30 days. If the issue is still occurring, please comment to let us know. Otherwise, this issue will be automatically closed in 30 days for housekeeping purposes.
This issue has been automatically closed due to 60 days of inactivity. If you're still experiencing this issue, please open a new issue with updated information.
This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.