Sources

Developer anticipation for this AI coding agent is immense, fueled by the promise of offloading tedious and routine tasks. Yet, this initial excitement is colliding with stark real-world challenges, from eyebrow-raising pricing structures to troubling questions about performance reliability and practical workflow integration. This deep dive navigates the critical user pain points, unfulfilled expectations, and pressing questions swirling around Codex. We'll explore its current capabilities and its potential standing in the rapidly evolving landscape of AI-assisted development. ... The buzz surrounding OpenAI's Codex is undeniably loud, but it's being met with equally potent user anxieties, primarily centered on its demanding cost structure and the perceived value it delivers. The $200/month Pro subscription required for early access has prompted many to question if its current AI-assisted development practices justify such an expense, especially when measured against existing OpenAI ChatGPT Plus subscriptions or a growing field of more affordable alternatives. … Beyond the sticker shock, early performance reports for Codex present a mixed bag. Developers venturing into its capabilities have encountered instances of the AI generating mere placeholder code, experiencing excessive processing times, or finding it falls short on genuinely complex coding tasks. Such experiences cast a shadow of doubt on whether the `o4-mini` model, which powers the Codex CLI, truly offers superior code generation or contextual code reasoning compared to other established models when applied to practical tests, like integrating outputs into project tracking systems such as Jira. "We were told Codex would be a revolution, but for many small teams, the initial $200/month hurdle feels more like a roadblock, especially with token costs for CLI usage still undefined." Despite the current challenges, the developer community holds onto a potent vision for Codex, imagining it as a transformative "software engineering agent." ... … |Deep IDE integration (e.g., robust plugin)|The absence of mature native plugins makes browser-based coding impractical for many serious development projects; users seek solutions akin to having an AI GPT Router embedded, directing tasks efficiently within their preferred environment.| |Secure & private code handling|Persistent distrust regarding the transmission of code and prompts to OpenAI servers, despite assurances of local file operations. Concerns are heightened when considering project files potentially exposed through integrations with services like Google Drive.| … The post-research-preview pricing model remains a significant unknown, inducing considerable anxiety among potential users. Will Codex be an affordable add-on, a token-based consumption service, or will users find themselves needing costly OpenAI GPT Assistants API access for full functionality? Similar pressing questions arise regarding the CLI: how will "API token usage for the Codex CLI" impact existing quotas and the overall cost of services, especially when compared to other AI: Text Generation tools that might be used for quick docstring generation, potentially incurring extra charges? Predictable pricing is critical for workflows. … A glaring pain point for developers exploring Codex is its present deficiency in deep IDE integration. The notion of coding complex applications within a browser tab feels profoundly impractical for serious software engineering endeavors, a sentiment loudly echoed by users accustomed to the power and efficiency of local development environments. The demand for dedicated plugins (for generic editor standards, not necessarily a specific solution for every variant of plugin integrations) or similar direct hooks is immense. … Despite OpenAI's assurances regarding local execution for direct file operations, a persistent and significant undercurrent of worry surrounds data privacy and security when using Codex. Developers handling proprietary or highly sensitive codebases express understandable reluctance to "outsource their code" to cloud-based AI agents. This concern is magnified when considering the implications of managing secure credentials required for integrations with external services, such as financial data systems like Xero, which are integral to real business operations. The fundamental unease stems from the understanding that code snippets, detailed prompts, and high-level contextual information about the repository are inevitably transmitted to OpenAI servers for processing by the AI model. Lingering questions about how OpenAI might utilize this data—even if anonymized and not specifically for unrelated services like OpenAI Image Generation—for training future models or for generalized system learning persist. This ambiguity fuels anxiety, especially without more granular, easily accessible privacy policies specific to Codex and its secure sandbox environment.

While Codex offers exciting possibilities for streamlining tasks and accelerating development, it’s not without its limitations. Developers, especially those working on complex or cutting-edge projects, should take a closer look before fully integrating Codex into their toolset. A Game-Changer for Developers—With Caveats Codex promises to handle everything from simple functions to full codebases, automating repetitive tasks and freeing up time for more creative challenges. It’s like having an extra pair of hands that never get tired. But as with any powerful tool, it’s important to understand where it shines—and where it falls short. Here are five key considerations developers should keep in mind when using Codex: 1. Outdated Knowledge: No Internet, No Updates Codex is trained on a static dataset, meaning it doesn’t have access to real-time information or the latest updates in frameworks, libraries, or tools. If something was released after its training period, Codex won’t know about it. This is a significant limitation for developers working in fast-moving environments. While Codex performs well with established technologies, it may struggle with newer APIs or modern development stacks. Bottom line: Codex is great for legacy systems and well-documented tools, but don’t rely on it for cutting-edge development. 2. Limited Context Handling: Good for Simple Tasks, Not Complex Systems Codex excels at generating boilerplate code and handling straightforward tasks. But when it comes to complex, multi-step logic or maintaining context over a long function or workflow, its performance can drop sharply. The AI may return incomplete or incorrect code when the task requires deep contextual understanding. Developers often need to stitch together multiple components, manage dependencies, and think several steps ahead—areas where Codex still struggles. Bottom line: Codex is a helpful assistant for prototyping or writing small functions, but it’s not ready to build robust, production-level systems on its own. 3. Security Concerns: AI-Generated Code Isn’t Always Safe Codex was trained on publicly available code, which means it can unintentionally replicate insecure practices or outdated patterns. It may even generate code that contains known vulnerabilities if those were present in its training data. This is particularly concerning for applications with strict security requirements. Developers must remain vigilant and conduct thorough reviews of any AI-generated code. Bottom line: Treat Codex’s output as a draft. Always audit the code, especially when working on secure or sensitive systems. 4. Legal and Ethical Implications: Licensing Matters Codex’s training data includes a wide range of open-source code, some of which comes with specific license agreements. If Codex generates code that closely resembles licensed material, it could raise legal concerns about copyright infringement. While OpenAI has taken steps to minimize this risk, developers are ultimately responsible for ensuring that any code they use complies with licensing terms. Bottom line: Be cautious. Understand the licensing implications of the code Codex generates and avoid blindly incorporating it into your projects. 5. Risk of Over-Reliance: Don’t Let AI Replace Skill Development One of the subtler risks of using Codex is the temptation to rely on it too heavily. Junior developers might miss out on learning opportunities, while experienced developers could fall into the habit of using Codex as a shortcut. Codex can’t teach you how to write clean, maintainable code or help you understand the architecture of your system. It lacks the intuition and experience that human developers bring to the table.

## The Catch: Why Codex Isn’t a Silver Bullet ### 1. **Code Hallucinations** While Codex is incredibly powerful, it sometimes generates code that looks good but doesn’t actually work. This is especially risky in production environments where undetected bugs can be costly. ### 2. **Lacks Full Context** Codex isn’t yet context-aware in the way a senior dev would be. It doesn’t always understand project-specific dependencies or architectural nuances, meaning it can output functional but flawed code. ### 3. **Security Is Not Built-In** Codex does not inherently prioritize secure code practices. Unless explicitly prompted, it can suggest insecure patterns or miss vulnerabilities altogether. ### 4. **Risk of Over-Reliance** There’s a danger of developers leaning too heavily on Codex and neglecting to learn the “why” behind the code. Like any tool, Codex should complement your skills—not become a substitute for them.

## The real test: Daily production use When I wrote my initial review in May 2025, Codex felt promising but rough around the edges. The kinds of tasks it could handle reliably were limited, error handling was poor, and multi-turn conversations often derailed. Fast-forward to March 2026, and I'm using Codex as a core part of my development workflow both personally and at WorkOS. The difference isn't subtle—it's night and day. … ## What still frustrates me ### Model selection opacity You still can't choose which model handles your task. Codex picks internally based on task complexity, repository size, and probably other factors I'm not privy to. As someone who understands the trade-offs between different model sizes and capabilities, this lack of control is annoying. Sometimes I want to throw GPT-5.2 at a complex architectural decision—it's their heaviest thinking and reasoning model—and sometimes I just need Codex to generate a simple CRUD interface where a smaller, faster model would be fine.

div I’ve been using **Codex in the web version since the very first day of its release**, and I can confidently say that **until around mid-September it was an outstanding tool**. After the release of what seems to be the **GPT-5-Codex** update, things have gone downhill fast. … - Codex **no longer completes tasks reliably** — in roughly **two-thirds of all cases**, tasks either hang indefinitely or end with *“I could not do this task.”* - It **makes a huge number of mistakes and regressions**, even in simple, previously stable workflows. - **The new “code review” feature**, which is supposed to help, now actually highlights how bad things have become — it finds **bugs and logical inconsistencies in almost every piece of code generated by Codex itself**, forcing me to **rerun and re-fix the same task over and over**. - Because of this, **two-thirds to three-quarters of all consumed limits go not into building, but into cleaning up Codex’s own mistakes** — and that’s on top of the fact that **the new usage limits no longer allow running tasks at the same pace and volume as before.** - Front-end generation has become absurd — it **ignores provided designs** and outputs something completely unrelated. To give some context — in **August**, I wrote over **300 000 lines of solid code** with Codex. Now, more than a month later, I **can’t even isolate one persistent bug**, and I’m unable to render a mobile UI without launching separate tasks **for every single component**. Honestly, the best decision right now would be to **roll everything back to the late-August state** and rebuild from there. Because right now, **you’re losing developers who were genuinely invested in this tool** — when **the GPT-5 model embedded in third-party agents performs better than OpenAI’s own core Codex service**, that’s a serious signal that something is fundamentally broken. … Long-running sessions are meaningless when after 1–2–3 hours of continuous work **there’s no guarantee the code isn’t riddled with hidden bugs.** Because of that, we’re forced to **overcomplicate our prompts** — making Codex re-check and re-verify its own output multiple times, trace affected code paths, and cross-check logic. That in turn **overloads your servers even more.** And with the constant **task freezes and “I couldn’t do this task” messages**, we end up running **the same job 4, 8, or even 12 times in parallel** just to get one usable result. So yes — technically Codex can “run for 6 hours,” but practically, **it can’t finish a 6-minute job reliably anymore**. … These days, even tweaking one CSS rule can hang for 15 minutes, so I end up force-quitting it… I was loving it - but I am ready to look for other options at this point. I have to ask it 3x to fix a simple thing, then when even CODEX realizes it keeps attempting the same fix, it says I give up.. I cant do it.. and then refuses even if I give it an alternate working fix. I also have been using Codex Web since it was still fresh, and its usefulness has diminished greatly since GPT5, and especially 5.1. Refactors used to be painless, now they are painful Its in it’s infancy. I recently posted something i never thought it could do but it did. ... Codex used to work really well, but after the recent update it’s become slow, unreliable and full of mistakes. Tasks fail, the code review tool catches problems codex creates itself , and even simple UI works now takes too much time. This is slowing down my work and i just want Codex to go back to the stable version we had before. It was a great tool, and i hope it gets back to that level soon. … It has become abundantly clear to me that feedback is of no importance to the Codex and OpenAI teams; their communication with users and developers is an utter failure. Judging by the situation, either the team has lost its engineering expertise, or it is being stifled by marketers.

But while Codex sounds like a dream come true for speeding up workflows and automating repetitive tasks, there are some crucial limitations that advanced developers and designers should be aware of before fully embracing it. Touted as the next big thing in AI-driven software development, Codex can generate everything from simple functions to entire codebases with ease. ... … ## 1. Outdated Knowledge Base: No Internet Access, No Updates One of the first issues you’ll notice when working with Codex is that it operates entirely based on a training dataset that’s frozen in time. Codex doesn’t have access to the internet, which means it can’t pull in updates on new libraries, frameworks, or tools that have emerged since its training cutoff. For those of us who live and breathe the ever-evolving landscape of development, **this is a pretty big deal.** While Codex is great for working with widely-used, established frameworks and libraries, it struggles to handle the latest tech stacks, APIs, or versions that could be essential for a modern project. Imagine coding with tools from 2021 while everyone else is using the cutting-edge technology of 2025. Not exactly ideal for developers building next-gen applications, is it? **The takeaway:** Codex is fantastic for legacy code or well-documented frameworks, but don’t expect it to keep you on the bleeding edge. ## 2. Handling Complexity: Great for the Basics, Not So Much for the Nuance Codex excels at generating boilerplate code and automating straightforward tasks, like setting up basic functions or structuring simple algorithms. However, when the complexity ramps up—whether it’s in a long, convoluted function or an intricate multi-step workflow—Codex can falter. The AI struggles with maintaining context over long chains of thought. It’s like asking a colleague to solve a multi-part problem without providing the full context. **You’ll often get results that are incomplete or outright wrong**. As developers, we know that coding is rarely as simple as it seems, and complex problem-solving requires a deep understanding of how different parts of a system work together. **The takeaway:** While Codex can speed up development for smaller, isolated tasks, when it comes to larger systems or intricate problem-solving, it’s no replacement for a human touch. Think of it as a useful helper for prototyping, but not for building production-ready systems from scratch. ## 3. Security Risks: AI-Generated Code Is Not Foolproof Let’s talk about one of the biggest concerns when it comes to AI-generated code: security. Codex is trained on publicly available code repositories, meaning **it can inadvertently generate insecure code or replicate bugs and vulnerabilities** present in the data it learned from. This can be especially problematic if you’re working on applications that require a high level of security. Even seemingly benign snippets of code could introduce subtle bugs or, worse, security holes. For example, Codex could reuse outdated or vulnerable patterns from open-source projects that have since been patched. In high-stakes environments where security is non-negotiable, relying on an AI without doing a thorough security audit might be asking for trouble. … For example, if Codex generates a code snippet that’s closely derived from open-source software under a restrictive license, there’s a possibility of infringing on copyright. This creates potential legal headaches for developers who may unknowingly deploy AI-generated code that violates licensing agreements. **The takeaway:** Advanced developers and designers should remain vigilant about the licensing implications of using Codex. Legal issues are rarely black-and-white, so make sure you’re familiar with the licenses of any code Codex generates for you. ## 5. Over-Reliance on AI: Where’s the Code Craftsmanship? One of the more subtle dangers of AI in development is the potential for over-reliance. As more developers and designers start using Codex, there’s a risk of losing some of the core skills that have traditionally defined great software development. Codex can write code for you, but it can’t teach you how to write clean, maintainable code, nor can it help you develop a deep understanding of how your system works. For junior developers, this could lead to a situation where they lean too heavily on Codex, sacrificing the opportunity to learn and improve their coding skills. For senior developers, while it may be tempting to use Codex as a shortcut for repetitive tasks, the real value in development comes from problem-solving and system design. Codex doesn’t have the intuition or experience that seasoned developers bring to the table, and it certainly doesn’t teach the craft of clean code architecture.

But while Codex sounds like a dream come true for speeding up workflows and automating repetitive tasks, there are some crucial limitations that advanced developers and designers should be aware of before fully embracing it. Touted as the next big thing in AI-driven software development, Codex can generate everything from simple functions to entire codebases with ease. It’s like having a supercharged pair of hands to churn through repetitive tasks, freeing developers up for more creative work. … ## 1. Outdated Knowledge Base: No Internet Access, No Updates One of the first issues you’ll notice when working with Codex is that it operates entirely based on a training dataset that’s frozen in time. Codex doesn’t have access to the internet, which means it can’t pull in updates on new libraries, frameworks, or tools that have emerged since its training cutoff. For those of us who live and breathe the ever-evolving landscape of development, **this is a pretty big deal.** While Codex is great for working with widely-used, established frameworks and libraries, it struggles to handle the latest tech stacks, APIs, or versions that could be essential for a modern project. Imagine coding with tools from 2021 while everyone else is using the cutting-edge technology of 2025. Not exactly ideal for developers building next-gen applications, is it? **The takeaway:** Codex is fantastic for legacy code or well-documented frameworks, but don’t expect it to keep you on the bleeding edge. ## 2. Handling Complexity: Great for the Basics, Not So Much for the Nuance Codex excels at generating boilerplate code and automating straightforward tasks, like setting up basic functions or structuring simple algorithms. However, when the complexity ramps up—whether it’s in a long, convoluted function or an intricate multi-step workflow—Codex can falter. … ## 3. Security Risks: AI-Generated Code Is Not Foolproof Let’s talk about one of the biggest concerns when it comes to AI-generated code: security. Codex is trained on publicly available code repositories, meaning **it can inadvertently generate insecure code or replicate bugs and vulnerabilities** present in the data it learned from. This can be especially problematic if you’re working on applications that require a high level of security. Even seemingly benign snippets of code could introduce subtle bugs or, worse, security holes. For example, Codex could reuse outdated or vulnerable patterns from open-source projects that have since been patched. In high-stakes environments where security is non-negotiable, relying on an AI without doing a thorough security audit might be asking for trouble. **The takeaway:** Always treat AI-generated code as a starting point. Make sure to manually audit and review everything it produces, especially for production code. Security audits should never be skipped. ## 4. Ethical and Legal Issues: Copyright and Code Licensing We’re all familiar with the complexities of code licensing—whether it’s MIT, GPL, or proprietary licenses. With Codex, things get a little murkier. Codex was trained on a vast dataset of publicly available code, much of which is open-source with specific licensing terms attached. While OpenAI has taken steps to mitigate risks, there’s still a real concern about generating code that violates these terms. For example, if Codex generates a code snippet that’s closely derived from open-source software under a restrictive license, there’s a possibility of infringing on copyright. This creates potential legal headaches for developers who may unknowingly deploy AI-generated code that violates licensing agreements. **The takeaway:** Advanced developers and designers should remain vigilant about the licensing implications of using Codex. Legal issues are rarely black-and-white, so make sure you’re familiar with the licenses of any code Codex generates for you. ## 5. Over-Reliance on AI: Where’s the Code Craftsmanship? One of the more subtle dangers of AI in development is the potential for over-reliance. As more developers and designers start using Codex, there’s a risk of losing some of the core skills that have traditionally defined great software development. Codex can write code for you, but it can’t teach you how to write clean, maintainable code, nor can it help you develop a deep understanding of how your system works. For junior developers, this could lead to a situation where they lean too heavily on Codex, sacrificing the opportunity to learn and improve their coding skills. For senior developers, while it may be tempting to use Codex as a shortcut for repetitive tasks, the real value in development comes from problem-solving and system design. Codex doesn’t have the intuition or experience that seasoned developers bring to the table, and it certainly doesn’t teach the craft of clean code architecture.

When you ask Claude Code to analyze your codebase, it's "reading" thousands of lines of code, which consumes tokens and costs money . This is also why the rate limits can be a major source of frustration. On a heavy development day, it's not uncommon for users on the Pro or even Max plans to hit their usage caps and be forced to wait for hours . … It represents a paradigm shift from AI-assisted coding to AI-delegated development. ... However, I must be transparent about its downsides. It is not a magic bullet. It has a steep learning curve, the costs are real, and the usage limits can be maddeningly restrictive on a busy day. … The learning curve is steeper than for simple autocomplete tools. The usage limits on lower-tier plans can be very restrictive and interrupt workflow. And like all AI, it can be slow and sometimes produces incorrect or suboptimal code that requires human oversight. Is Claude Code secure for enterprise use? Yes, it was designed with enterprise security in mind.

### Style fanciness (by default): D I described the project as vanilla Python, and I mostly want to keep it that way. Claude on the other hand enjoys using the full range of language constructs available. The most striking example was when I realized that where I would just make a function call, Claude would sometimes spawn a subprocess and make a function call within that, for no good reason that I could see. This led me to write style guides that I have Claude re-read every so often (see Instruction Following below). ### Robustness over correctness: D I suspect that Claude Code was quite literally trained (fine-tuned?) not to write code that crashes. Since my overall system largely runs in batch mode most of the time with me as operator, I don’t care much about uptime, but care intensely about correctness. Subtle silent bugs are killers. If something has gone wrong I want the code to fail with an informative fatal error. Claude, on the other hand, wants to write code like this: … Claude lacks this intuition (which of course leads to the question why I have the expectation that Claude should have reasonable intuitions, or any intuitions at all. It’s because it is so competent otherwise!). But I’ve had the experience of having Claude write and modify web-crawling code, and respond to a newly-broken crawler with a theory that 1000 different independent webservers have simultaneously been reconfigured to return HTTP 500 errors in response to all requests. Or respond to newly-broken code that determines directory paths for file lookup by guessing that someone has maliciously deleted all the files since the last run. … Server: “Excellent choice!” Where Claude fails is not in willingness to following instructions, but in remembering what they were, and sticking to them. Whenever I start a new session Claude is asked to re-read a style guide that tells it not to multiprocess unnecessarily, avoid cascading fallbacks, only put import statements at the top of code files, don’t explicitly catch signals like cntrl-c – a long list of prescriptions that are a mix of basic Python style and my own idiosyncratic desires for how the project should be structured and coded. … ### Machine Learning Methodology: A Claude has clearly seen enough ML projects to have a well-developed sense of all the methodological pitfalls: feature sparsity, overfitting, data leakage, and so on. If you are working on an AI or ML project it’s definitely worth just asking Claude to spin through your codebase not just to find code bugs but also to audit your ML code for soundness.

## Downsides of Claude Code ### Technical constraints and performance boundaries Claude Code demonstrates occasional inconsistency with complex architectural patterns. Particularly: event-driven systems, microservices with intricate communication patterns, or applications using cutting-edge frameworks underrepresented in training data. Context and scope limitations affect effectiveness with extremely large codebases or tasks requiring domain-specific business logic understanding. Key technical limitations: - **Context degradation** during extended sessions requiring periodic conversation history clearing - **Rate limiting** on enterprise accounts based on aggregate load - **Occasional security vulnerabilities** requiring immediate patching ## Security considerations for enterprise deployment Claude Code sends code context to Anthropic’s servers, meaning sensitive code and business logic are transmitted over networks. The tool can inadvertently access environment variables and configuration files containing API keys. Security researchers identified vulnerabilities including *CVE-2025-54794* (path restriction bypass) and *CVE-2025-54795* (command injection), both now patched. … ### Usage management best practices Organizations must track and manage consumption carefully. Community reports indicate sudden usage blocks affecting even Max plan users when consumption patterns trigger undocumented thresholds. The lack of transparent usage dashboards and advance warnings about policy changes creates uncertainty for teams relying on Claude Code for production workflows. Successful implementations establish monitoring workflows using community tools, plan intensive work around weekly reset cycles and select appropriate model complexity for each task. Organizations should:

A recent advisory from Check Point Research revealed details of a trio of vulnerabilities that could allow code to be run remotely or allow hackers to steal API keys by taking advantage of automation and other built-in tools. The flaws shouldn't come as a surprise, given how quickly AI coding tools have been introduced to the industry, said Check Point. ... … "These platforms combine the convenience of automated code generation with the risks of executing AI-generated commands and sharing project configurations across collaborative environments." All three bugs have already been fixed after the security firm disclosed them to Anthropic over the course of several months last year. ... In a blog post detailing the flaws, Check Point said Claude Code introduced a new attack vector by trying to make work easier for developers. The tool is designed to embed project-level configuration files directly within repositories, researchers explained, automatically applying them when a dev opens the tool within any given project directory. While this is a convenient feature, researchers noted that in some instances cloning and opening a malicious repository would be enough to trigger hidden commands, slip past safeguards, and expose active API keys. … The second centered on Model Context Protocol (MCP), an industry system for letting AI models work with external tools. With this flaw, designated CVE-2025-59536, Check Point found that repository-controlled configuration settings could override safeguards that require users approval, letting remote code be executed. "When code runs before trust is established, the control model is inverted – shifting authority from the user to repository-defined configuration and expanding the AI-driven attack surface," the researchers said. The third flaw, tracked as CVE-2026-21852, takes advantage of those repository-controlled configuration settings, researchers said. If a hacker meddles with those, it's possible to redirect API traffic to an attacker controlled server before security protections kick in. That could allow attackers to steal a developer's active API key and other credentials.

### It Enables Me to Do More but Makes Me Lazy and Less Confident I fully understand the issues that many take with actual *vibe coding*. At least for now, I’m relatively sure that in the hands of people without any software development experience, tools like Claude Code will do more harm than good. They produce functional products, but these products, more often than not, are unmaintainable, and we don’t even have to talk about the security implications. … ### It Finds Hacky Ways to Solve Complex Problems In my experience, code (and documentation) reviews are crucial when working with Claude Code. While, at least in my opinion, it rarely fails to create a working solution, it very often finds extremely overengineered and hacky solutions to relatively simple problems. Hence, “it works as intended” is most definitely not a good metric for AI generated code. Instead, I’m constantly reviewing the changes Claude Code introduces and “solve the underlying issue and don’t create a hacky solution” has become a staple of sorts. This also leads to a situation in which keeping the project maintainable becomes an even more constant task than it already is. The moment I start using Claude Code, especially for larger features, I’m likely introducing new maintainability issues. Of course, this is a trade-off and through good prompting and reviews, these issues can be reduced quite well. … Put differently: Despite using `CLAUDE.md`, spec files, etc., I am having a really hard time keeping Claude Code within the scope of what I want it to do. ### It Gets Stuck in the Current Context and Ignore the Rest Generally speaking, it’s a good idea to keep Claude Code’s context as focused as possible. I usually reset the context after each completed task or feature. If done consistently, many common issues are gone! However, a related, and significantly trickier, challenge is the fact that Claude Code often only selectively looks at files and parts of the project. For instance, when I request a refactor of existing code based on a recent change (e.g., a change in data structures), there’s a good chance it will only refactor about 1/3 of the affected files, all while claiming it has reviewed the entire project with utmost confidence. Although this issue can be mitigated through the use of tools, I simply don’t trust Claude Code to consider the whole project, even with a relatively small codebase.