With Anthropic’s impending IPO debut, I wanted to write this article as a thank you for the developers and researchers who were able to make Claude possible. On the eventual release of Mythos, Opus and Sonnet have served to wonderful models assisting as personal assistants. Claude Code has become an essential part of my workflow as a engineer and has been fundamentals towards the speed of my development. This article is to brief my viewers the way I have utilized Claude, challenges that have occurred, and the potential skills that I wish to grow in this current time.

Claude - The AI Assistant

I have been using Claude since the summer of 2025. Before that, I had been reliant on ChatGPT for basic knowledge and utilizing bits and pieces of vibe coding. Going into CS3000: Object Oriented Design, I found myself to come across the popularity of Claude. After a few lessons, I decided to give Claude a try as I personally felt the performance of ChatGPT began to decline. Claude had shown itself to approach with a more minimalist style design, utilizing a render engine for code being written by itself and delivered increasingly impressive results. After multiple instances of utilizing Claude with less hallucinations, I had decided to utilized Claude as my main AI assistant.

I found Claude to be excellent in its delivery of summarized informations. Throwing multiple PDFs as a study guide, it has been able to generate tremedous amount of summarized information as well as provide the foundations on what to learn if I was struggling. Treating AI as an assistant had saved me hours of dissecting large amounts of information in a short time, allowing for myself to focus on other needs.

As a result, Claude had subconsciously pushed myself to focus more on system architecture and design. Seeing it as an assistant to do more mundane tasks, I found myself to be able to focus more on how a system works rather than making it work. This eventually drifted me to utilizing Claude Code much more frequently to develop features, scaling systems, and having myself act as a supervisor to ensure high quality production. Shifting myself from spending hours on minor developments, projects can be created in weeks and have myself training more as a senior level engineer to focus on drawbacks and decisions.

Claude Code

I had initially been opposed against Claude Code at the beginning of its release. Acknowledging other tools such as Cursor and Copilot, I had never truly embraced the full potential of vibe coding as I felt that I would be losing out on learning the essential skills of writing code. However, seeing many engineers I have spoken to discuss their usage of agentic code editor tools, I had shifted to utilizing the tools in situations that I felt would be fitting.

Across multiple activities, I had decided to install Claude Code to develop projects and systems, fully realizing the true capability. In just a few months, fully functional systems have been developed that would take years to normally build. As a result, speed of progression in activities had drastically increased and were able to focus on more activities that would continue to help us grow substanially.

However, even now, I am still not utilizing Claude Code to its full potential. I am currently working on exploring new types of strategies and features in order to increase my workflow even more. Just recently, I had learned Claude was capable of writing shell scripts, giving me the ability to now switch between multiple Claude accounts for different memory, context, and usage. As I further go on in this summer, I am looking to see what else Claude can be able to do as it serves as a fundamental tool in our current era of software engineering.

Claude Plugins

Claude Plugins were a topic introduced when I was researching Claude Hooks. I had been introduced briefly to Claude Superpowers and had further gotten into it during my experience of watching other engineers use senior level agents to deliver high code quality. I also heard of Claude Skills as well and am currently using very few Claude Plugins. However, I have been recommended a frontend Claude plugin and will be utilizing it soon to develop responsive frontend code.

Claude Hooks

I had learned of Claude Hooks briefly in a conversation with another engineer. I didn’t fully understand what they were at the time but I had realized it came in development while creating my agents for my agentic workflow to develop systems. The hooks acted as powerful scripts that fired in the Claude Code lifecycle to help it generate the code that I had wanted specifically rather than leaving it to open intepretation. This was able to help me review over the quality of code in familar format without having too much trouble with memory and context. I have not fully explored the usage of Claude Hooks and am currently looking into it.

Claude Agents

Agents have been a hot topic due to their ability to autonomously complete tasks with or without user supervision. When I had first started using Claude Code, I had not realized the full potential of using agents until completing an entire system. After taking a travel break for a week, I had been thinking about ways to improve workflows due to some repetitive prompting within my agents.

As a dive into a workflow, I had been using Claude to interpret and complete a task with details and requirements. After completion, I had found bugs within the code and had Claude to test it again as well as resolve any bugs found. This could be a process that goes back and forth over and over until finishing the task for review. I believed there was a more efficient way to work through this without the usage of humans. Moreover, I believed that some of the code quality generated was poor and had looked for a solution.

Upon some discovery, I had found out other members of my team had been using their own agents as a sort of quality checker for the code. This was a substanial increase to my workflow as I felt this could speed up the progress in development and decided to design a new agentic workflow. As a result, my general workflow became: utilizing an interpreter/planner to understand any prompts given -> have a senior level coding agent write out the code -> having an iterative process to ensure bugs were not an issue. This has substanially increased the quality as the code being distributed had much less bugs that could be resolved through 1 or 2 passes and helped deliver out the final product in a faster manner.

Challenges

As I had used Claude, the system can never be perfect. Claude, itself, is an AI model that is subject to hallucinations. While I have had less hallucinations with Claude as well as a few memory context issues when developing study guides, Claude is still subject towards with some of my needs. More recently, I had utilized an enterprise account with a budget limit and had created this very website using only Claude. Upon checking the final usage, I had utilized 70% of the entire budget due to a mix of poor prompt engineering, using only the Opus model, and Claude Code generating and removing code that was unnecessary. Burning through 70% of the credits given, I had decided to learn how to utilize Claude in a more budget friendly way as the prices of Claude tokens will only increase as time goes by.

Hallucinations

Hallucinations are common in all AI models. Just like humans, AI can not always meet the needs of others perfectly each time. I had strong hallucination issues when developing a study guide for a statistics class and developing this website.

Within the study guide, Claude had interpreted around 20+ PDFs, extracting necessary images and generating sample questions for a cheat sheet. Repeatedly, Claude had struggle to grasp certain concepts and keep it through the multiple iterations of the cheat sheet. This turned from an expected 1 hour of building a cheat sheet and study guide into 6 hours of going back and forth to ensure the information was there. This marked as one of the first times I have had a long iterative conversation to ensure a high quality document would be processed, highlighting an issue within my own prompts as well as the limits of the current model.

Developing this website, I had not used Claude Design previously as I typically use niche and unique designs for any relation in artistic development. As a result, I did not feel confident initially when I had asked for Claude to develop the timeline portion on the home page. However, on the first prompt and my description, it was able to develop a strong starting point with the timeline having the yearly ticks and small events being marked. Gaining some confidence, I had further asked for more details which Claude had been successfully able to construct a complete and satisifactory timeline that I would be able to ship, even if incomplete from the intended design. However, upon trying to explain details on these features, Claude ran into many issues, taking many tokens to understand the context of the new detail and creating various bugs that violated my intended design (collisions between cards and event bars, removing the background entirely). While eventually completing the full design, I feel I needed to research more on the limits of AI as I felt that it had used 35% more of the credits than necessary if: the model was better, utilizing different models for certain context, or having better prompt engineering.

This had rose a mindset shift on how to utilize AI in the future as there will be limits to budget.

Upcoming Skills to Grow

Humans are always evolving. I have been learning the skills I am lacking despite being an active AI user for aorund 2 years. The skills below are things I am currently working on to improve my usage of AI.

Prompt Engineering

Around this time marks the 1 year that I have started using Claude. Nonetheless, I have come to realize recently with the development of this website that I do not have the best prompt engineering skills. I had believed prompts could be more lazily spoken in order to achieve a very similar result. Seeing jobs such as prompt engineers, I had initially believed them to be more of “filler” jobs that could be done by anyone. Introducing a budget limit for Claude usage had changed my perspective entirely as I had leanred that I had to be careful to not reach the limit so quickly. This was apparent considering the fact that it was only the personal website that I had worked on in the first week of June and I had already utilized 70% of my credits. If a project that required a much larger scale needed to be completed in the same month, I would have to resort back to manually coding until the limit was refreshed. Hence as a result, progression speed would be slower and I am currently re-evaluating my usage as I would like to create projects later on this month that are much more complex than a unique UI. Fundamentally, I will start to focus further on making my prompts to be descriptive in order to prevent many hallucinations and nuances that can appear in development.

Claude Automation

As earlier mentioned, I had learned to utilize agents within my coding workflow to act as engineers that could review over the quality of code to allow for time saved when reviewing it over and ensuring a system could function. However, even then, speed can be slow due to the need for manual intervention in case the workflow needs approval or issues come arise. I am developing a new personal workflow for coding cases to utilize in development. One of my proposed workflows for complex systems involves:

• A planner/interpreter who is able to process tasks and details to send over to other agents.
• Senior level engineer agents that are able to write code with high quality, informing other agents of potential drawbacks in their code and design choices.
• Testing agent to develop and run throguh tests through the system, looking for edge cases and telling the quality engiener of any issues
• Senior quality engineer agents that review through the code to look for bugs and potential issues
• Returning back to the engineer to reinforce the code with higher quality, should bugs come arise, and address any issues with their architecture
• System advisors that ensure that all agents are following the requirements without violation
• A superagent that supervises all agents to ensure the workflow runs smoothly and can stop in a severe scenario to communicate with the agents or wait for user approval

This is a proposed workflow for speeding up quality and having a fully autonomous system with the superagent to make the final decision and inform the user of any issues that arise. I feel that this would allow for work to run autonomously without my personal intervention isn’t necessary and allow for me to focus on other tasks that require my physical presence.

Claude Shell Scripting

Upon developing a short shell script to allow for multi-Claude account usage, I am now looking forward to finding other shell scripts that can be developed with Claude. Initially, I thought I would need to develop an entire customized client for switching between Claude accounts as I did not want different accounts to receive the context and memory of others. Upon asking Claude out of curiosity as I had believed it would involve their API with authentication, Claude explained the script only required 3 lines due to built in systems within itself. This makes the possibility of having quick Claude developments for quality of life within the operatin system to be exciting and I am looking to see what else can be done.

Claude Design

As I had personally never used Claude Design before, there have been several reasons on my avoidance. Claude has been the only AI that I have utilized to consist of a frontend previewer. Despite this, I have found Claude to deliver relatively poor, simplistic, and repetitive designs. As a result, I had chose not to use Claude Design due to my own personal design skills to be difficult to explain. Furthermore, seeing many videos of Claude Design has also consisted of the same repetitive designs. However, I am still hopeful for Claude’s improvements in design due to this website’s design being completely fully through Claude Code with no designs given. I believe that there will be a day that I can describe in full detail to Claude and it will be capable of designing to meet my requirements.

Other Explorations

Apart from all the information given, here are other things I found interesting that are still related on the topic of AI.

Claude Mythos

The model that will “change the world” is currently being tested due to security reasons. Nevertheless, I remain to be interested in the true capacity of Mythos as it has been rumored to potentially one of the most powerful, innovational, and destruction models.

Claude’s Backbone Foundation: Constitutional AI

I am currently exploring how Claude has been foundationally built due to its excellent performance in comparsion to current ChatGPT. Upon quick search, I had learned of their methodology of using Constitutional AI (CAI) as its training model to revise its own responses. In comparsion to ChatGPT, CAI does not utilize human input and reduces human lableing ofr safety. This allows for the principles to be readable and helps the model with bias, due to the concerns of using human input for reinforcement learning. The issue with human input for reinforcement learning is it strictly requires the human to be right as well which can be an issue in complicated scenarios. This inspired me to further continue reading AI research papers as I am very interesting in seeing the paper on how Claude has been foundationally built.

Other AI Models

I have experimented with ChatGPT, Gemini, and Copilot throughout my developer experience. I made the conclusion of using different AI for different use cases as I have found all of them to have their own isssues. ChatGPT has had slow performance and many hallucinations recently, leading my experience with them to be poorer and poorer by the day. However, I believe that it is capable and easy when answering general questions quickly, mostly providing accurate answers when prompted. With Gemini, I personally found the code quality to be subpar, matching the same as ChatGPT but not at the level of Claude. However, when asking for research as well as solving mathematical problems, Gemini has had the highest quality of results. I was never a strong fan of using Copilot as I had still avoided general vibe coding at the time. I had found the autocompletion to give mediocre results and sometimes, not anything that I had wanted. However, at the time I did use it, I found it’s ability to write repetitive code with variations to be sufficient enough for achieving tasks (though if possible, an abstract class would have delivered a similar result).