I have been using AI since ChatGPT made its debut. Here’s how I use and think of it.
What I worked on
It started as merely a chat service back then, so for my programming tasks, I was doing old school back-and-forth copy-pasting with AI, not much different than doing it with StackOverflow, but doing it with AI requires more context and gets instant (not guaranteed correct) responses in return.
A lot of 3rd-party wrapper coding assistants and tools emerged, but I didn't bother to check them. Compared to the seemingly satisfying hands-off experience, I prefer the approach of discussing, understanding what is going on to prevent any potential mess it creates.
Fast forward, Claude has rolled out its first AI application layer killer product, Claude Code. I instantly adopted it and started establishing my workflow.
My workflow
For the past few months I have been instructing Claude Code to work for my open source task, cBioPortal Navigator, a MCP Server that constructs precise cBioPortal URLs for users directly from their messages. I’m on a Pro plan, so I have to kind of dance in fetters, with Sonnet model most of the time. My workflow goes like this:
-
Understand requirements
It all starts with some problems or requirements, but sometimes they're well defined and sometimes it needs discussion to be made clear. During this time it will explore codebase for relevant components (I like how these work are distributed to sub Haiku agents) and explain the problem back to me. I maintain a README and a CONTEXT doc for it to familiarize the project faster. Then I’ll review how it sees the problem and make sure it is not missing any critical details.
This first step is where the current chat session gets steered towards the optimal solution. From a local optimum perspective, this is when the chat started to head downhill towards a dreamy valley. If it goes sideway, now is a good time to pull it back before it gets lost in the woods, or simply, waste too much token.
-
Identify solutions and make design decisions
Sometimes I come to it with solutions already in mind, yet it’s still very important to walk it through the problem with Chain of thought (CoT) to make sure it understands the problem well. And sometimes I only have some rough sense or even no idea at all, but the previous discussion can pave the way for clear, possible solutions. It will propose a few options and I’ll clarify and make decisions with design in mind. I often ask for examples on how things would actually work out to understand better.
This is a step when, if not done properly, the architecture starts to get rigid and tech debt starts to build up. This is also when software engineering and taste come into play. AI has a million ways to make things work, then the question becomes - how would you want it to work? There is no real way of avoiding “making decisions” - not making decision itself is a bad decision made.
-
Actual technical implementation
“* Flibbertigibbeting… Done. Here’s what I did…”
With two previous steps laying a solid foundation, it can be expected for Claude Code to produce elegant and efficient code in one go. One caveat is that since model is rolled out version by version, it might have a gap on knowledge about certain new frameworks or best practices, but online docs can easily fix that. In general, this is where Claude Code is best at. I will miss you, coding.
-
Review and test
It is no joke when people say “software engineers have transitioned from code writer to code reader”. I would check git diffs to ensure implementation is done as designed, perform tests to ensure features are working as expected, and then update progress to the CONTEXT doc for easier references next time.
If without earlier steps guiding AI properly, it often generates seemingly correct results but can't stand scrutiny. Therefore this is a responsible step to take to prevent AI slop. I’m very cautious to not offload such burden of quality control to the fellow code reviewer. Offloading work to AI is one thing, offloading work to human is another.
So the whole workflow goes like: requirements → decisions → implementation → verification. It has proven to be efficient for me when building a big project, with its divide and conquer approach, take a big problem apart into smaller manageable requirements and verify results after each implementation.
It’s worth noting that often times when people feel AI dumb or hallucinate, it’s simply due to AI doesn’t have all your context, and unlike us, it’s only thinking when talked to. So context matters. Make sure it knows what you know. Make sure it follows your thoughts. Your expression limit is AI’s understanding limit.
By no means this is an optimal workflow, everyone does it differently and at current fast pace everything can change overnight. But it helped me observe these shift in the traditional software development paradigm:
- Requirements can be taken with more flexibility
- It becomes easier to find optimal solution for a given task
- Coding is becoming a hobby
- More need of code review
What about …?
Other than the Claude Code I used, endless applications around AI have emerged. What about them?
What about Perplexity, GitHub Copilot, Cursor, Manus, Midjourney, Sora, Nano Banana, Codex, Agent, MCP, Skills, OpenClaw… you name it?
AI Trend
They are the living proof that AI is rapidly developing. And now this development has entered the stratification phase, meaning the base model layer have already settled down to 3 main pillars (OpenAI, Anthropic and Google), and the current fight is happening in one upper level, the application layer. Anthropic notably, has come down into this arena cage, fighting against a lot of wrapper 3rd-party applications with its own official variants: Claude Code, Cowork, Chrome support, Office support and is currently scoring with MCP and Skills. They are doing great on building moat so I’d expect their revenue to grow.
Should we try to keep up with everything? Absolutely not. Plus you can't. It is good to explore new tools in general. (Here comes a…) But attention is a scarce resource. It’s rather important to evade the hallucinations of learning, figure out what do you want to build first and then look for the right tools for it, not the other way around. Everything looks like a nail if you’re holding a hammer.
What’s happening
In this AI trend, questions were asked, insights were shared. I came across a few.
Software Industry
AI trend started as a strong wave to replace human workers. I'm Worried It Might Get Bad combined AI layoffs with other worrying scenarios. I think AI is just an excuse for layoffs, and other things just come and go with the grand economy cycles, for better or for worse. Is AI a bubble? It does feel like the dot com bubble with endless startups emerge and AI companies investing each other to the moon. But hey I haven't been born when the dot com bubble started, so this does not constitute investment advice.
Such damage, significantly, has been made to junior developers, with numbers in The Junior Hiring Crisis confirming it. Not only it shows the need to juniors has greatly shrunk, but it also reveals the growing need of seniors, which means there will be a gap in the future, a time when current seniors disappear and no one takes their relay baton, a fault zone that I don't clearly see the fix right now. (Anecdotally AI seems very good at teaching stuff.)
Everything has two sides and AI is no different. It's becoming an era where it rewards high agency, actively learning, exploring. Maybe it has always been. The Last Programmers showed that, as coding load taken off developers’ shoulder, they split into two groups, one pushing workload as much to AI as possible to be faster and faster, the other one still trying to understand what's in the code and have it under control. I agree with it that in a fierce competing, consumer facing market, time is of the essence. But the knowledge of what's under the hood still matters, otherwise we will just end up with more COBOL. And I agree fully with other skills mentioned important in there.
Contradictory to what we are feeling with massive AI layoff wash out, The Efficiency Paradox: Why Making Software Easier to Write Means We'll Write Exponentially More showed another possibility that as AI lowers the entry barriers to knowledge work, more needs can be satisfied with more ambitious systems. We are seeing more and more AI created content or small team providing customized solutions to validate this point. But it all comes down to that simply question - “Is it profitable?” - that will need time to shake out.
AI Progress
How is AI now? Reflections on AI at the end of 2025 emphasized the importance of CoT which aligns with my experience, and believes AGI can happen via multiple approaches. Then Something Big Is Happening caught a lot of attention, saying AI has evolved to the point where it not only did the job well, but also developed taste, and it’s definitely capable of replacing us. I agree to most of it. For the taste part, even though I don’t think it can have perfect taste because there’s no such thing, it doesn’t need to have one. Most of people also just do things insipidly, “it works”. Later on, AI Thoughts calmed these worries by presenting a few challenges for AI companies on model development, hardware limit, legal risk, etc. It also casts lights on what AI can bring to us, such as better models which lead to more cures and better businesses.
Why does CoT work? Anthropic has carried substantial research on how model work internally, recorded in Anthropic’s Interpretability Research. Most notably one, Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet, successfully extracted features in the model that represent concepts. My understanding is that by taking CoT, users message activate these features, thus aligning the model to think in a similar way as current user.
Future
Will AGI happen? From the hardware perspective, there is an interesting debate between Why AGI Will Not Happen and Yes, AGI Can Happen – A Computational Perspective, mainly on whether the computation limit posed can be resolved or not. The former says linear progress needs exponential resources, and GPUs no longer improve, while the latter argues that the models are still underutilized and newer hardware is coming to rescue. Both have their points, but I want to point out one missing factor here: current AI doesn’t have a body like we do. Surely they have mapped all of our words into parameterized features, but those features have no meaning to it. It wouldn’t know how things fall down to ground is due to gravity, nor why apples taste sweet, it seems to know because it was told so. It can’t figure things out on its own. It can’t learn by doing stuff. So I’d say embodied AI will be the final missing puzzle on our way towards the ultimate AGI.
Regarding AI replacing human workers, people’s sentiment is pretty clear. One camp worrying that it will happen or is happening already, the other being optimistic about what AI can enable us to do. And it’s valid to think in either direction. But I think in the end it all comes down to revenue and cost. Will AI providers find a way to be profitable? Will AI adopters make more money than using human workers? Only time will tell. Think of all the human workers in farms, mines, factories. They are there not because we can’t design a machine to replace them.
What do we do
AI has definitely passed its “iPhone” moment. Has it passed the “iPhone 4” moment? It might: the legendary Donald Knuth solved a 30-year problem with Claude: Claude’s Cycles. He says:
“It seems that I’ll have to revise my opinions about “generative AI” one of these days. What a joy it is to learn not only that my conjecture has a nice solution but also to celebrate this dramatic advance in automatic deduction and creative problem solving.”
More and more evidence show that AI is capable of real thinking now. And Everyone around me is starting to try out Claude and get genuinely excited about it.
So here we go, the greatest Pokémon battle of all time, begins now!
Pokémon Master
Now that we’ve determined to join this epic Pokémon battle, how do we become a Pokémon Master? Here are a few points that matter to me:
- As a software developer, continue to grow experience. Gather context, weigh options, make decisions, control quality. If AI is making software developing 10x faster, then it’s also making 10x people generating 10x AI slop, which means 100x tech debts, or 100x chance that it will collapse at some point. Senior experience will be more and more valuable to save that from happening. The Next Two Years of Software Engineering gave really good suggestions regarding potential outcomes I discussed above, one especially, is to become a T-shaped engineer: broad adaptability with one or two deep skills. To learn these broad skills, You Need To Become A Full Stack Person has very good directions, like taste, agency, and other important skills.
- Find your passion. It works no matter how the world change. There will always be things you see that can be improved, needs you notice that can be fulfilled. Build for those, not for the fear of missing out. While passion is a great initial momentum, high agency can help you go even further: High Agency In 30 Minutes is very good article to help build that mindset.
- Build taste. I noticed myself liking certain ways of code instead of “it just works”, and I attribute that to taste. Endless blogs I referenced above have mentioned this point. In software, here is a good read on this topic: What is "good taste" in software engineering? But it goes beyond software, into our daily life. It even leads to my judgement that even if AI is so capable to the point it becomes an omniscient and omnipotent existence, it can get lost in the sea of possibilities. There would never be enough constraints for this optimization problem. It would never know which way could resonate with human the best. But we do, because we are human, and that’ll be the last thing that sets us apart from that silicon-based life.
May we all become a Pokémon Master. See you at the Pokémon Gym. Or the boss battle.