Jinwoo Je

Learned how to use more abstracted, developer-friendly cloud platforms. Very different experience from AWS which is infrastructure-focused and difficult to use (B2B, industry use). Platforms like Vercel and Supabase offer more abstraction with different pricing/business models.
Importance of bird's eye view - understanding how things connect versus focusing on details. Learning the right tools and languages for specific jobs. Been narrowly focused on Python (90% time series data), but building full-stack applications requires other tech stacks, especially JavaScript. Picked up Next.js for the ZipScape project frontend.
Gemini 2.5 Pro is insanely good. Currently my go-to model for most needs - blazing fast, unlimited context, and shares thinking tokens. Second month on Gemini Pro subscription, and DeepResearch has been surprisingly effective. Completed a project I'd been thinking about for years thanks to it. Image generation is also impressive with coherent output and strong object consistency.
Deliberately spent more time using LLMs across all fronts: chat interfaces, integrated IDEs, thinking models, deep research, MCPs, and CLI agents. First time using all 500 fast requests in my Cursor subscription and all $30 credit with Claude Code.
One-shot, ambiguous prompting is difficult beyond toy use cases. However, agentic workflows show promise. Planning, designing, and systematic prompt management (templates, specs) helps significantly. Research is essential - similar to working with a team of experts where you act as PMO. Must understand the problem space, distill required components, and build effective prompts.
Model, context and prompt are crucial. Using a strong model, providing appropriate context, and writing clear prompts makes all the difference.
Spent some time testing out MCPs. While MCP servers get the hype, I see significant potential in MCP clients. Claude desktop app is good but limited to Claude models. Other models are available and claude easily gets rate-limited without Pro subscriptions. Cursor works well but its interface is coding-specific, but isn't purposed for usage beyond coding. ChatGPT announced MCP adoption but haven't seen the client in their desktop app yet. I really hope gemini adopts MCP and creates a desktop client.
Prompting seems much more effectife when step-by-step. I try to avoid one-shot prompts in most cases ("build me an app that does..."). Document contains prompts and iterations used during frontend development. Still figuring out better prompt management systems or workflows - currently keeping notes, versioning prompts, and tracking iterations.
Learned valuable lessons building ZipScape. Proper Git versioning, including what to include and exclude. Lost a weekend's work due to a Git mistake where code in .gitignore was deleted and couldn't be recovered. Learned to be cautious in agentic mode - understanding what the LLM is doing and course-correcting when necessary. Deployment itself is complex but fascinating, involving configuration, environment variables, build processes, and package installation.
Realized LLMs aren't good at initiating code projects. Found much more success when setting up the environment and basic folder structure before handing off to LLMs (e.g., using uv init project, pnpm create-next-app).
Always test after each step. An area for improvement - currently doing manual testing (running the server and testing personally after code changes) which is time-consuming. Need to learn programmatic testing, creating tests for each prompt.
Clarity on what to do and what not to do is essential when working with LLMs.
First encounter with a dataset too large to load conventionally. Spent three days attempting to load parquet files via vertical scaling without success. Required a different approach (distributed processing) which I lack IAM permissions for. Created a workaround by reducing row count through filtering and direct ETL processing. Planning to learn partitioning, EMR clusters, Hadoop, Spark, and related technologies by end of Q2.
Business acumen, product sense, and critical thinking seem increasingly important compared to coding and technical skills, as LLMs abstract away technical details. Similar to how high-level code replaced machine code or low-level programming.
Some notable lessons from Naval Ravikant: True intelligence means achieving desired life outcomes by knowing how to get what you want and wanting the right things. Avoid autopilot living by making deliberate decisions. Spend significant time on major life choices. Balance optimism with quickly cutting losses and use objective reasoning rather than self-limiting labels.
Alaska Airlines is launching direct flights from Seattle to Seoul starting September, which is exciting as they're my favorite airline.