AI is everything. In every industry. In every workflow. In my scenario (biz/software dev) Code-gen tools have become sophisticated enough that they are necessary parts of any development process. better, faster, or some combination of the two is a killer value prop. GPT Enterprise (GPT5 is amazing don’t listen to the ppl on reddit who’d tell you otherwise), cursor, claude-code, MCP tools like linear-mcp and postgres-server can collapse tasks that would take days into hours.
Even LampWorld in this first version is maintained partially by AI, and soon will include AI-enabled features to make the site a lot cooler. We have Sentry set up for comprehensive logging and error monitoring, allowing “agents” to access error logs using MCP immediately while debugging. Coming soon in the feature realm: RAG-powered semantic search, which will use Pinecone to structure the lamp bios ya’ll imported into a vector database that’s easier to maintain and reference.
It’s magic. Even in applications beyond work. Trip planning, meme generation, etc. One massive caveat though: if used properly.
There’s something bizarre and almost dystopian about how we humans interact with AI. I’ve seen a buddy of mine use ChatGPT like a therapist. Recently, I find articles about AI delusions on a near daily basis. So, as my first post here, I thought I’d share a few do’s and don’ts I’ve found helpful to avoid the common “gotchas” and minimize model quirks.
Prompting and context management:
DO always include your goals and be as verbose as possible. Talk to the interface in long form. Things like using the right words and being descriptive have proven to improve the performance of these tools. If you’re coding, always add the up-to-date codebase and database-schema or API route documentation. Attached a promp design from OpenAI’s best practices documentation that’s great.
DO NOT be vague. This increases the probability of mistakes, incorrect assumptions, etc that get users trapped in a feedback loop based on shaky foundations/logic.
Distinguish between an AI problem or not an AI problem:
AI is great for brainstorming, but different standards apply for more rigorous tasks.
DO remember the ELIZA effect: Humans attribute emotions and intentions to these LLMs where none exist. It’s easy to do so; these model providers encourage that behavior. BUT, AI across the board is inherently lazy at the present. Prompt to surface reality checks, avoid unverified assumptions, and encourage reflective interactions. They will not take those steps unless explicitly asked.
DO NOT use AI to fully vibe-code, do vibe-physics, write vibe-documents, or whatever. Anything that involves just pasting whatever the tool says and moving on. Even if the code runs or the AI-generated summary of whatever idea looks compelling, always check the outputs with your own eyes and validate that it is correct to the best of your knowledge. In the long run, independently verifying outputs will still be faster than performing the task from scratch, so you’re still saving time even with this extra step.
Cross check with multiple tools:
DO (whenever possible) use multiple models against the same task. Ex- I have found that Claude is really good at prompting GPT. It is better at framing and reframing instructions whereas GPT tends to just dive straight into execution unless you explicitly ask it to step back and structure the task.
DON’T rely on a single AI output whenever possible just as you would not rely on only one person’s advice/peer review/feedback.
TLDR: AI is not a substitute for independent thinking. If you were sitting at a bar and someone asked you if it was raining outside, would you get out your phone and call the weather bureau in lieu of simply looking outside? Of course not. The same general principle applies to interacting with AI. These tools are super powerful, but no tool (or nobody for that matter) is yet capable of explaining to you a concept, the words on a piece of paper, or the acts caught on a complete video better than you can explain it to yourself (quite the word salad there but the fact remains…)
Attaching below an example prompt and prompt designs that I've found to maximize AI for tasks of all sorts. Curious to hear what practices others in this group have adopted.
Wow.