As much as we’d like to say “Make me a website that does X, Y, and Z” and then taking a nap, I’ve found breaking it down is best, even if each step is done by AI.
I started with just ChatGPT for the preliminaries.
I wrote a description, loose set of requirements then had the AI review the requirements for completeness and consistency, then draft them into a requirements.md file.
Then I gave the requirements to the AI and had it generate an architecture and design doc.
Then I had the AI review the design against the requirements a few times, refining them.
I also did a couple of rough paper sketches of layout and photographed them (digital or scan would be fine/better of course).
Only then did I switch the agentic coding AI.
Again, I broke it down into steps.
First, mock up the UI/UX with dummy data.
Implement the backend needed to support the UI.
Hook it all up.
Refine and add features.
Have it do code reviews to remove dead code, consolidate redundancies, etc. Always make sure it meets its requirements and design (modify design if necessary so that it documents the real state). It can also maintain your Readme.md.
I did this in a long afternoon to create a utility site that tracked Azure DevOps PRs across multiple projects and repos, creating links to YouTrack tickets, and providing basic searching and sorting (so a lot of API calls and state persistence, but no local database). I only had a step “fail” once or twice, usually due to misunderstanding or vagueness. No real errors.
By the end of the afternoon, I gave the tool to my team and it’s been being used without error since. a few new features were requested and I did the last few steps for each new features - add feature, code review, requirements review, update design and Readme.md.
I used the Junie AI coding agent by JetBrains with Claude and ChatGPT engines to create a utility web page and service to track PRs by devs across multiple repos and tied to our ticketing system.
I did it as an experiment with my constraint being that I refused to edit code, but I did review the code it made and made it make fixes.
I didn’t do it as a one shot. Roughly, I:
* sketched out a layout on paper and photographed it (very rough)
* I made a list of requirements and has the AI review and augment them
* I asked ChatGPT outside of the IDE to come up an architecture and guidelines I could give to the agent
* I presented all of that info to the AI as project guidelines and requirements
* I then created individual tasks and had it complete them one by one. Create a UI with stubbed API calls and fake data, Create the service that talks to AzureDevOps and test it, create my Node service, Hook it all up, Add features and fix bugs.
Result, fairly clean code, very attractive and responsive UI, all requirements met.
My other developers loved and immediately started asking for new features. Each new feature was another agentic task, completed over 1-3 iterations.
So it wasn’t push button automatic, but I wrote 0% of it (code wise) and probably invested 6-8 total hours. My web dev skills are rusty, so I think the same thing would have taken 4-5 days and would not have looked as nice.
There is no “social” media anymore, there’s just algorithmic feeds of dubious news, memes, ads, AI generated content (and AI generated ads), promoted content, and short form videos by people you don’t know. You have to really dig to keep up with your friends and people who have interesting things to contribute unless they have the numbers to break through. Or you click and filter a lot.
I just moved to a house with a barely finished basement. White walls, white painted floor, exposed ceiling joists and ductwork painted black. I’m experimenting with cheap projectors and lighting effects (using clamps to attach to the joists as if they were a truss) and furniture on wheels to create a configurable virtual space with full wall projections, sound, and lighting to match (but not overpower) the video. My plan is to make a camera/light platform with a cheap projector, and Raspberry Pi, and directional LED lighting so that I can coordinate all of them over the network. It’s also my office, library, game room and I have some awesome ideas on how to use the space to augment D&D games. But the white concrete floor has got to go - too bright, too cold, too hard, and too loud.
Only if the goal is to run the result and never have to update it or add features. Several of the good test projects I’ve made from scratch with AI (my title at work needs to be “Speaker to Silicon” because I’m usually tasked with experimenting with AI tools) have worked and looked great. Then someone wants a new feature. No problem, it adds it. Then you say, add that feature to this other part of the program, and it does it, but if you don’t look at the code, you realize it re-implemented it, so if you go back in a month and request a change, it only gets applied to the first place it finds. I had to constantly say “DRY! Don’t implement it twice, share the code!”
I haven’t done any serious web coding in years, so when I needed a little web page dashboard, I thought I’d do it 100% vibe coded.
Problem statement: We have four major repos spanning two different Azure DevOps servers/instances/top-level accounts. To check the status of pull requests required a lot of clicks and windows and sometimes re-logging in. So we wanted a dashboard customized to our needs that puts all active pull requests on each repo into a single page, links them to YouTrack, links them to the Azure DevOps pages, auto-refreshes, and flags them by needing attention for approval, merge conflicts, and unresolved comments. And it would use PATs for access that are only stored locally and not in the code or repo.
AI used: I began by describing the project goals to ChatGPT 5 and having it suggest a basic architecture. Then I used the Junie agent in JetBrain’s WebStorm to develop it. I gave it the ChatGPT output and told it to create a Readme and the project guidelines. Then I implemented it step by step (basic page layout, fill with dummy data, add Azure API calls, integrate with YouTrack, add features).
By following this step by step iteration, almost every step was a one-shot success - only once that I remember did it do something “wrong” - but sometimes I caught it being repetitive or inconsistent, so I added a “maximize code reuse and put all configuration in one place” step.
After about 3 hours, some of which was asking it code to my standards or change look and feel, I had a very full featured application. Three different views - the big picture, PRs that need my attention, and active PRs grouped by YouTrack items. I gave it to the team, they loved it and suggested a few new features. Another hour with the Junie Agent and I incorporated all the suggestions. Now we all use it every day.
I purposefully didn’t hand edit a single line of code. I did read the code and suggested improvements, but other than that, I think a user with no programming experience could have done it (particularly if they asked chatGPT on the side, “Now what?”). And it looked a helluva lot better than it would have if I coded it because I’m rusty and lazy.
Overall, it was my biggest success story of AI coding. We’ve been experimenting with AI bug triage, creating utility functions, and adding tests to our primary apps (all .NET Maui) but with a huge code base, it often missing things or makes bad assumptions.
But this level of project was near perfect capability to execution. I don’t know how much my skills helped me manage the project, but I know that I didn’t write the code. And it was kinda fun.
My rule is that if other human beings eat something for pleasure (and not out of desperation, a dare, or to show off), then I should at least try it a few times as long as I don’t have ethical qualms about it.
I tried salted licorice. Granted, I don’t really like sweet licorice, or anise, or fennel, or any of the liquors that use that flavoring, but I tolerate them. The salted licorice was the worst thing I’d ever tasted.
So I bought a whole bag of it and ate a piece every day or so. After a week, I wasn’t cringing as much. After two or three weeks I started craving it. By the end of the month, I liked it. I don’t love it, but I did buy another bag when that one was done. And yes I know the health risks, but I’m never going to be eating a bag or two a day.
The weirdest, though, was cilantro. I’m in the genetic group that thinks it tastes soapy. And yet, after trying it enough, I love it.
With me, it has been story/novel ideas. The AI is a genuinely useful tool for brainstorming through ideas and giving historical and scientific background. I don’t let it write the stories, but I throw ideas at it and it riffs on the ideas which gives me new ideas and so on. Useful, but you realize it’s 4AM and you’re obsessively plot outlining a trilogy and sketching out characters and inventing a new economy when you connected to ask a personal finance question.
I’ve found it useful, but I recommend a “give me an honest critical evaluation as if you were an editor/agent/publisher” and “Is this derivative of anything?”
Just an anecdote, but back in college, I had an algorithms professor who gave us a classifier problem like the square and triangle boundary problem. His English was poor and nobody understood the problem as he stated it. I got an okay score on it, but never understood it very well.
Anyway, it’s 40 years later and I just read this article and said, “Oh! Now I get it.” A little too late, for Dr. Hippe’s class.
WDE! I remember calling the Foy desk one night (this would have been late 80’s) because someone at a party said “It says Crun-chy” in a Paul Lynde voice and we couldn’t remember the name of the rat from Charlotte’s Web (Templeton).
Reminds me of the time I turned myself into a Van de Graff generator at work.
I was a theater projectionist, back when you had 20 minute reels you had to constantly change, while babysitting two high-voltage, water-cooled, carbon arc projectors. Sometimes the film would break and you’d have to splice it. So when the theater got a print in, you had to count and log the number of splices for each reel, then the next theater would do the same and retire the print when it got too spliced up (plus, sometimes if it was the last night of a run, some lazy projectionists would splice it in place with masking tape and then you’d have to fix it). Sometimes you had to splice in new trailers or remove inappropriate ones as well.
Anyway, you counted splices by rapidly winding through the reel with a benchtop motor with a speed control belted to a takeup reel while the source spun freely. Then, while letting the film slide between your fingers, counting each “bump” you felt as it wound through. I was told to ground myself by touching the metal switch plate of the speed control knob with my other hand. One night I forgot and let go until my hair started rising. I’d gone through most of the reel at a very high speed and acquired its charge.
I reached for the switch plate and shot an 8-10” arcing discharge between the plate and my fingers.
Lesson learned, I held the switch plate from then on.
Yeah, completely automatic DMing just makes it like an open-ended computer game. On the other hand, the AI generated NPC responses are better than I could have come up with quickly, which makes me think co-DMing with an AI partner might be fun, particularly if they get better at remembering the details. That would let the human DM control the flow of the story without worrying that Thug #4 has taken 3hp of damage and is Dazzled. Of course, I most often use Foundry or Roll20 these days and they do a lot of that for you.
I wonder how well it could do for game summaries and recaps?
I don’t think it would have an impact great enough to explain the discrepancies you saw, but some chess engines on very low difficulty settings make “dumb” moves sometimes. I’m not great at chess and I have trouble against them sometimes because they don’t make the kind of mistakes humans make. Moving the difficulty up a bit makes the games more predictable, in that you can predict and force an outcome without the computer blowing it with a random bad move. Maybe part of the problem is them not dealing with random moves well.
I think an interesting challenge would be looking at a board configuration and scoring it on how likely it is to be real - something high ranked chess players can do without much thought (telling a random setup of pieces from a game in progress).
I found the audiobook to be a superior experience to reading the book as well. It think PHM is an excellent primer on that type of SF for someone who hasn’t read something like it before. My daughter, who never reads hard SF, loved the audiobook.
I once commented on Twitter that the Anansi Boys audiobook read by Lenny Henry was better than the book. Neil Gaiman responded, “I agree”.
I read Solaris years ago and was unimpressed. The philosophy and science’ish elements were poorly described and it felt like a lit-fic attempt at trippy SF by someone ignorant of the genre.
Then, a year or two ago, I read about how bad the early translations were, so I picked up a new English translation. Wow, what a difference. Now it’s one of my favorites.