TL;DR: I've shipped 26+ projects by directing AI agents, and I still couldn't write a Python program that prints one through ten. Applying to Stanford's Code in Place X meant typing my own code for the first time. The logic ended up right. The AI inside it was wrong four times in a row.
The all-time leading scorer for the Los Angeles Lakers is Kareem Abdul-Jabbar.
My program printed that on the first run. I wrote the part that asked the question, and I was very proud of it.
But it isn't Kareem. It's Kobe.
I needed help printing Hello World
Every tool I've shipped over the last year was typed by an agent. In I Didn't Write the Code. Does It Count? I asked whether directing agents counts as learning. I landed on yes, just different. This time I wanted to find out what the typing would teach me.
What I had never done was write the code myself. I took a basic Python course on DataCamp and repeated the practice prompts until they were done, and none of it connected. When I sat down with Code in Place X, Stanford's free six-week intro course, I would still have needed help printing "Hello, World."
The application asks if you're a beginner, and being a beginner is a requirement. If you already know the concepts, even in another language, they want you to teach instead. So I took a five-question gut check. From a blank screen, with no AI, could I write a program that:
- prints the numbers one through ten
- asks your name and says hello back
- prints "high" if a number is over 100, and "low" if it isn't
- prints a list of three courtrooms, one per line
- takes two numbers and returns their sum
No on all five.
That made the application pretty clear. I qualified.
The application ends with short coding lessons. I worked through them with Claude as a tutor, with one rule. I typed every line. Claude could point at a line and ask me what it did. It couldn't write it for me.
The text in the quotes is the question
The first task: ask the user for a color, an adjective, and a goal, then fill in a sentence. My first line was this.
name = input("Green: ")
I had put the answer where the question goes. Whatever sits inside input("...") is what the program shows the person. They type their own answer after it. My program would have greeted everyone with the word "Green:" and waited.
My second try had the prompt right and a minus sign where the equals sign goes. Python would have tried to subtract.
The wrong answers and typos kept stacking up. I was discouraged. I kept going anyway, because I wanted to hit submit on that application.
The computer counts every space
My first version of the sentence would have printed this:
...and the air feltsmelly. I decided today I will finally eat fewer bugs
Missing spaces inside the quotes. A period outside the quotes, which would have stopped the program cold. Then, after I fixed those, a double space I added by accident.
When it finally ran, the tester marked it Incomplete. All three tests. It took me a minute to find out why.
Expected: A goal you would like to achieve:
Observed: A goal you would like to acheive:
One swapped letter in a prompt. Code checks every character, including the ones in the question.
One equals sign per line
The second task was a haiku generator. Take a name and a topic, send them to GPT, print the poem. I tried it myself first and I got the skeleton right. Two inputs, a call_gpt, a variable to hold the answer.
Then I got lost. My prompt to GPT didn't include the name or the topic at all, so GPT had nothing to write about. When I added them, I used = inside the prompt where + belonged. Twice.
At one point I typed this to Claude: "i don't understand i'm so confused which is why i need this class!" Poor Claude. lol
What helped was a picture. The prompt is a chain, and + links each piece. Text, then a variable, then text, then a variable. The = shows up once, at the very start of the line, where the finished chain goes into its box. I changed one character and the line ran.
response = call_gpt("Write a haiku with 5 syllables, 7 syllables, 5 syllables for " + name + " about a " + topic + ".")
print(response)
I sent "it ran!!!" with three exclamation points. I stand by all three.
My code was right. GPT was wrong four times.
The last lesson was open ended. Build something with AI that's fun or useful to you. I read the instructions and wrote this one on my own before asking for any help. It was almost right on the first try, with a few small fixes.
Four of those lines are mine. The rest came with the course's starter file.
Same bugs as before, too. GPT received "leading scorer forbasketballon thislakers" and answered anyway. It's very forgiving about spaces. It's less careful about facts.
According to NBA.com's franchise scoring leaders, the order is:
| Rank | Player | Lakers points |
|---|---|---|
| 1 | Kobe Bryant | 33,643 |
| 2 | Jerry West | 25,192 |
| 3 | Kareem Abdul-Jabbar | 24,176 |
GPT picked number three, with the right point total for number three. I ran it four times. I got four different wordings of the same wrong answer. One run added "Please verify this information with up-to-date sources, as statistics might change over time." The hedge was about time passing. Kobe retired in 2016, years before the "October 2023" date GPT kept citing, so time wasn't the problem.
I tried adding "Research and" to the front of the prompt. Nothing changed, because this GPT has no internet access. It can't research. It can only remember, and it remembered wrong with total confidence.
My tutor wasn't immune either. Claude first told me Kareem was number two. It found Jerry West when it checked NBA.com before putting the numbers on the card above.
You have to laugh at this point. I wrote the program that called GPT, and GPT fumbled the one job I gave it. Another Pyrrhic victory in vibecodeland.
At my day job, an error like this wouldn't cut it. Court outcomes need more than a confident guess.
On my last post, a commenter named @mansio wrote that the typing was never the load-bearing part. I agreed with him then. I still mostly do. But four lines taught me things a year of directing hadn't. Spaces inside quotes are text. An equals sign fills a box once. And the part that held this program up was the same as always: checking the answer.
The more I learn, the more my answer to "does it count" moves. I expect it to move again in October when I build the last project for the AWS AI/ML Learners program.
I left the program as is. It does what I wrote it to do. The answer is GPT's.
So, your first one
That was the last lesson in the application. I submitted it on September 21. If I get in, the course starts October 12.
I'm curious about your first handwritten program. Not the tutorial. The first one you wrote because you wanted it to exist. Did it work?
Mine did. Mostly.
Quick context if you are new here. I work in the California courts, running court operations for the county. I started building with AI in July 2025 and I have been learning in public ever since. For most of my builds, I direct, the agents generate, and I validate and decide. This one is different. I typed it. I build the Clew Suite, a set of civic tech tools for making complex systems easier to inspect.
AI Assisted. Human Approved. Powered by NLP.
Top comments (1)
Spot on. The irony here is that Python did its job, but the prompt formatting was messy and the LLM filled the void with a confident hallucination. Typing code builds the mental syntax map you need for debugging, but without a ground-truth data source to validate against, no amount of prompt engineering saves the output.