About this draft
This is an in-progress document. It covers the first seven weeks of Fall 2026: all of 15-113 except the Project 2 survey, and none of 15-114. It will be extended as the remaining surveys come in and finished at the end of the course.
Like Spring's first draft, this one was generated with AI (Claude Opus 5.5) from the anonymized survey responses. Given the need to produce this immediately prior to our last lecture, I have only done a light manual review, and will do a more thorough review later. However, every quotation was checked by script against the survey files, the statistics were computed directly from them. Sorting free-text answers into themes, and the counts attached to each theme, are an AI's reading and deserve a critical eye.
It builds on the Spring 2026 report, with the summary moved to the front, and compares the two semesters wherever they did the same assignment.
How to read this report
Part one is the summary: an updated list of best practices, the tools students used, what has changed since Spring, and what is still to come. Part two covers the intake survey and four in-class discussion prompts: who is in the class, why they are here, and their opinions on AI. Part three goes through the four assignments surveyed so far in detail.
The stat strip. Averages from the rating questions. Satisfaction, code read, and understanding are on a 1 to 7 scale; everything else is on a 1 to 5 scale. The small line under each number is Spring 2026's value for the same assignment, where the same question was asked.
Themes. What students wrote in the free-response fields, grouped. Each theme is tagged with how many students raised it:
Example theme
A dominant theme (5 dots) was raised by a third or more of respondents on that survey. Strong (4 dots) means a fifth or more. Common (3 dots) means a tenth or more, and at least four students. Some (2 dots) means two or three students, and one student (1 dot) marks a single observation. One answer can count toward more than one theme.
Quotes are verbatim, typos included, and cited as [S## · HW#]. A leading or trailing ellipsis means the excerpt starts or stops mid-sentence. The student identifiers are new this semester and do not correspond to the ones in Spring's report. Identifiers were manually removed from the intake survey prior to analysis, so its quotes are cited as [HW1].
Course context marks a note about what was said in lecture or changed in an assignment, where that bears on how to read a theme. Several of the most repeated tips (be specific, build incrementally, ask for a plan) also appear in the assignment writeups in both semesters. Note that students had access to ChatGPT/Codex Pro trials, along with CMU-provided Gemini in the browser, but did not have access to Claude, suggesting many students already carried paid subscriptions. The intake survey bears this out: 22 of 55 reported a paid or provided Claude plan (Figure 4).
Part one · summary
Best practices so far
Spring's list, updated with what this fall's students have said through week 7.
Fourteen practices, ordered roughly by how often this fall's students raised them. Twelve come from Spring's list and keep their Spring titles, with the Spring number noted; two are new. Each tag says whether this class has echoed the practice yet. Practices 12 and 14 rest on Spring assignments this class has not reached.
Feed the AI concrete references, not descriptions.
The most repeated theme on Project 1 (10 of 48), and raised on every assignment since. Five students added that the AI's own design sense is weak without a direction.
"AI's design ability was very poor, but given a clear direction and visual target, it succeeds very well at creating what you want." [S19 · P1] · "showing ai the actual issue in an image seems to work much better than typing the issue" [S31 · HW4]
Build incrementally. Get to an MVP before you get fancy.
Raised on all four assignments, most on HW 2 (9 of 39). One student there would skip the simple version next time.
"Starting from the bare-bones working prototype and then building functionality around it." [S22 · HW3] · "Prompt little by little" [S14 · HW2]
Write a detailed plan: ideally a SPEC.md, before any code.
Being specific and settling the design first came up on every assignment. A design spec appears on HW 2 and spec files on HW 4; in Spring's report, SPEC.md first appears on HW 8. Under HW 2's 30-minute clock, four students said planning cost them the sprint.
"Majority of time spent should be discussing design details and really specifying everything to run into way less issues later." [S7 · P1] · "…I tried using SPEC-driven development, and it helped to come up with categories of things I haven’t really thought of." [S22 · HW4]
Use multiple models deliberately, for what each one is best at.
Raised on every assignment, most often as one AI writing the prompts for another.
"Prompt using a different LLM than the one you’re using to build the project" [S34 · HW3] · "I tried getting AI to generate a prompt for me and it worked really well for this implementation…" [S17 · HW4]
Ask the AI to ask you questions before it builds.
Raised by 2 students on Project 1, 6 on HW 2, and 1 on HW 3, and by 8 of 38 describing their ideal workflow. It was on a lecture slide in both semesters; Spring's report has no theme about it.
"Tell your ai to ask any questions before it starts" [S5 · HW2] · "I didn't ask the model to ask me some questions I think that caused me to make a lot of mistakes" [S26 · HW2]
Move into IDE-integrated agentic tools by mid-semester.
44% worked mainly in an IDE on Project 1 and another 15% in Codex or a similar tool, despite a browser recommendation. By HW 4, 86% worked outside the browser. Many arrived that way: at intake, 19 of 55 already described an agent in their editor or terminal as their usual workflow.
"Using Claude in vs is by far my favorite way to use Claude so I can actually look at the code easily" [S23 · HW3]
Separate concerns. One chat (or agent) per module.
Five students on HW 4. Two had handoff files or a prompt log carry the state between agents, where Spring's students briefed each chat by hand.
"Having agents write detailed handoff files to each other was very useful to have frontend connect to backend seamlessly with separate agents" [S7 · HW4]
Protect what already works: it adds and overwrites things nobody asked for.
Seven students on Project 1 (four on unrequested additions, three on protecting working code) and one on HW 4. In Spring's report this first appears on HW 6.
"…first state what they did right and then ask for changes so they don’t accidentally change features already working/wanted." [S41 · P1] · "My Ai agent had a tendency to touch code that I did not ask it to, so any code I wrote would be overwritten" [S36 · HW4]
Understand the platforms, not just the code.
Deployment trouble was a strong theme on Spring's HW 4. This fall three students said it went more smoothly than expected and two described friction.
"I had to get the API key from TMDB myself and I couldn’t use AI to get it" [S32 · HW3] · "Render was easier than I expected" [S26 · HW4]
Guard your secrets and your rate limits.
Credits and usage limits came up on Project 1, HW 2, and HW 4. Trust in the security of a well-configured backend averaged 3.1 of 5.
"Kiro uses a lot of credits so many people switched to other models." [S19 · HW2] · "My classmate was also confused how to set up Github Pages with API key hidden." [S8 · HW3]
When the chat gets stuck, start a new one.
Two students on Project 1 and two on HW 4. On HW 2 the restart was more often a switch of tools.
"…sometimes it’s easier to start back at the beginning of where you were with that part and giving it a fresh prompt instead of trying to just keep editing" [S31 · P1] · "…I restarted with Claude and immediately got better results" [S17 · HW2]
Keep reading the code.
Spring's warning came from the agentic assignments (HW 8 and HW 9), which this class has not reached. At intake, 12 of 55 described not being able to follow or verify AI-written code, and a few students have since said that understanding the code matters. Code-read ratings on HW 2 to HW 4 averaged 3.0, 3.6, and 3.3 of 7 (Spring: 2.9, 3.5, 3.8).
"…I do not know exactly what the AI is doing or how to develop without using AI." [HW1] · "Definitely try to understand what code changes the AI made after each prompt." [S42 · P1]
Before pasting an error, spend 20 seconds reading it.
No one this fall has said to read the error yourself first. Students described how to hand it over.
"Being able to very accurately describe errors in detail makes it very simple to fix bugs in one shot" [S7 · HW3] · "it’s easier to debug with ai quickly if you give it the problem but also any errors in the console, terminal, etc" [S35 · HW4]
Document for the next person and for the next you.
Spring's evidence came from HW 7 (the code handoff) and HW 9, both still ahead. The closest so far is documentation written for an AI to read.
"…as long as I kept updating the prompt log I would tell it to read that so it knew where we were." [S23 · HW4]
The toolkit so far
Two pictures of the first four assignments, next to Spring's numbers for the same four.
Figure 1
Where students worked and which tools they named, Project 1 through HW 4
Fall 2026 · where students mostly worked
Fall 2026 · model families found useful
Spring 2026 · tools named among the most useful
Figure 2
How students rated the same four assignments, this fall and in Spring (1 to 7 scale)
Fall 2026 Spring 2026
Table view
| Average rating | Project 1 | HW 2 | HW 3 | HW 4 |
|---|---|---|---|---|
| Satisfaction (of 7) | 5.3 S26 5.3 | 5.1 S26 4.4 | 5.2 S26 5.1 | 5.3 S26 4.8 |
| Understood the code (of 7) | 4.2 S26 4.4 | 3.3 S26 3.6 | 3.6 S26 3.9 | 3.4 S26 3.6 |
| Read the code (of 7) | not asked | 3.0 S26 2.9 | 3.6 S26 3.5 | 3.3 S26 3.8 |
| No-bugs confidence (of 5) | 3.8 S26 3.8 | 3.0 S26 2.2 | 3.3 S26 3.0 | 3.3 S26 2.8 |
| Understanding sufficient (of 5) | 3.2 new this fall | 2.9 new this fall | 3.3 new this fall | 3.2 new this fall |
| Responses | 48 S26 42 | 39 S26 41 | 42 S26 39 | 36 S26 37 |
What changed since Spring 2026
On the same four assignments.
- Where the work happens. 58% worked mainly outside the browser on Project 1 and 86% by HW 4, with the same browser recommendation as Spring. On Spring's Project 1, 10% named GitHub Copilot and 5% named Cursor. Many arrived that way: at intake, 19 of 55 described an editor or terminal agent as their usual workflow (Figure 4).
- Which models. Gemini led all four assignments in Spring (31% to 51%); this fall 6% to 15% named it. Claude was named by 26% to 34% in Spring and 52% to 64% this fall; ChatGPT by 26% to 36% and 35% to 57%. The preference predates the course: at intake, 31 of 55 named Claude as the model family they used most for coding, and 22 reported a paid or provided Claude plan.
- Agentic habits arrived early. A design spec appears on HW 2, and spec files, several agents, and handoff notes on HW 4. In Spring's report these belong to HW 8 and HW 9. Spec-driven development and multi-agent review were introduced in lecture in week 5 this fall.
- A new repeated tip. Asking the AI to ask questions first (practice 5). The slide suggesting it was shown in both semesters.
- Fewer deployment complaints on HW 4 (practice 9). This fall's writeup added a section on testing locally.
- Two complaints showed up sooner. Unrequested additions and weak design sense both appear on Project 1. In Spring's report they first appear on HW 6 and HW 3.
- Ratings. Satisfaction and bug confidence are level or higher on all four assignments. Self-reported understanding is lower on all four: 4.2, 3.3, 3.6, 3.4 of 7 against Spring's 4.4, 3.6, 3.9, 3.6. See Figure 2.
Keep in mind when comparing
- A different cohort. 15-113 now requires 15-110; Spring's class required 15-112. Even so, 39 of 55 intake respondents had finished 15-112 or a later course, and 9 reported neither that nor another rigorous course (Figure 3).
- HW 2 is a different assignment. The other three kept their briefs, with longer time estimates on HW 3 and HW 4.
- The tools question changed from one combined question to two: model families, and working environment.
- This class has seen Spring's report. It was shared in week 5, before the HW 4 survey.
- Theme counts are a reading of free text. A student who did not write about something may still do it.
Still to come
- Project 2. Still in progress when this draft was written; its survey has not been given.
- 15-114. The second half of the course has not started. In Spring it covered databases, a code handoff, an agentic build with AI review, phone apps, and the final capstone.
This report will be updated as those surveys come in and finished at the end of the course.
Part two · motivations and perceptions
The intake survey and four in-class discussion prompts. Spring's report has no equivalent sections.
Week 1Who is in the class
HW 1 intake survey · coding background and AI habits before the first project
HW 1 asked students about their programming background and how they already use AI. Student identifiers were manually removed prior to analysis, so its quotes are cited as [HW1] and cannot be matched to the S## numbers used elsewhere. It did not ask about majors.
Responses
55
Finished 15-112 or beyond
71%
Self-rated at 15-112 level or above
76%
Taking another programming course
45%
Have a premium AI plan
69%
Figure 3
Coding background coming into the course
Ability to build without AI, self-described
Most advanced course completed
Most familiar language
25 of 55 were taking another programming course at the same time.
Table view
| Answer | Students (of 55) |
|---|---|
| Ability without AI | |
| Could not build Tetris, even with help | 4 |
| Tetris, with difficulty and lots of help | 9 |
| Typical after 15-112 or 15-122 | 34 |
| Nearly graduated CS major | 7 |
| Professional, production-ready apps | 1 |
| Most advanced course completed | |
| 15-122 or beyond | 19 |
| 15-112 | 20 |
| Another rigorous course only | 7 |
| 15-110 only | 2 |
| None of the listed courses | 7 |
| Most familiar language | |
| Python | 41 |
| C, C++, or C0 | 9 |
| Other or mixed | 5 |
Figure 4
How students were already using AI
How often they use AI to…
Almost neverOnly very occasionallyMore often than notAlmost alwaysRather not say
Number at right: share answering "more often than not" or "almost always".
Model family used most for coding
Premium plans they pay for or are given
A student can have more than one.
Usual coding workflow
Sorted from each student's free-text description.
Table view
| How often they use AI to… | Almost never | Only very occasionally | More often than not | Almost always | Rather not say |
|---|---|---|---|---|---|
| Summarize an unfamiliar topic | 5 | 6 | 31 | 13 | 0 |
| Edit or debug personal projects | 5 | 10 | 19 | 21 | 0 |
| Check understanding of a new topic | 8 | 9 | 23 | 15 | 0 |
| Write code or text for personal projects | 5 | 13 | 19 | 18 | 0 |
| Compose or improve emails and messages | 8 | 12 | 15 | 17 | 3 |
| Edit or debug coursework | 9 | 18 | 20 | 7 | 1 |
| Explore how AI behaves or breaks | 10 | 16 | 13 | 13 | 3 |
| Write code or text for coursework | 20 | 10 | 21 | 3 | 1 |
| Answer | Students (of 55) |
|---|---|
| Model family used most for coding | |
| Claude | 31 |
| ChatGPT | 17 |
| Gemini | 5 |
| Other or none | 2 |
| Premium plans | |
| Claude | 22 |
| ChatGPT or Codex | 19 |
| Gemini | 6 |
| None reported | 17 |
| Usual coding workflow | |
| An agent in the editor or terminal | 19 |
| A chat window, copying code across | 15 |
| Tool or setting not stated | 12 |
| No workflow yet | 9 |
Figure 5
How students saw themselves, their classmates, and AI
"I am confident that I can…" (the five course objectives)
Strongly disagreeDisagreeNeitherAgreeStrongly agree
Number at right: share who agree or strongly agree.
Has AI helped or hindered the ability to…
Greatly hinderedHindered a littleNeither helped nor hinderedHelped a littleGreatly helped
Number at right: share answering "helped a little" or "greatly helped".
Worry about keeping up, from 1 (never) to 5 (always)
1 (never)2345 (always)
Number at right: share answering 4 or 5.
Table view
| I am confident that I can… | Strongly disagree | Disagree | Neither | Agree | Strongly agree |
|---|---|---|---|---|---|
| Build ambitious projects faster with AI, and keep code quality | 2 | 4 | 11 | 29 | 9 |
| Navigate ethical considerations | 5 | 6 | 8 | 27 | 9 |
| Articulate a personal framework for when to use AI | 2 | 5 | 13 | 32 | 3 |
| Choose the right AI tools for a task | 2 | 15 | 12 | 20 | 6 |
| Critically evaluate AI-generated content | 2 | 12 | 16 | 20 | 5 |
| AI has helped or hindered the ability to… | Greatly hindered | Hindered a little | Neither helped nor hindered | Helped a little | Greatly helped |
|---|---|---|---|---|---|
| Create impressive projects · you | 2 | 0 | 7 | 17 | 29 |
| Create impressive projects · your peers | 1 | 1 | 10 | 18 | 25 |
| Learn challenging skills · you | 1 | 8 | 6 | 15 | 25 |
| Learn challenging skills · your peers | 1 | 6 | 14 | 19 | 15 |
| Find new opportunities · you | 0 | 1 | 17 | 21 | 16 |
| Find new opportunities · your peers | 1 | 0 | 14 | 27 | 13 |
| Rating | 1 (never) | 2 | 3 | 4 | 5 (always) |
|---|---|---|---|---|---|
| How often I worry I am less proficient than others | 3 | 9 | 15 | 17 | 11 |
| How often I think other students worry | 0 | 10 | 36 | 7 | 2 |
Themes from the free responses
I can build it, but I cannot always follow or verify it.
"I am producing output that I was never able to do before, such as full-stack applications, however I do not know exactly what the AI is doing or how to develop without using AI." [HW1] · "I find it hard to verify the output at times and make my work reliable and reproducible." [HW1] · "…its code is often very complicated and difficult to understand." [HW1] · "…not being confident with the code as it is not written by us but just a faith in a chatbot." [HW1]
It makes me lazier, or takes the work away from me.
"I like how it can write a lot of boilerplate code, but dislike how it takes away ownership from my code, and it makes me lazier. I don't know how to learn well with AI." [HW1] · "I like that I'm being guided but also I see myself being less creative than I used to be." [HW1] · "How to use AI but not be controlled." [HW1]
Fine on small tasks, shaky on big ones.
"Solving small programming syntax questions, building simple prototype is good. Building a production app with a database and security taken care of I am unsatisfied with." [HW1] · "…when I've used ai to vibecode, the code has become too long eventually and when I gave the ai the code and the issue, it coudn't solve the problem anymore." [HW1] · "…when it comes to more complex things such as real game mechanics, I feel that it struggles." [HW1]
Classmates: leaning on it too much, or ahead of me.
28 of 55 rated their own worry about keeping up at 4 or 5; 9 rated their classmates' worry that high (Figure 5).
"I think my peers fall into two categories; either they really know how to use AI, or are ruining their learning with it." [HW1] · "I think many of my peers do not take the "think first, AI second" approach, and use AI to replace the thinking portion which hinders their learning." [HW1]
"My peers likely use AI for coding in more advanced and in-depth ways." [HW1] · "I think some of my peers have really leaned into using AI efficiently. I'm not as satisfied with how proficient i am with these tools." [HW1]
Other observations.
"I know that it is hard to be specific enough in many prompts and I wish that clarfiying prompts for the user (me) were given." [HW1]
"Its feels like sometimes its a blurry line when talking about vibe coding in SWE interviews…" [HW1]
Week 1Why students enrolled
8/24 · first day of class · reported second-hand
Students asked a neighbor why they were taking the class and wrote down the answer, so each response describes a classmate. Names are left out. Asked directly in the intake survey what they hoped to learn, 33 of 55 wrote about using AI more effectively, 10 about a portfolio or career, and 10 about understanding or keeping ownership of what the AI writes.
Responses
45
Themes
To get good at using AI, and to understand it.
"Just copying and pasting code into ChatGPT doesn’t work, so learning how to use it effectively is what she wants to get out of it." [S22 · Enrolled] · "Already have experience with using AI to code in a professional setting and wants to get better at using AI with code and understanding its limitations" [S39 · Enrolled] · "Learning how to navigate Ai designed implementations to learn how to edit them and understand how they work" [S31 · Enrolled]
A portfolio, and projects worth showing.
"They want to make an interesting project while improving their portfolio for recruiters." [S3 · Enrolled] · "Saw the final projects of the course and wanted to gain the ability to be able to use AI to build impressive projects for portfolio" [S52 · Enrolled]
It fit the schedule, or followed naturally from 15-112.
"Want to take a chill course that can deepen their experience with AI" [S42 · Enrolled] · "They thought it was a good next step after a challenging course like 15-112 and was glad they could fit it into their schedule this semester." [S44 · Enrolled]
Internships and jobs had already made the case.
"…she found it hard to prompt effectively during her internship experience. She would like to know how could people code effectively and keep integrity." [S58 · Enrolled] · "They spent the summer doing a start up and vibe coded a lot so they want a deeper understanding of tools and practices" [S23 · Enrolled]
A friend who took it, or last semester's projects.
Spring 2026 was the first offering, so this route did not exist then.
"Heard good stuff from the friend who took it last semester and wanted to make the portfolio!" [S55 · Enrolled]
Week 2How students felt about AI
8/31 · before the first project was due
Students compared notes with a neighbor on how they feel about AI, then wrote down one thing they agreed on and one place their experiences differed. Names are left out.
Responses
50
Themes
Where neighbors differed: how much they use it, and for what.
"We all had different use cases and different amounts of use." [S16 · Sentiment] · "I use AI to write emails while the person next to me doesn’t." [S57 · Sentiment] · "So I prioritize speed while my neighbor prioritize depths." [S45 · Sentiment] · "Most of us just use AI on a conversational level, less actual agentic use experience" [S15 · Sentiment]
Useful, and a risk to learning and independent thinking.
The most common point of agreement. At intake, 40 of 55 said AI had helped their own ability to learn challenging skills and 9 said it had hindered it (Figure 5).
"We both agreed that AI is very useful for studying and coding, but relying on it too much could reduce independent thinking." [S13 · Sentiment] · "We all agreed that its super useful for understanding information, yet its worrisome how much we are offloading our critical thinking." [S62 · Sentiment] · "People who are new to coding will lack practices if they overused AI. But they can produce nearly-professional level code." [S2 · Sentiment]
Jobs, especially entry-level ones.
"Many jobs will disappear but new ones will be formed. The bar for entry level jobs will increase." [S21 · Sentiment] · "Mainly about no entry level jobs and everything becoming ai where humans aren’t really needed" [S12 · Sentiment] · "I believe that AI will help the job market, which the others don’t agree on." [S47 · Sentiment]
Privacy and personal data.
"We think that privacy is an important topic with AI since we never know where our information is stored or will be sent to." [S41 · Sentiment] · "Worry about the loss of our personal data, whether it’s being used for training or just sold or other companies" [S34 · Sentiment]
Check what it tells you.
"…AI could be making things up or finding non existent sources, while staying within the sources you have ensures the accuracy more." [S20 · Sentiment]
One exception:
"…I think that ai is probably more reliable than human when it comes to favts" [S12 · Sentiment]
The environmental cost.
"We discussed that although AI can be very useful at what environment cost is it worth using it at" [S8 · Sentiment] · "…we also talked about how we often don't think about the physical costs for a single AI query." [S62 · Sentiment]
It feels unavoidable.
"Unavoidable, lots of pressure to use it to be on the same playing field (eg for recruiting)…" [S25 · Sentiment]
Week 4What students took from an industry guest
9/16 · a software engineer at Google, and CMU alum, on how AI is changing the job
Students wrote down something they learned or were surprised by. About a third of the answers were career advice; the themes below are the ones about AI and the work itself.
Responses
39
Themes
Engineers are still needed, now to manage the agents.
"It was refreshed to hear that SWE roles aren’t going anywhere!" [S34 · Guest] · "AI will not replace software engineers, but instead need them to manage these agents." [S41 · Guest] · "I'm surprised to learn how much work is involved in maintaining ai systems, ie creating separate agents, adding skills, and reviewing code." [S36 · Guest]
Soft skills count for more than expected.
"Learned and confirmed my belief that soft skills are becoming more important as AI evolves." [S63 · Guest] · "I was surprised that she focused more on soft skills and extracurriculars than hyperfocusing on technical abilities…" [S44 · Guest]
How much AI is in the daily work surprised people, in both directions.
"I was surprised by the fact that all the workflows revolve around AI." [S32 · Guest] · "There is much less LLM usage than expected in day to day." [S7 · Guest]
Describing the problem well is becoming the skill.
"…know how to use natural language to describe a problem and goal is more important, especially in a human-AI co-op environment." [S15 · Guest]
Fundamentals still matter.
"That it’s important to learn one language very strongly. Technical skills plus ability to use AI is important" [S47 · Guest]
Week 5The workflow students say they want
9/23 · during the lecture introducing spec-driven development
"Imagine that AI can potentially complete any task, though it still makes mistakes from time to time, and can't read your mind. What do you think your ideal workflow looks like?"
Responses
38
Themes
Several agents, with some checking the others.
Course context Asked right after slides describing a spec handed to a builder agent and a set of parallel reviewer agents.
"My ideal workflow would include multiple agents that could work on different parts of the task and then coordinate with each other and cross check each others work." [S57 · Workflow] · "The idea workflow would be to use ai to confirm my idea first and then have the codex or other agent to do all the work with another agent checking to make sure there’s no mistake." [S33 · Workflow]
The AI asks me questions, and checks in when it needs me.
"My ideal workflow is having AI prompt me questions or probing ideas to qualify my specifications more." [S22 · Workflow] · "Set up a loop where it keeps building and improving and fixing things, then asks me for design/feature decisions and pings me when it needs me" [S7 · Workflow] · "It’ll probably Turn it into a plan and then execute it and only ask me questions for things that are not clarified." [S59 · Workflow]
My idea and my plan first. Then the AI builds.
"Draft outline by myself. Give AI brief ideas first(the big picture). Manually check and fine tuning with specific instructions." [S41 · Workflow] · "I plan the process and tell it my in detailed what my plan is. Let it finish my plan step by step" [S11 · Workflow]
One exception:
"1 AI does all the brainstorming and planning (come up with an idea is hard, so AI should do it)…" [S45 · Workflow]
Step by step, back and forth.
"Prompt, test, fix, repeat" [S20 · Workflow] · "I think a back and forth workflow is good. Like generate something, but we talk it through issues chunk by chunk on a high level would be great" [S4 · Workflow]
Stay in the loop: understand it, check it, make it mine.
"I think ideally AI does most of the busywork and actual execution (ie. writing code) while I am mostly there to check its work and make sure it makes sense." [S21 · Workflow] · "I would try to make the outputs my own by editing it" [S34 · Workflow] · "…personally I want to use ai to support my creative process not to do it for me." [S35 · Workflow]
Failure modes they want gone.
"…when you try to prompt the ai it ping longs between two broken implementations…" [S36 · Workflow] · "…sometimes it will just claim it’s working on it and doesn’t actually change anything at the end." [S57 · Workflow]
Holding back.
"I'm pretty happy with my minimal usage of AI right now…" [S19 · Workflow] · "My ideal workflow is something that makes me feel like my information is mine alone" [S26 · Workflow]
"I think I find value in using AI to help me understand how a task is completed, so ideal workflow for me is to have multiple agents running: one to plan, one to check the plan for holes, another to implement it, another to test, etc. and along the way each of these would check in with me to make sure I understand the steps of what they completed." [S24 · Ideal workflow]
Part three · assignments in detail
Project 1Personal Portfolio Website
Weeks 1 to 2 · first major assignment · same brief as Spring
Build and deploy a professional portfolio website with About, Projects, and Contact sections. Responsive, live on GitHub Pages, AI usage documented in code comments, plus a prompt log (more explicitly required this semester). As in Spring, the writeup and the first week of lecture recommended starting in a browser chat tool; an IDE demo followed in week 2, before the deadline.
Responses
48
S26 · 42
Satisfaction
5.3/7
S26 · 5.3
Understood code
4.2/7
S26 · 4.4
Understanding sufficient
3.2/5
new this fall
No-bugs confidence
3.8/5
S26 · 3.8
Can update later
4.3/5
S26 · 4.3
Tools the class reached for
Model families found useful
Where they mostly worked
Spring 2026, one combined question: Gemini 31%, ChatGPT 26%, Claude 26%, GitHub Copilot 10%, Cursor 5%.
Themes
Show it what you mean: reference images, screenshots, wireframes, templates.
Also a theme on Spring's Project 1.
"it was useful to feed ai actual images because when i asked it to just create a flower for example, it turned out pretty bad and i had to give it a reference image." [S28 · P1] · "Make wireframes that clearly identify what you want the visuals to be" [S5 · P1] · "AI was not very good at taking a different website link in the initial prompt and actually using it for inspiration. I found I had to give ChatGPT a screenshot of what the website I was getting inspo from looked like instead." [S8 · P1]
Two limits:
"Adding reference images for creating a vibe of a website is super helpful! But AI is less good at mimicking the image itself to, for example, create icons" [S25 · P1] · "Having AI work off of someone else’s template makes it much more consistent, however be careful as it will tend to directly copy what the other person did." [S17 · P1]
Be specific, and settle the design before you build.
"Give detailed explanation of your idea. Give examples if possible" [S1 · P1] · "Try to be specific in prompts, especially regarding visual components" [S21 · P1] · "providing the exact context and specifying the layout and structure you want in your portfolio website is always useful to provide to avoid letting the AI make those assumptions for you since it tends to over do things…" [S52 · P1]
Spring quoted three students who preferred loosening the reins. This fall, one:
"…it looked very ai generated at first, but when I pointed it out and wanted it to look more immersive, it did so very well without me giving it specific ways" [S12 · P1]
Small pieces, one change at a time.
"Giving problems in bite sized pieces, or building anything new in bite sized pieces often works better than throwing a large/complex prompt at it" [S39 · P1] · "Don't type everything you want to do in one prompt" [S40 · P1] · "Make small edits at a time" [S14 · P1]
Its own design sense is weak. Give it a direction, or tune it yourself.
Not a Project 1 theme in Spring's report.
"AI's design ability was very poor, but given a clear direction and visual target, it succeeds very well at creating what you want." [S19 · P1] · "One suggestion was giving ai a clear theme so it doesn’t go for a modern look." [S3 · P1] · "…i just told it what i wanted to change and asked it to give me the corresponding variable names so i could fine tune it myself" [S35 · P1]
It adds things nobody asked for.
In Spring's report this complaint first appears on HW 6.
"It created the website well, but it kept adding strange details like a Konami Code Easter Egg, or weird random text that I wouldn’t want on my website." [S16 · P1] · "…there were a lot of random elements that the AI added to my website which I had to remove." [S32 · P1]
Use different tools for different jobs.
"Choosing differently-powered models for different purposes. For example, I found that Claude was better at generating an SVG than ChatGPT was." [S22 · P1] · "Use gpt or other ai to make a good prompt for ur coding agent." [S13 · P1]
Protect what already works.
"When prompting AI, try to first state what they did right and then ask for changes so they don’t accidentally change features already working/wanted." [S41 · P1] · "Dont use AI to modify a good code" [S11 · P1]
Have it confirm what it understood, or ask you questions, before it changes anything.
Course context On a lecture slide in both semesters. Spring's report has no theme about it.
"When a prompt is given, ask the AI to confirm its understanding before letting it actually change the code" [S45 · P1] · "…ask the A.I. To ask you clarifying questions." [S38 · P1]
Other observations.
"Something very useful I used was an extension called “live server” which I used to see the changes I did in my website in real time before deploying on GitHub." [S30 · P1]
"…sometimes it’s easier to start back at the beginning of where you were with that part and giving it a fresh prompt instead of trying to just keep editing" [S31 · P1]
"Definitely try to understand what code changes the AI made after each prompt." [S42 · P1]
"Majority of time spent should be discussing design details and really specifying everything to run into way less issues later. You can do a lot of trial and error in this phase without committing to building anything." [S7 · Project 1]
HW 2Crossy Road
Week 3 · a 30-minute in-class sprint, then finish the game at home · changed since Spring
Spring's version was a single one-hour build in Python with any tool. This fall it began with a 30-minute in-class sprint in Kiro, an AI IDE everyone had to use for that half hour. Students then had up to about four hours at home, with any tools, to finish a playable JavaScript game and publish it on their portfolio. The survey came after the finished game was due and asked about the sprint separately.
Responses
39
S26 · 41
Satisfaction
5.1/7
S26 · 4.4
Read code
3.0/7
S26 · 2.9
Understood
3.3/7
S26 · 3.6
Understanding sufficient
2.9/5
new this fall
No-bugs confidence
3.0/5
S26 · 2.2
Could build full 2.5D game
3.6/5
new this fall
Spring's ratings describe a one-hour build, so the two are not like for like.
Tools the class reached for
Model families found useful
Where they mostly worked
Spring 2026, one combined question: Gemini 44%, Claude 34%, ChatGPT 34%, Cursor 7%, GitHub Copilot 7%. This fall, Kiro counts under Copilot or Kiro.
Themes
Kiro, the required sprint tool, drew the most complaints.
New to the course, so there is no Spring comparison.
"I did not get very far in the 30-minute class exercise because my Kiro crashed and was very slow." [S8 · HW2] · "Kiro got stuck on generating a design file for several minutes." [S22 · HW2] · "Kiro uses a lot of credits for very little actual work." [S21 · HW2] · "I think the cheap models of kiro free are very limiting compared to the models I’m used to…" [S7 · HW2] · "Kiro did not provide me a good solution for the 2.5D view and was unable to solve the issues, I restarted with Claude and immediately got better results" [S17 · HW2]
Build in small pieces, mechanics before looks.
The dominant theme on Spring's HW 2.
"I would use the same incremental approach again because building the mechanics first made debugging much easier." [S42 · HW2] · "…i would start with a shorter and smaller prompt instead of feeding it all at once" [S31 · HW2] · "First make the game working and then change the art or any concepts visually" [S29 · HW2]
Two exceptions:
"…I would start with 2.5D prompt directly because 2D version does not contribute any help" [S45 · HW2] · "…if I try to change the graphics sometimes the mechanics get troubled up too." [S44 · HW2]
Planning: a time sink in a 30-minute sprint, a time saver otherwise.
Course context A pre-sprint slide warned that a one-shot prompt "may not finish its first prompt in 30 minutes."
Used up the sprint:
"I did not get very far, I thought my one prompt wouldn’t take that long, but it ended up taking the whole class and giving a mid tier output." [S16 · HW2] · "…the second prompt where I asked it to code based on the planning documents it created took the whole class and didn’t give a developed output at all." [S25 · HW2] · "i would skip the planning stage a bit in a short amount of time if given the opportunity." [S46 · HW2]
Made the build go smoothly:
"Having a full design spec doc with tech stack, requirements, full planning, etc. before you build will save you so much time." [S7 · HW2] · "In the future, I would start with more planning to decrease technical debt for future additions." [S22 · HW2]
Tell the AI to ask you questions before it starts.
Not a theme anywhere in Spring's report.
"I didn't ask the model to ask me some questions I think that caused me to make a lot of mistakes" [S26 · HW2] · "I got pretty far by including links/screenshots in my planning and then also prompting the planning bot to ask me a bunch of questions. Once it started actually building it pretty much did everything I wanted on the first try." [S23 · HW2] · "Always ask the AI to ask clarification questions if confused." [S36 · HW2]
The model matters.
"The difference between models is massive!" [S47 · HW2] · "One suggestion we observed was that the model you use really makes the difference so sometimes it’s worth it to splurge." [S3 · HW2] · "I think some applications are better at coding and other at understanding/framing your ideas." [S38 · HW2]
Be specific about what you actually need.
"be specific and precise. don't use too vague words" [S2 · HW2] · "Make your prompt detailed on things that are really needed" [S48 · HW2]
One exception:
"I think i tried to direct the ai too much, and when it attempted to match what i described, it made mistakes." [S35 · HW2]
Screenshots, links, and things that already exist.
"It helps to use screenshots! Like feed the AI screenshots if it’s a more advanced model that can interpret them." [S24 · HW2] · "Adding references/replicating existing things was a lot easier for the AI" [S25 · HW2]
What you know still matters.
"I found that I had to pull details of crossy road from my head based on my past knowledge of the game, which is why domain knowledge is still important while coding with AI." [S32 · HW2] · "I think we should understand some code to better interact with" [S1 · HW2] · "…i’d make it give me the relevant variables and i’d just change them myself" [S35 · HW2]
Other observations.
"Sending the pdf instructions directly to the Ai works really well for me." [S30 · HW2]
"Even if coded with AI, it takes a while for AI to edit the code which ended up taking more time than I expected" [S20 · HW2]
"AI isn't the most creative yet." [S10 · HW2]
"I started with Kiro and it went horribly and kept breaking and resetting. I then tried Claude code and prompted it to ask me any clarifying questions before beginning. In 25 minutes I had a perfectly working game" [S5 · HW 2]
HW 3Explore an API
Week 4 · first contact with keys, endpoints, and external data · same brief as Spring
Find a public API and build a small script, game, or web feature that does something interactive with its data. A README, a prompt log, and a short demo video are required, and keys must stay out of the repo. The time estimate grew from two to three hours in Spring to up to about six this fall.
Responses
42
S26 · 39
Satisfaction
5.2/7
S26 · 5.1
Read code
3.6/7
S26 · 3.5
Understood
3.6/7
S26 · 3.9
Understanding sufficient
3.3/5
new this fall
No-bugs confidence
3.3/5
S26 · 3.0
Could use other APIs
3.9/5
new this fall
Idea was hardest part
3.5/5
new this fall
Tools the class reached for
Model families found useful
Where they mostly worked
Spring 2026, one combined question: Gemini 51%, ChatGPT 36%, Claude 26%, GitHub Copilot 23%, Cursor 10%, Antigravity 8%. Spring's first assignment with IDE tools above 30% combined.
Themes
Be specific and clear, starting with the first prompt.
"Being more specific and direct with your first prompt will allow the progress a lot faster, which is odd because it was not this case with crossyroad" [S31 · HW3] · "Giving the ai instruction without specifics or using vague language will usually lead to the ai making decisions that are not intuitive." [S36 · HW3] · "Being able to very accurately describe errors in detail makes it very simple to fix bugs in one shot" [S7 · HW3]
Ideas were the hard part, and many looked alike.
22 of 42 agreed (4 or 5 of 5) that coming up with a good idea was the hardest part. They split on whether AI helps.
"People had similar ideas like quizzes recipe stuff etc" [S20 · HW3] · "Many peoples UI look pretty similar, but we should try to design our own since AI struggles with design features." [S19 · HW3] · "One useful suggestion was to not draft ideas with AI because they were very basic and uncreative." [S3 · HW3] · "Using the AI to brainstorm ideas really help…" [S30 · HW3]
Keys, hosting, and the steps the AI cannot do for you.
"I had to get the API key from TMDB myself and I couldn’t use AI to get it" [S32 · HW3] · "My classmate was also confused how to set up Github Pages with API key hidden." [S8 · HW3] · "…search a lot by yourself just in case there’s some mistakes that AI probably makes" [S27 · HW3]
One AI to plan or write the prompts, another to build.
A common theme on Spring's HW 3.
"Prompt using a different LLM than the one you’re using to build the project" [S34 · HW3] · "…ask ChatGPT in browser to give prompts to give to codex in VS code" [S12 · HW3]
Start from a bare-bones working version.
"Starting from the bare-bones working prototype and then building functionality around it." [S22 · HW3] · "Found it useful to build incrementally, setting up the structure first…" [S24 · HW3] · "it is helpful to break down problems and implement each feature individually" [S9 · HW3]
Give it the documentation, and read what the API returns.
A common theme on Spring's HW 3.
"Sometimes AI wasn’t able to correctly access an API and needed more prompting from API Documentation" [S42 · HW3] · "Understand the full information that the API gave you" [S26 · HW3]
Does a stronger model help? Students disagreed.
Four students on HW 2 said a stronger model made the difference.
"Using a stronger model does not necessarily improve the code it generates." [S45 · HW3] · "Using more complex models is a bit overrated when you have a solid vision of what you want to do (and a good foundation for the code being used)" [S40 · HW3] · "a simple prompt with a smarter / slower model is better than a complex prompt and or a quicker model" [S35 · HW3]
Work in the editor, where the code is visible.
"Using Claude in vs is by far my favorite way to use Claude so I can actually look at the code easily" [S23 · HW3] · "…VS code plus Claude setup hasn’t failed me, it is really good!" [S30 · HW3]
Other observations.
"I think if you have a md file that says when you get a question, ask ne before you implement, that usually does very well" [S4 · HW3]
"API calls can be unpredictable" [S37 · HW3]
"I encourage mixing models. Codex helped me make the guidelines including what API to use and what endpoints and I organized the info and put it into Claude Code." [S44 · HW 3]
HW 4Frontend + Backend
Week 5 · a server-side backend talking to a static frontend · same brief as Spring
Build a simple backend service on Render, keep any secrets there, and have a frontend on GitHub Pages call it. A README, a prompt log, and a short video are required. The time estimate grew from about three hours in Spring to up to about six this fall.
Responses
36
S26 · 37
Satisfaction
5.3/7
S26 · 4.8
Read code
3.3/7
S26 · 3.8
Understood
3.4/7
S26 · 3.6
Understanding sufficient
3.2/5
new this fall
No-bugs confidence
3.3/5
S26 · 2.8
Could build another backend
3.8/5
new this fall
Trust backend security
3.1/5
new this fall
Tools the class reached for
Model families found useful
Where they mostly worked
Spring 2026, one combined question: Gemini 51%, ChatGPT 32%, Claude 27%, Cursor 14%, GitHub Copilot 11%, Antigravity 8%.
Themes
Split the work across agents, tasks, or chats, and pass notes between them.
A strong theme on Spring's HW 4, where the student briefed each chat by hand. Two students this fall had handoff files or a prompt log do it.
"I think building it incrementally with several agents is the best way to go" [S4 · HW4] · "Making it explicitly clear separate tasks for the AI to focus on allowed it to complete them more accurately" [S25 · HW4] · "Sometimes I felt like the context was lacking between chats especially since it would forget that I already set up the back end but as long as I kept updating the prompt log I would tell it to read that so it knew where we were." [S23 · HW4]
Have one AI write the prompts for another.
"…explained everything to gpt and having it create prompts for me to tell codex and I communicate the results back and forth" [S12 · HW4] · "I tried getting AI to generate a prompt for me and it worked really well for this implementation, essentially no oversight was needed" [S17 · HW4]
Know the architecture yourself, and have the AI walk you through it.
"Having a clear understanding of the architecture of the project, for example understanding what front end and backend is supposed to do, how many moving parts the project has and what render does is helpful…" [S34 · HW4] · "Ask ai to give you a walkthrough of its plan and explain the steps it has accomplished, it will help boosting your understanding about what’s happening…" [S15 · HW4] · "I think let AI teach you how to use backends is useful" [S27 · HW4]
Deployment went better than expected for some.
Deployment trouble was a strong theme on Spring's HW 4.
Course context This fall's writeup added a section on testing locally and a caution about scope.
"Render was easier than I expected" [S26 · HW4] · "I was surprised to find Claude was pretty accurate when giving me instructions for Render…" [S24 · HW4]
"…when I first would open the Github pages the call to my backend api was really slow…" [S8 · HW4] · "I didn’t realize that you had to have credits in your OpenAI account to use the API." [S32 · HW4]
Tell it what it is building on, from the first prompt.
"One suggestion was to let AI know which service you’re using first so it can properly build around that." [S3 · HW4] · "it would be easier for the future steps if stating that this is a backend + frontend project at the beginning when you are putting the first prompt out" [S29 · HW4]
Show it the bug: screenshots and console errors.
"showing ai the actual issue in an image seems to work much better than typing the issue" [S31 · HW4] · "it’s easier to debug with ai quickly if you give it the problem but also any errors in the console, terminal, etc" [S35 · HW4]
Be detailed, and state the security rules explicitly.
"Be very specific when promoting and provide examples" [S14 · HW4] · "similar to last project, ai did pretty well on security and api conservation when told explicitly to follow instructions" [S19 · HW4]
Spec files appear.
Course context Spec-driven development was introduced in lecture on 9/23. In Spring's report, SPEC.md first appears on HW 8.
"…I tried using SPEC-driven development, and it helped to come up with categories of things I haven’t really thought of." [S22 · HW4] · "I created a spec Md this time and created new chats more often." [S23 · HW4]
Other observations.
"Have Claude control browser to audit settings was helpful for render to make sure security was good" [S6 · HW4]
"My Ai agent had a tendency to touch code that I did not ask it to, so any code I wrote would be overwritten" [S36 · HW4]
"Sometimes talking in a chat with no context is better since it can explain things better" [S16 · HW4]
"Having agents write detailed handoff files to each other was very useful to have frontend connect to backend seamlessly with separate agents" [S7 · HW 4]