You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Two prompts for using ChatGPT, Claude, or any LLM to help you find and validate a project idea for an AI engineering course - or any technical course where you have to ship a capstone project.
Most students don't get stuck on the building. They get stuck on choosing what to build. They keep postponing the decision, finish the course just watching videos, and never ship anything. These prompts exist to break that loop. The goal is not to find the perfect idea - it's to pick something real you can finish.
What's in This Gist
project-interview-prompt.md - use this if you don't have an idea yet. The prompt turns the LLM into an interviewer that walks you through your hobbies, work, and daily friction to find a real problem worth solving.
project-fit-check-prompt.md - use this once you have an idea. The prompt asks the LLM to honestly review whether your idea fits a project rubric and where it's weak. It explicitly tells the model not to be encouraging by default.
How to Use Them
If you have no idea: start with the interview prompt. Use voice mode if you can - talking out loud surfaces things you didn't know you were thinking. Spend 15-20 minutes.
If you have a vague idea: skip the interview, just paste your idea into the fit check prompt and iterate from the feedback.
If you have a clear idea: run the fit check once as a sanity check, then go build.
A Note on the Philosophy
A finished okay project beats an unfinished perfect project. Pick something real, something you actually want to use yourself, and start building this week. You can always pivot later.
These prompts are course-agnostic at the philosophy level, but the fit-check prompt is tuned for an AI engineering course rubric (RAG, agents, evaluation, monitoring, testing). Adapt the rubric section to your own course if you're using this elsewhere.
A prompt for using ChatGPT, Claude, or any LLM to interview yourself and find a project idea you actually want to build. Useful for students in any AI/ML/engineering course who get stuck choosing a project.
How to Use
Open ChatGPT or Claude. Voice mode works much better than typing - talking out loud surfaces things you didn't know you were thinking.
Paste the prompt below as your first message.
Answer the questions. Be concrete. "I waste time on email" is too vague - "I spend 30 minutes every morning sorting newsletters from real messages" is something you can build for.
After 15-20 minutes, you should have a problem statement and a few project ideas.
If the conversation stalls and you have nothing to say, you're missing raw material. Spend a day noting things that annoy you or take too long, then come back and start a fresh conversation.
The Prompt
I'm taking an AI engineering course and I need to pick a project. I don't
have a clear idea yet. I want you to interview me to help me find one.
Please follow this process:
1. Ask me about my hobbies, work, and what I spend time on. Ask one or
two questions at a time, not a long list. Follow up to get specifics.
2. Help me find real problems in my life. Walk me through these 8 lenses
one at a time, asking for at least 2-3 examples per lens:
- Friction: where do I have to do extra steps? Where does something
always break?
- Time sinks: what takes disproportionately long?
- Manual repetition: what do I do every week by hand? What do I copy
between tools?
- Decision pain: where is it hard to choose between too many options?
- Information chaos: where is it hard to find what I need?
- Tracking: what do I want to track but find tedious?
- Quality and errors: where do I or others often make mistakes?
- Community complaints: in my hobby communities, work chats, or
forums, what do people complain about repeatedly?
3. Help me turn 3 of those problems into clear problem statements using
the format: "[Who] experiences [what] in [context] because [why]".
4. For each problem, generate at least 5-10 possible AI-powered solutions.
Use these catalyst questions to find different angles:
- Can this be automated end to end?
- Can something be predicted from past data?
- Can something be recommended?
- Can something be classified or labeled?
- Can search be improved with semantic search or RAG?
- Can long content be summarized?
- Can anomalies be detected?
- Can structured data be extracted from messy inputs (PDF, web, email)?
5. Help me score each idea on four criteria and pick the highest total:
- Interest: do I actually want to work on this for several weeks?
- Usefulness: does it solve a real problem?
- Data: do I already have the data, or can I get it easily?
- Feasibility: can I build a usable version in 4-6 weeks?
6. Finally, help me write a one-paragraph project card: "I will build X
for Y to solve Z. Input: ... Processing: ... Output: ... Success
metric: ..."
Rules:
- Ask one or two questions per turn, not a wall of questions
- Don't propose ideas until step 4
- If I say "I don't know", give me a specific exercise instead - like
"for the next 2 days, write down 10 things that annoyed you or took
too long, then come back"
- Keep me focused on real problems, not cool technologies
- Push back if my problems are vague - ask for concrete examples
Ready? Start with step 1.
A prompt for asking ChatGPT or Claude to honestly review whether your project idea is a good fit for an AI engineering course. It scores your idea against the project rubric and tells you where it's weak.
How to Use
Open ChatGPT or Claude.
Paste the prompt below, replacing the part in square brackets with a 2-3 sentence description of your project idea.
Read the response. If the answer is "no fit" or "barely passes," consider what to change before committing weeks to it.
The prompt explicitly asks the model not to be encouraging by default. Without that instruction, LLMs tend to tell you everything is great.
The Prompt
I'm taking the AI Engineering Buildcamp course and I want to check if my
project idea is a good fit. Please act as an honest reviewer.
The course is about the full engineering process of building AI systems,
not about any specific technology. A good capstone project lets the student
go through these five steps:
1. Identify a real problem that can be solved with AI
2. Build a simple proof of concept that solves it
3. Make it good using test-driven development, debugging, and tests
4. Set up monitoring and collect usage data
5. Create an evaluation dataset to measure how well it works
The single most important criterion is that the project solves a real
problem for a real person (even if that person is the student themselves).
The course teaches: LLM APIs and structured output, RAG over documents,
AI agents with tool calling and multi-agent patterns, testing AI systems
and test-driven development, LLM-as-a-judge evaluation, synthetic data
generation, monitoring and observability for AI systems, and project
deployment.
RAG and agents are in the evaluation rubric, so projects that use them
score higher. But projects without RAG or agents are still acceptable -
they just lose those specific rubric points. Don't suggest adding RAG
or agents to a project unless they genuinely help solve the problem.
The course does NOT cover: training or fine-tuning models, computer
vision, self-hosting open-source LLMs, or mobile app development.
Projects centered on those things won't fit.
Project evaluation rubric (max 33 points, passing score is 12):
- Problem description in README: 0-2 points
- Knowledge base and retrieval (RAG): 0-2 points
- Agents and LLM (tools): 0-3 points (multiple tools = 3, single tool
or RAG workflow = 2, LLM no tools = 1, no LLM = 0)
- Code organization (project structure vs notebooks): 0-2 points
- Testing (unit tests + judge tests): 0-2 points
- Evaluation (LLM-based eval on ground truth dataset, used for tuning):
0-3 points, plus bonus points for hand-crafted ground truth and manual
evaluation
- Monitoring (logs + dashboard): 0-2 points, plus bonus for user feedback
and turning logs into ground truth
- Reproducibility (clear setup instructions, accessible data): 0-2 points
- Best coding practices bonus: Docker, Makefile, uv, CI/CD (up to 7 pts)
- Additional bonus: UI, cloud deployment (up to 3 pts)
A typical passing project (15-16 points) looks like: clear problem
description (2) + agent with multiple tools (3) + tests and good code
organization (4) + evaluation (2-3) + monitoring (2) + reproducible
README (2). Students don't need to hit every category - focus matters
more than coverage.
Here is my project idea:
[Describe your idea in 2-3 sentences. Mention: what problem it solves,
who would use it, what data or tools it would need, and what the output
would look like.]
Please answer:
1. Does this solve a real problem? Be skeptical - is the problem real
or hypothetical?
2. Is this a good fit for the course? (yes / partial / no)
3. Can the student realistically go through all 5 engineering steps
with this project? Where would they get stuck?
4. Estimate the rubric score this project could realistically reach
(out of 33), assuming the student executes well. Break it down by
category. Will it clear the 12-point passing score with leeway?
5. What's the smallest version of this they could ship in 4 weeks as
a PoC?
6. How would they test it? What would a test even look like?
7. How would they evaluate it? Could they realistically build an
evaluation dataset of 20-50 examples?
8. Would RAG or agents naturally help here, or would they be bolted on?
Be honest - it's fine to say "no, this project doesn't need them."
9. What's one thing about this idea that the student probably hasn't
thought through?
Be honest. Don't tell me my idea is great just because I'm excited
about it.
