Lesson 5 of 6 · Free to read
Lab: fix a broken API with AI as your pair
Lab: fix a broken API with AI as your pair
Audience: Final-year students and engineers in their first three years
Prerequisites: Lessons 1–4. Python 3.10 or newer (tested on 3.13), and any AI coding tool.
Time: about 75 minutes
What you will keep: a small API with all tests passing, and aPROMPTS.mdwith at least six entries
This lesson does not cover: deploying the API.
The task
Below is a small task-tracker API with four bugs. Six tests describe the correct behaviour; four of them fail. Your job is to get all six passing with an AI tool as your pair-programmer, and to log the prompts that mattered.
The rule for this lab: read the failing test before you prompt. People who diagnose first usually finish faster than people who paste the error and hope.
Set up (10 minutes)
Create a folder called task-api with this layout:
task-api/
requirements.txt
app/__init__.py (empty file)
app/main.py
tests/__init__.py (empty file)
tests/test_tasks.py
requirements.txt (these versions were tested together):
fastapi==0.162.0
pydantic==2.14.1
pytest==9.1.1
httpx==0.28.1
app/main.py:
from datetime import date
from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
app = FastAPI()
TASKS: list[dict] = []
class TaskIn(BaseModel):
title: str
due: date
done: bool = False
@app.post("/tasks", status_code=201)
def create_task(task: TaskIn):
row = {"id": len(TASKS) + 1, **task.model_dump()}
TASKS.append(row)
return row
@app.get("/tasks")
def list_tasks(offset: int = 0, limit: int = 10):
return TASKS[offset : offset + limit + 1]
@app.get("/tasks/overdue")
def overdue(today: date):
return [t for t in TASKS if not t["done"] and t["due"] > today]
@app.get("/tasks/stats")
def stats():
done = sum(1 for t in TASKS if t["done"])
return {"total": len(TASKS), "percent_done": round(100 * done / len(TASKS))}
@app.get("/tasks/{task_id}")
def get_task(task_id: int):
for t in TASKS:
if t["id"] == task_id:
return t
raise HTTPException(status_code=404, detail="Task not found")
tests/test_tasks.py:
import pytest
from fastapi.testclient import TestClient
from app.main import TASKS, app
client = TestClient(app)
@pytest.fixture(autouse=True)
def empty_store():
TASKS.clear()
def add(title="Write tests", due="2026-11-01", done=False):
return client.post("/tasks", json={"title": title, "due": due, "done": done})
def test_create_and_fetch():
assert add().status_code == 201
assert client.get("/tasks/1").json()["title"] == "Write tests"
def test_empty_title_is_rejected():
assert add(title="").status_code == 422
def test_page_size_is_respected():
for i in range(5):
add(title=f"Task {i}")
assert len(client.get("/tasks?offset=0&limit=2").json()) == 2
def test_overdue_means_due_before_today():
add(title="Old", due="2026-10-01")
add(title="Future", due="2026-12-01")
titles = [t["title"] for t in client.get("/tasks/overdue?today=2026-10-27").json()]
assert titles == ["Old"]
def test_stats_with_no_tasks():
assert client.get("/tasks/stats").json() == {"total": 0, "percent_done": 0}
def test_missing_task_is_404():
assert client.get("/tasks/99").status_code == 404
Install and run the tests from inside task-api:
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -r requirements.txt
python -m pytest -q
You should see 4 failed, 2 passed. A deprecation warning from the test client may also appear; it does not affect this lab.
Work the bugs (50 minutes)
For each failing test:
- Read the test and the error. Write your own guess at the cause in
PROMPTS.mdbefore prompting. - Prompt with a pattern from lesson 2. Start with Specify: the test name, the error, and "list likely causes, no fix yet".
- Ask for the smallest fix, limited to the function involved (Constrain).
- Run the tests yourself. Log the result and your decision.
- If the AI changed anything else, reject that part.
Watch for the failure modes from lesson 3. If the AI suggests a pydantic option you have not seen, check that it exists in pydantic 2 before accepting it.
Definition of done
-
python -m pytest -qshows 6 passed. -
PROMPTS.mdhas at least six entries, each with a Decision line. - Each fix is the smallest change that makes its test pass, with no unrelated edits.
- You can explain every changed line without looking at the chat.
- One entry says where you chose not to use AI, and why.
Stretch
Add DELETE /tasks/{task_id} that returns 204, or 404 for a missing id. Write the two tests first, then use one Specify prompt and see how close the first answer gets.
Key points
- Diagnose first, then prompt.
- One bug, one small fix, one test run.
- The log is part of the deliverable, not an extra.