Lesson 5 of 5 · Free to read
Mini-project: a prompt library you will actually use
Mini-project: a prompt library you will actually use
Audience: Final-year students and engineers in their first three years
Prerequisites: Lessons 1–4
Time: about 60 minutes
What you will keep: a five-prompt library for one task you do every week, tested and documented, ready to show in an interview
This lesson does not cover: sharing prompts through a company tool.
Why this matters
Most people's best prompts live in old chat history, which means they write them from scratch every time. A small, tested library turns your best prompts into tools. It also makes a good interview story: "Here is how I use AI on code reviews, here is how I checked it, and here is what I changed when it failed."
The trap is building an impressive library of prompts you never run again. So this project starts from a task you actually repeat.
Step 1 · Pick one weekly task (5 minutes)
Choose something you do at least once a week, for example:
- reviewing a pull request
- turning a ticket into tests
- writing a commit message or release note
- debugging from a stack trace
- writing a README or API doc for something you built
Write one sentence: "Every week I ___, and it takes me about ___ minutes."
Step 2 · Write five prompts (20 minutes)
Create a folder prompts/<task>/ with one Markdown file per prompt. Use the patterns from module 02 and the parts from lesson 1. A typical set of five:
- Specify: the main prompt that does the task.
- Few-shot variant: the same task with two examples of good output (lesson 2).
- Constrain: a follow-up for when the answer goes too broad.
- Critique: a review prompt that checks the main prompt's output.
- Chain or compose: a short chain that runs the task in steps (lesson 3).
Each file uses the same template:
# <name>
**Use when:** <one sentence>
**Pattern:** Specify / Iterate / Constrain / Critique / chain
**Inputs:** {diff}, {ticket}, ... (the parts you paste in each time)
**Prompt:**
<the prompt, with {placeholders}>
**Output format:** <what a good answer looks like>
**Check:** <how you verify it: run tests, checker script, read against the ticket>
**Last tested:** <date> on <what input>
Step 3 · Test on real inputs (25 minutes)
Run each prompt on at least two real inputs from your recent work. For any prompt whose output you would not use as it is, change the prompt and run it again. Record changes in a CHANGELOG.md: what failed, what you changed, and whether it worked.
Use the checks from lesson 4. If a prompt returns structured output, run your checker on it. If it returns code, run the code.
Finish when:
- all five prompts have been run on two real inputs each;
- at least two prompts were changed after a failure, and the changelog says why;
- every prompt has a Check line that someone else could follow.
Step 4 · Write the README (10 minutes)
Add prompts/<task>/README.md with:
- which prompt to start with, and when to reach for the others;
- the time the task takes now compared with your estimate in step 1, if you measured it;
- one honest limitation: a case where the library does not help and you do the work yourself.
That limitation line matters. It shows you know where the tool stops, which is the judgement this whole module is about.
Going further
Swap libraries with a classmate or colleague who does a similar task. Run one of their prompts on your input and note what you would borrow.
Key points
- Build the library around a task you really repeat.
- Every prompt carries a check and a last-tested date.
- Keep a changelog of failures and fixes; that is where the learning is.