Course A · Module A2 Prompting · Lesson 5 of 17

Your Prompt Library

Stop rewriting the same prompt from memory. Save it once as a template, and every future run takes thirty seconds.

~10 min · Have a notes app or doc open · Prereq: Lesson A3 (your chosen task)

You've already written a prompt that works. The problem is you'll write it differently next time — slightly shorter, a different instruction — and get a worse result. A prompt library solves that: you store the prompt that worked, mark the parts that change each time with [BRACKETS], and paste it forever. That's the single win from this lesson: one recurring task turns into a fill-in-the-blank template you never rebuild from scratch.

Why templates beat memory every time

Think of a great recipe card. You don't rethink the whole dish each time — you follow the same card and swap the variable ingredient for what's in season. A prompt template works exactly the same way: the fixed instructions stay in the card; only the [BRACKETED] slot changes per run.

Bonus: when a template stops working, you edit one place. Every future run improves instantly — no hunting through chat history to remember how you phrased it.

The bracket method

Write your best prompt once. Then look at every word that changes from one use to the next — a client name, a topic, a date, a tone — and wrap it in [BRACKETS]. Those become your fill-in slots. Everything outside the brackets is locked, tested instruction text.

# Template: Weekly project-status email

You are a clear, professional assistant.

Write a brief status email for a client.

Project: [PROJECT NAME]
This week: [2–3 bullet points of what was completed]
Next week: [1–2 planned steps]
Blockers: [any issues, or "none"]
Tone: [professional / warm / brief]

Keep it under 150 words. Use the client's first name: [CLIENT FIRST NAME].

Every [BRACKET] is a decision made once. The surrounding instructions — word count, tone, structure — never change. That's what makes the output consistent.

Where to keep your library

Works well

A Google Doc or Notion page titled "Prompt Library"
One section per task type (email, social, research, client)
Copy-paste directly from the doc into ChatGPT / Claude
Version notes: "v2 — added tone slot, works better"

Watch out for

Keeping prompts only in chat history (you can't find them)
No version notes (you won't know which one worked)
One giant unsorted doc (add headers from day one)
Sharing a doc with credentials pasted into prompts

Build one now — for the task you picked in A3

Go back to the automation task you chose in Lesson 3. Write your best current prompt for it. Then do three things:

Turn it into a template:

Write the full prompt as you'd send it today, with real example data.
Bracket the variables — circle every word that changes run to run, replace with [DESCRIPTIVE LABEL].
Test it — fill in the brackets with fresh data and run it. Does the output match what you got before?
Save it — paste it into your library doc with a name and today's date.

Naming convention that saves you later: put the task type first — EMAIL — invoice reminder v1, SOCIAL — product launch v2. Sorting alphabetically groups all your email prompts together, all your social ones together. Simple and effective from day one.

Check yourself

Answer before opening. Recalling it now is what makes it stick.

What do the [BRACKETS] in a prompt template represent?

The parts that change each time. Brackets mark the variable slots — a client name, a topic, a date — so the stable instruction text around them never needs rewriting.

You get a great result with a prompt. Two weeks later you can't reproduce it. What went wrong?

You didn't save the exact prompt. Small wording changes produce different outputs. The fix: copy the winning prompt straight into your library doc before closing the chat.

Your "draft cold outreach email" template starts producing generic results. What's the fastest fix?

Add more specific context slots. Bracket out things like [ONE SPECIFIC REASON WE'RE REACHING OUT] and [RECIPIENT'S INDUSTRY] so you're forced to supply the details that make output feel personal. Generic prompt in, generic output out.