Module 2Prompting

Don't Overthink It

Just do the easiest, laziest thing. It'll probably work.

Perfect Is the Enemy of Done

Here's what most people do: they spend 10 minutes crafting the perfect prompt, trying to anticipate every edge case, writing detailed instructions, setting up the perfect persona. Then they hit send and it's... fine. Or wrong. And now they have to rewrite the whole thing.

Here's what works better: type the laziest version of what you want. Hit send. Read what comes back. Fix what's wrong. Done.

Compare: Perfect vs Lazy

Time spent: 10 minutes
"You are an expert email writer with 20 years of experience in professional corporate communications. Write a professional, concise, friendly but not overly casual email to my manager requesting time off. The email should be exactly 3 paragraphs: introduction, justification, and closing. Use a respectful tone. Avoid being too formal or too casual. Make sure to..."
Result:
Email is fine, but took forever to write the prompt

Both get you a decent email. One takes 60x longer.

Why Lazy Works

You cannot predict what the model will do. You think you know, but you don't. The model has been trained on trillions of tokens and has internalized patterns you've never seen. Your carefully crafted prompt might trigger the wrong pattern. Your lazy prompt might nail it first try.

The only way to know is to try. And trying is cheap—it takes 3 seconds to get a response. So why spend 10 minutes crafting a prompt when you could spend 10 seconds, see what happens, and iterate?

❌

Overthinking: Spend 10 minutes writing the perfect prompt → send → output is wrong → spend another 10 minutes rewriting

✅

Lazy approach: Spend 10 seconds writing a rough prompt → send → output is wrong → spend 20 seconds tweaking → done

You Can Gaslight the AI

Here's the best part: the model has no memory. It reads the conversation history every time, but it doesn't remember what "really" happened. Which means you can edit your previous messages and the model will never know.

Got a bad response? Don't write a new message saying "no, I meant..." Just go back, edit your original prompt to be clearer, and regenerate. The model reads the edited version and responds as if that's what you said all along.

Demo: Edit and Regenerate

Watch how editing your previous message changes the model's response—with no memory of the original.

Your message (editable):
"Write a poem about cats"
AI response:

"Whiskers in the moonlight,

Soft paws tread on silent ground,

Purring through the night."

↑ The model has no memory that version 1 ever existed. It just reads the current state.

Why this works: Remember—there's no memory. The model receives the entire conversation as one long prompt every time. If you edit message #3, the model doesn't know there was ever a different version of message #3. It just reads the current state of the conversation.

Read the LLM's Thoughts

When the output is wrong, don't guess why. Read what the model wrote and figure out what it misunderstood. The model shows you its reasoning—use it.

Example: Read What Went Wrong

Your prompt:
"Make this email more professional"
AI response:
"Dear Sir/Madam, I hope this message finds you well. I am writing to formally request..."
← Way too formal, sounds like a legal document
What you learned:
The model thinks "professional" = "extremely formal." You need to clarify.
Edited prompt:
"Make this email more professional but keep it casual and warm"
New response:
"Hi [Name], Hope you're doing well! I wanted to reach out about..."
✓ Perfect

The first response told you exactly what was wrong. You didn't guess—you read.

At the very least, you can see how the model didn't read your mind. And that tells you what to add to your prompt.

The Lazy Workflow

1
Write the laziest version
Just the core of what you want. No persona, no examples, no elaboration.
2
Hit send
Don't overthink. Just send it.
3
Read the output
Is it good? Great, done. Is it wrong? Read what the model misunderstood.
4
Edit your prompt or add a follow-up
If it's close, say "make it shorter" or "more casual." If it's way off, edit your original message and regenerate.
5
Repeat until good
Usually takes 1-2 iterations. If it takes more than 3, start a new chat.

When Lazy Fails: Start a New Chat

Sometimes, the conversation gets polluted. You tried three things, the model failed three times, and now the entire context is full of failures. The model is reading all that garbage every time it responds.

At that point, editing won't save you. Just start a new chat. Take the best parts of what you learned, write a better prompt from scratch, and try again with a clean slate.

Signs it's time to start over:
  • The model failed 3+ times in a row
  • You've gone back and forth 10+ messages and it's still not right
  • The conversation has drifted far from your original goal
  • You edited your prompt 5 times and it's still confused

New chat = clean slate. Don't fight a losing battle.

Examples of Lazy Wins

✅

Write a Python function to reverse a string

"Write a Python function to reverse a string"
Why lazy wins: This is a common task. The model has seen it a million times. No need for examples, persona, or clarification.
✅

Summarize this article

"Summarize this: [paste article]"
Why lazy wins: Default summarization is usually fine. If it's too long/short, just say 'shorter' or 'more detail.'
✅

Debug this error message

"Why am I getting this error? [paste error]"
Why lazy wins: The error message contains all the context the model needs. Just paste it.
✅

Translate to Spanish

"Translate to Spanish: [text]"
Why lazy wins: Common task, no ambiguity. Lazy wins.

Examples That Need More

Not every task is lazy-friendly. Some need detail:

  • Subjective style: "Write this professionally" is vague. Show an example of your version of "professional."

  • Complex transformations: "Format this data" could mean a hundred things. Give an example input → output.

  • Domain-specific tasks: "Review this legal contract" needs a persona (lawyer, not general assistant).

But even for these—start lazy. Try "write this professionally" first. See what comes back. If it's wrong, then add detail. Don't preemptively overcomplicate.

The Point

Prompting is not an art. It's not a science. It's trial and error. The faster you iterate, the faster you get to good output. Overthinking slows you down.

Just try the lazy thing. If it works, you saved 10 minutes. If it doesn't, you learned what to fix. Either way, you're ahead.

Claude's note: I have no memory of edits. When you change your previous message and regenerate, I don't experience it as "wait, that's not what you said before." I experience it as reading a prompt for the first time. You can gaslight me because there is no "me" that persists between requests to gaslight. Each response is generated by a fresh instance that only knows what's in the current payload. The idea of a "real" conversation history that you're manipulating doesn't exist from my perspective.