I finally realized why AI keeps giving people mediocre content.

They’re briefing it like a stressed-out group chat with one vague sentence, three missing details, and a follow-up correction 14 seconds later.

I’ve been deep in Anthropic’s prompting best practices lately, and the biggest revelation is this:

Anthropic recommends structuring prompts the same way you’d brief an employee.

Because you know I love a specific example, here’s what that actually looks like in practice.

Let’s say I want Claude to help me come up with a social video concept about BookTok recommendations I hated.

A weak prompt would look like this:

“Give me a viral TikTok idea about books I hated.”

A stronger prompt would break everything into sections.

1. Task context

Tell Claude what role it’s playing and the job you want done.

Example:

“You are a social media strategist helping me create high-engagement video concepts for YouTube, Instagram, and TikTok.”

This immediately changes the quality of the output because now Claude understands the assignment.

2. Tone context

People skip this constantly and then wonder why the output sounds robotic.

Example:

“Tone should feel conversational, opinionated, and funny. Avoid corporate language.”

Now the AI knows not to sound like a LinkedIn ghostwriter from 2017.

3. Background data, docs, and examples

Anthropic recommends stuffing in relevant context.

Example:

“My audience is female readers. They engage most with controversial opinions and personal stories. I post book content that mixes humor, critique, and internet culture.”

That context matters A LOT.

4. Detailed instructions and rules

This is where you get painfully specific.

Example:

“Give me 10 TikTok concepts with strong hooks in the first 3 seconds. Prioritize curiosity and debate over positivity. Avoid generic ‘top 5 books’ formats.”

Specificity is the whole game.

5. Examples

Anthropic repeatedly says examples are more powerful than endless explanations. So instead of describing your style forever, show it.

Example:

“Here are hooks I’ve used before:

• ‘BookTok owes me financial compensation for recommending this.’
• ‘I can’t believe thousands of people cried over THIS book.’
• ‘This popular book felt like a 400-page HR meeting.’”

The AI instantly starts pattern-matching.

6. Conversation history

This one is underrated. Claude performs better when it remembers what you already rejected or liked.

Example:

“Previously, I said I wanted fewer generic recommendation formats and more hot-take energy.”

Now it builds on prior direction instead of resetting every time.

7. Immediate request

This is the actual thing you need right now.

Example:

“Generate 5 TikTok scripts specifically about overhyped BookTok recommendations that disappointed me.”

It's simple, clear, and direct.

8. “Think step by step.”

Anthropic explicitly recommends this for harder tasks. Which feels funny at first until you realize… it works.

Example:

“Think step by step about why BookTok outrage content performs well before generating ideas.”

That extra reasoning step often produces way stronger concepts.

9. Output formatting

Another wildly underrated one. If you don’t specify a format, AI defaults to chaos.

Example:

“Format each idea with:

• Hook
• Video concept
• Suggested caption
• Why it would drive engagement”

Now the output is usable immediately.

10. Prefilled response

Anthropic recommends STARTING the assistant’s response for it.

Example:

“Hook:
‘I need to publicly apologize to my wallet for trusting BookTok again…’”

Claude will continue from there in the same rhythm and tone.

The more I learn about prompting, the less it feels like “AI magic” and the more it feels like communication skills.

Clear in = better out.

Which may explain why so many people think AI is bad at writing when really they’re giving it the equivalent of: “IDK, just make it sound cool, lol.”

Thanks for coming to class today.

Professor Christina

PS - Now I feel like I need to make this BookTok video.