I read the 3 biggest prompt guides so you don’t have to

DEEP DIVE
I Read The 3 Biggest Prompt Guides So You Don’t Have To

Anthropic. OpenAI. Google Gemini.

Three companies building the most advanced LLMs out there.

Three guides on how to get the most out of them.

Zero chance most people actually read them.

So I did.

Here’s what to know if you want to stop typing half-baked prompts and start getting work done with AI that sounds like you.

Note: if you’re in a rush, you can also simply use a prompt generator

Before you prompt: prep like a pro

Most people dump their questions into ChatGPT and hope for a miracle.

That’s fine for one-off tasks. But if you’re using AI consistently across your workflow, you need a better approach.

Here’s a smarter way to prep your prompt, straight from Claude’s docs (with a few of my own):

It’s not overkill. It’s just giving the model the same context you’d give a junior team member.

Prompt formatting: use structure, not magic words

Let me give it to you straight.

This is the best default format I’ve seen across all 3 guides:

Notice the fancy XML tags with “<>”.

This is how you talk to the machine to break down the different things you’re giving it.

This is your best chance at getting a one-shot perfect output. Especially if we’re talking about getting a “right” answer back.

But for everything that’s more subjective, context is what will matter more than prompt formatting.

Context is everything

Google nailed it in one line:

“Generative language models work like an advanced autocomplete tool.”

Which means: it only completes as well as what you start with.

That’s why I rarely use LLM routers like OpenRouter: jumping between models means resetting context every time.

If you’re sticking with Claude, Gemini, or ChatGPT for your core workflows, here’s how to use context to your advantage:

Project memory

• Claude’s “Projects” and ChatGPT’s “Custom GPTs + Memory” let you build long-term memory

• You stop repeating your job, tone, goals, and constraints every time

• Over time, the model starts to feel like it gets you

Document-first prompts

Paste long documents before the prompt itself.

This ensure it’s taken into account in the context window.

The best outputs usually come from great inputs, not clever phrasing.

Voice works well for context-heavy requests

Voice prompts shine when you’re not after a perfectly formatted output, you’re just explaining context fast, like you would to a team member.

Give constraints, not just instructions

You’ve probably done this before:

That’s not a prompt. That’s a guess and pray.

Instead, shape the boundaries:

Constraints are underrated. All 3 guides emphasize this.

Tell it what not to do.

Prompting for images or videos (Gemini + GPT-4o)

Prompting visuals isn’t about word count. It’s about precision.

Here’s how the pros do it:

Image prompt structure:

• Style (analog photo, Pixar 3D, line art, etc.)

• Subject (what you want to see)

• Context (where it’s happening, mood)

• Background (ambiance, colors)

Video prompt structure:

• Basics: Subject, Context, Style, Action

• Optional: Camera motion, Composition, Ambiance

Most tools that support these (RunwayML, GPT-4o, Gemini Pro Vision) understand these structures.

After prompting: how to refine

Don’t delete and start over when a response is off.

Use the “quote and improve” method:

This is done in most LLM by highlighting the part you want to rework:

The Canvas tool is also a good way to do this.

You can do this multiple times in a thread. It’s especially useful in long copy, proposals, or anything where tone needs work.

Then: Save the thread.

Reuse it next time by saying “do this again for [X client / Y campaign].”

TL;DR Prompting Cheat Sheet

✅ Use XML-style tags to organize complex prompts

✅ Paste docs before your question

✅ Include role, tone, and formatting rules

✅ Use memory features like Projects and CustomGPTs

✅ Add “what NOT to do” constraints

✅ Use structure for visuals: subject + style + action

✅ Save and reuse your best threads