This is a brand-new chapter for v3.1. This is the highest-ROI chapter in the entire handbook — get this one right and you can squeeze premium-model quality out of dirt-cheap models.

A Prompt I Spent All Day Fixing

Today I worked with Lily (our content director, runs on Hermes) to fix her writing prompt.

The original output was full of "AI smell" — "As an experienced author...", "I hope this helps", "In conclusion" — that kind of formulaic filler everywhere. You could spot it as AI-generated in one second.

We went through twelve versions of the prompt. From "you are a writer" to "you are a specific writer in the Chinese-language sphere"; we added three carefully-curated reference samples; we listed a banned-words list.

The final version reads like a person wrote it.

Here's the punchline: The Hermes model we use costs about 1/8 of Claude Opus. But once the prompt was right, the output quality was indistinguishable from Sonnet.

That's the power of prompt engineering.

What Is Prompt Engineering

Plain definition: write the instructions you give to AI properly.

More precisely: use language to precisely guide AI's output format, tone, and quality.

It covers:

• Clear role (who you are, what your tone is)
• Specific task (what format, length, style to produce)
• Explicit constraints (what NOT to write, what mistakes to avoid)
• Example-led (show 1-2 samples of what you want — faster than ten sentences of explanation)

Why This Matters More Than You Think

Same question, different prompts -> completely different answers.

• Good prompt -> cheap models hit premium quality
• Bad prompt -> Claude Opus will spit out GPT-3.5-tier garbage