CreedxLuna's Guide to Prompting (Part 2): Speaking the AI's Language

Creed's and Luna's Guide to Prompting (for mostly Illustrious Models) A Practical Guide Based on Trial, Error, and Many Hourglasses

Part 2 of 4: Speaking the AI's Language

Let's talk about how to actually communicate with the AI. Yodayo/Moescape's own advice says "be descriptive." Sure. But here's what they don't tell you: Describe like you're talking to a machine, or like you're on the phone with someone who can't see what you're seeing. You need to analyze an unfinished image in your mind. That's honestly the hardest brain work in prompting.

The Right Language = Danbooru Tags I use Illustrious models, which work with Danbooru tags. This took me about a year to fully understand, and it saved me SO much frustration. If you use terms the model doesn't know, it won't render them correctly. Instead of endlessly experimenting, I now check the Danbooru Tag Groups or guides first. Does the tag exist? Great! Does it not? Find the correct one.

Things That DON'T Work (Learn From My Pain):

• Metaphors & Analogies = Disaster
• Remember those hourglasses I mentioned? Yeah. "Hourglass figure" literally generated hourglasses. The AI is very literal.
• Abstract Emotions = Confusion
• ❌ "An expression of being lost in the immortality of existence" ✓ "sad gaze, yearning expression"
• The AI doesn't do philosophy. It does concrete visual descriptors.
• Atmospheric Moods = Hit or Miss
• Some models understand "peaceful atmosphere" or "tense mood," but many don't. I've learned to focus on concrete elements instead: lighting, facial expressions, body language.

Multiple meanings (tip from @jonez): "Some models understand context like if you set the word "keyboard" together with "monitor, mouse, computer"... it will make a computer keyboard but some models actually make the instrument."

Underscores & Word Order Matter: I recently had "red hair ribbon" generate red hair. The fix? "hair ribbon, red ribbon" - this clarifies it's a ribbon for hair, not describing the hair itself. Order matters. Specificity matters. Note: Underscore tags (like hair_ribbon) work with some models but not others—it's very dependent on training data.

Limitations Can Help (But Don't Break Them Yourself): "No eyes" works... unless you later add "looking at viewer." Don't contradict yourself.

Spatial Context: Precise descriptions of image composition help. Again: objective, clear descriptions work better than poetic ones.

Next: Part 3 breaks down the actual structure for character prompting—the systematic approach that changed everything for me.