Here's the good news: in Layer, you don't need to be a "prompt engineer." When you describe what you want, Layer's Agent automatically rewrites your request into a detailed, model-ready prompt — adding the composition, lighting, materials, and technical settings that get great results. Your job is simpler: tell it clearly what you want, and steer as you go.
The Agent does the heavy lifting
When you send a request, the Agent expands and adapts it before generating. In the Generate your first asset guide, we typed:
A glowing magic potion bottle game icon, fantasy style, vibrant blue liquid, cork stopper, centered on a plain background
…and Layer actually generated from its enriched version:
Game icon: a glowing magic potion bottle, fantasy style. Vibrant blue liquid inside a clear glass bottle with subtle refraction, cork stopper, faint runes on the glass, soft bloom glow, clean centered composition, plain neutral background, crisp silhouette, no text, no watermark.
You can always see the exact prompt Layer used by opening an asset and checking its details.
Behind the scenes, the Agent automatically:
Expands sparse requests with concrete, visible detail
Adapts the wording to the specific model you're generating with
Moves size and format into settings, not the prompt text
Handles references and edits correctly (more below)
Keeps results production-ready — clean backgrounds, no stray text or watermarks
Keeps you IP-safe — it describes a look by genre and visual language, never by naming other studios' games
💡 Because the Agent handles the craft, the single most useful thing you can do is describe the outcome you want — not perfect the prompt wording.
Focus your effort here
A few things genuinely move the needle:
Say what it's for. "A health-potion icon for a fantasy match-3 game" beats "a potion." Telling the Agent the purpose lets it choose the right style, framing, and level of detail.
Name what matters — and your constraints. Be concrete about subject and style, and give constraints a reason: "isolated on a solid background so I can cut it out" is far clearer than "clean background."
Iterate in the chat. You rarely nail it in one shot — and you don't have to. Reply with "make the glow warmer," "give me three variations," or "try a flatter, more iconic style," and the Agent refines in place.
Show, don't describe. To match a look, attach a reference image or @-mention a reference set instead of describing a style in words. For a consistent character across many assets, a reference set keeps everything on-model.
For edits, describe the change. When tweaking an existing asset, say what to change and what to keep — "make the armor gold, keep everything else the same" — rather than re-describing the whole image.
Set your context once. Style guides, brand rules, and preferred formats set on your workspace or project are applied automatically, so you don't repeat them every time.
Ask the Agent for help. Not sure how to phrase something? Ask "make this prompt more specific" or "suggest a stronger prompt for this" — improving prompts is something it's built to do.
You can skip these
Old prompt "tricks" don't help in Layer — the Agent handles or ignores them:
Format codes in the prompt ("1:1," "4K," "16:9") — set the size in the generation settings instead
Filler words ("8K," "masterpiece," "ultra-detailed," "trending on…") — they add noise, not quality
Camera/photography jargon (ISO, f-stops) — modern models don't need it
Model-specific syntax — the Agent translates for whichever model you're using
Re-describing an image you've attached — the Agent already sees it
In short
Describe what you want and why, attach references when the look matters, and iterate in conversation. Layer's Agent turns that into the detailed prompt — so you can spend your energy on the creative call, not the keyword soup.
What's next
Stay consistent with reference sets → [Using reference sets]
Make your first thing → [Generate your first asset]
Pick the right engine → [Choosing a model]