Prompt Engineering

How to ask Nwiro for things so it does what you want. The tool layer is the same across both products; the AI agent driving it is what you need to steer.

General agent prompting

State the end goal, not the steps

The agent knows the 209 tools. You don’t need to tell it which tool to call. Describe what you want to end up with and it plans the steps itself.

Do:

Add a sprint feature to BP_ThirdPersonCharacter. Use an Enhanced Input action IA_Sprint bound to Shift, and change MaxWalkSpeed from 600 to 1100 while held.

Don’t:

Call read_blueprint("/Game/Characters/BP_ThirdPersonCharacter"). Then call create_input_action("IA_Sprint"). Then call edit_blueprint(...) with nodes […].

Both work. The first is shorter and lets the agent recover from mistakes.

Name assets, not paths

If you’re not sure where something lives, say the name:

Open the third-person character blueprint.

The agent uses find_assets to locate it. Only give full /Game/... paths when you have them handy and want to skip the lookup.

Paste screenshots when it helps

Ctrl+V an image into the chat (Pro) or drop it into your client (Integration Kit) when:

You’re describing a reference art style: “tree colors should look like this”.
You’re asking about something you see in the viewport or graph: “why is this connected wrong?”
You’re giving an error screenshot.

Vision-capable models read blueprint graph screenshots well.

Let iterations do the work

The agent is good at follow-ups. After a result lands:

Now do the same for BP_AICharacter.

Undo the last input action and use Q instead of Shift.

Revert that, it broke the animation bp.

Tool-level undo is limited by Unreal, but the agent is good at manually reversing its own edits.

Biome Generator prompts (Pro only)

The Biome Generator has its own pipeline: it plans layers, finds meshes by name, and produces a PCG graph. Its prompting style is different from general agent prompting: simpler, more like describing a photograph.

Less is more

Short and open produces better forests than long and specific.

Good


Dense tropical forest with palm trees
Autumn maple grove with wildflowers
Rocky alpine hillside
Overgrown abandoned garden with ivy

Bad (the pipeline respects your words literally)


Place exactly 50 palm trees with 3 meter spacing in a grid
Add SM_Palm_01 at (100, 200, 0) and SM_Palm_02 at (250, 300, 0)
Use 20% ground cover and 15% bushes

Biome keywords

Use at least one concrete biome word so the generator knows which mesh family to draw from.

Biome	Prompt hints
Alpine	”pine forest”, “mossy alpine”, “snowy slope”
Tropical	”tropical island”, “palm grove”, “jungle floor”
Forest	”autumn forest”, “birch grove”, “deciduous floor”
Grassland	”lush meadow”, “wildflower field”, “oak savanna”
Desert	”arid canyon”, “cacti plain”, “rocky desert”
Medieval	”overgrown courtyard”, “abandoned garden”, “creeping ivy”

Constraint keywords

Keyword	Effect
`around <X>`	Place elements encircling the reference
`near <X>`	Place close to reference
`avoid <X>`	Keep clear of reference
`clustered`	Group in patches
`scattered`	Spread evenly / randomly
`along the spline`	Follow the spline you’ll edit afterwards

Example combining them:

Dense pine forest along the spline. Clustered bushes near the trees. Scatter mushrooms around the fallen logs. Avoid the center clearing.

Iteration

Stay in the same chat, use natural-language deltas:


Make it denser
Add more rocks
Swap the pines for oaks
Remove the flowers near the spline
Use varied tree sizes

The generator updates the existing graph instead of rebuilding from scratch.

Common failure modes

Symptom	What to try
Generator picks wrong meshes	Add a biome keyword; check your asset naming
”Make it denser” didn’t change anything	Be more specific, like “twice as dense” or “three clusters of five”
Agent keeps asking for permission	Switch to Accept Edits mode if you trust the plan
Agent plans but doesn’t execute	Might be in Plan Mode; switch to Default
Cost feels too high	Use a smaller model (Sonnet / Haiku) for routine edits
First forest result is sparse	Your project has too few meshes. Download Demo Assets, or add more variety
Agent keeps re-asking something it just did	Chat is too long; start a fresh one and link any assets with `@`

Prompt debugging checklist

Before re-sending:

Does the message state an outcome, not a toolchain?
Are referenced assets named clearly enough for the agent to find them?
Did you include a biome word if using the Biome Generator?
Did you drop technical specs (coordinates / percentages) that lock the AI into bad choices?
Is this a case where a screenshot would explain faster than words?

Beyond the basics

Use Plan Mode to have the agent describe a proposed change before running anything. Good for reviewing refactors.
Use vision for art direction. “Style reference image: …” plus an attached screenshot does a lot.
Chain instructions with explicit “then” for multi-stage work. The agent still plans holistically but respects ordering cues.
Ask “why” when a result surprises you. The agent explains; often the explanation reveals a mistake on either side.