Procedural Levels That Don’t Feel Like Oatmeal: A Designer’s Playbook

Procedural generation has a public-relations problem. It promises infinite worlds, then sometimes serves the player twenty beige corridors and a fern rotated to 37 degrees. Technically different? Yes. Spiritually, the same bowl of oatmeal.

The cure is not “more random.” Memorable levels come from decisions: where the player feels safe, what they glimpse before they can reach it, when a route loops back, and why one room breaks the pattern. A good generator automates those decisions without surrendering the designer’s point of view.

Randomness is an ingredient, not a director

Noise is excellent at variation. It can scatter trees, choose room modules, nudge dimensions, or change loot. It has no opinion about anticipation, contrast, or whether the player has been walking through identical combat spaces for twelve minutes.

The foundational Procedural Content Generation in Games textbook treats PCG as a family of methods—constructive algorithms, grammars, search, planning, experience-driven systems, and mixed-initiative tools—not as a single “randomize” switch. That distinction matters. Pick the method that matches the design promise.

• Use noise and sampling for texture, density, and natural scatter.
• Use grammars for structures with relationships: a lock needs a key route; a balcony needs something worth seeing.
• Use search or rejection when a layout must satisfy measurable constraints.
• Use authored modules when composition, cover, sightlines, or traversal must be dependable.

Most enjoyable generators are hybrids. Humans author the vocabulary and dramatic rules; software recombines and tests them.

Design in three scales

Flat generators produce flat experiences. If every decision happens at “place a room,” the world has no larger rhythm and no smaller texture. Split the system into three scales.

Macro: the journey

The macro layer decides the experience graph: start, goal, optional branch, gate, reveal, recovery point, climax. It should answer “what changes during this run?” before it worries about wall panels.

A compact dungeon might follow this shape:

1. Teach one threat in a forgiving space.
2. Show the landmark that contains the goal.
3. Offer a risky shortcut and a safer long route.
4. Place the key or capability off the direct path.
5. Loop the player back to a recognizable hub.
6. Recombine earlier threats in the climax.

That is authored pacing even if the actual rooms change every run.

Meso: rooms and transitions

The middle layer assigns roles such as arena, overlook, connector, puzzle, resource pocket, secret, or rest room. Give each role constraints. An overlook needs a long sightline; an arena needs navigable combat space; a rest room should not spawn beside another rest room.

This is where a module-based “maze plus rooms” approach becomes powerful. In his GDC talk on Eldritch, David Pittman describes a process influenced by the maze-and-module logic associated with Spelunky. The lesson is not to copy a particular grid. It is to separate connectivity from the authored chunks that make each space playable.

Micro: dressing and surface variation

Only after the route works should the generator scatter grass, rubble, decals, props, and material variants. Modern tools make this layer inviting because it produces instant visual payoff. Unreal’s current PCG framework, for example, moves spatial data through a graph, attaches attributes such as density and seed to points, and provides node-by-node debug views.

That machinery is excellent, but one million well-scattered pebbles cannot repair a meaningless journey.

Give the generator a dramatic budget

When every room may become spectacular, spectacular becomes the wallpaper. Assign budgets per run:

• one dominant landmark;
• two or three strong vista moments;
• a limited number of high-intensity encounters;
• at least one quiet stretch after a major challenge;
• one controlled rule-break—the flooded library, upside-down chamber, or room with no floor.

The numbers depend on your game, but scarcity is the point. The generator must choose where to spend emphasis.

Landmarks are navigation tools, not garnish

Players build mental maps from stable anchors. A glowing tower, colossal tree, colored transit line, or unique sound can tell them where they are and where they want to go. Place a landmark at the macro layer, then protect its sightlines at the room layer.

Reveal it early. Hide it during part of the journey. Return it from a new angle. The player now experiences progress even if the intervening modules were selected procedurally.

Local landmarks help too. Give major regions different shape language, lighting temperature, ambient audio, or material families. Color alone is fragile; combine at least two cues so navigation survives different displays and accessibility needs.

Turn seeds into bug reports

A seed should reproduce every important random choice. Display it in development builds, attach it to crash and playtest reports, and make it easy to reload. “The cave felt unfair” is fog. “Seed 184291 put two snipers over the only health pickup” is actionable.

Then make the generator report its reasoning:

• which macro template it chose;
• which room role filled each node;
• why a candidate module was rejected;
• the shortest and longest route lengths;
• encounter intensity over distance;
• how often each rare feature appeared.

Epic’s PCG tooling exposes point inspection, debug rendering, and sanity-check nodes for exactly this reason: procedural systems need observability. A generator that only outputs a finished world is a slot machine with a build button.

Reject bad worlds before the player sees them

After constructing a candidate, run validators. Some are hard failures:

• The goal is unreachable.
• A required key can spawn behind its own lock.
• A jump exceeds the movement model.
• An encounter lacks the minimum clearance or cover.
• The navigation mesh fails to connect required zones.

Others are quality scores: repeated room count, route redundancy, line-of-sight variety, distance between rest points, landmark visibility, or consecutive minutes at the same intensity. Generate several candidates, keep the strongest, and record the score. This does not prove fun, but it stops obviously broken oatmeal from reaching the table.

Use hierarchy for both design and performance

The same three-scale model helps streaming and generation cost. Unreal’s hierarchical generation guidance uses large grid cells for prominent objects such as trees and boulders, then smaller cells for grass, flowers, and stones. The principle generalizes: resolve expensive, visible structure at a broad scale; generate disposable detail only where the player can benefit from it.

Do not reroll a cathedral because one flower moved. Give sub-systems their own deterministic seeds derived from the world seed. That keeps large forms stable while artists iterate on dressing.

Let humans interrupt the machine

Mixed-initiative design lets the generator propose and the designer edit, lock, or regenerate selected parts. The PCG textbook chapter on mixed-initiative creation frames procedural tools as collaborators rather than replacements for authors.

In practice, expose controls that speak design language: “more loops,” “protect this vista,” “keep this room,” “reduce combat here.” A panel full of random-number ranges may be mathematically honest, but it leaves the creative intent trapped in someone’s head.

A small generator worth building first

Before promising an infinite continent, build a ten-minute slice with:

1. two macro journey templates;
2. six clearly defined room roles;
3. three variants per common role;
4. one landmark and one rare exception per run;
5. hard connectivity checks plus a simple pacing graph;
6. a seed browser that shows fifty generated maps at once.

Play the worst-looking seeds, not only the hero seed. Fix rules before adding content. Once the structure reliably produces understandable journeys, visual variety becomes valuable. New props, biome dressing, and material sets—including tileable surfaces from the PLAYTEX AI Texture Generator—can broaden the world without pretending decoration is design.

The goal is authored surprise

The best procedural level feels surprising in the moment and inevitable in hindsight. The player sees variation; the designer sees a carefully fenced possibility space. That is the real craft: not generating everything, but deciding what the machine is allowed to invent, what it must protect, and when it should roll the dice again.