What it is. An invisible system that decides what threat appears, when, and how much — from hand-placed spawns, to budget-priced waves, to a reactive director that reads player state and drives a tension curve. It is the author of pacing you never see.

Player fantasy / why it’s fun. The world feels alive and out to get you — pressure that swells and releases like breathing, a lull that’s earned and a peak that’s survived. The best directors feel authored, never random: every spike reads as a story beat the player will retell.

When to use / when NOT

Use when: replayable content (roguelike, genre-horde-survival, genre-extraction-shooter) where hand-placing every threat is impossible; when you want the same map to feel different each run; when pacing must adapt to how well the player is doing.

Don’t use when: a tightly authored genre-precision-platformer or genre-grid-puzzle where every element is a designed problem — a director robs you of the deliberate difficulty ramp. Also skip for a first-timer’s opening minutes: hand-place the system-onboarding so the tutorial is legible, then hand control to the director.

Variants

Budgets and directors compose: price threats with a budget, then let the director decide when to spend. See system-encounter-design for the per-encounter grammar the director assembles from.

Tuning levers

• Intensity curve. The shape the director aims for over a session — the target it steers toward. Peak-and-valley, not a flat line or a monotone climb. See pattern-pacing-and-tension.
• Peak-and-valley rhythm. L4D’s model: build-up → peak → relax → ensure a lull → repeat. The valley is not idle time; it is the setup that makes the next peak legible. Cut it and the player numbs.
• Budget cap. Ceiling on threat priced per beat. Prevents a bad roll from stacking three elites in one room. Cap scales with run length, not real time.
• Cooldown / lull floor. Minimum quiet between peaks. Guarantees the player can reload, heal, breathe — the anti-frustration valve.
• Spawn distance & sightlines. Where threat appears governs fairness as much as how much. Off-screen-but-audible telegraphs (see system-telegraphs) beat teleport-behind-you every time.
• Composition weights. Ratio of fodder to specials to elites. Shifts the texture of a peak, not just its height. Pull from system-enemy-archetypes.
• Reactivity gain. How hard the director leans on player performance. High gain = rubber-band (system-difficulty-and-dda); low gain = mostly scripted with light adaptation. Tune conservatively — players smell rubber-band.

Twist seams

• Director but it targets the strongest player (risk-reward) — in co-op, the director concentrates threat on whoever is dominating: the fed player draws the elite, the swarm, the special. Rewards mastery with pressure, protects the struggling teammate, and manufactures the clutch save. See pattern-risk-reward, system-coop-and-competition.
• Spawns but the player sets the budget for a reward (perspective) — flip authorship: the player dials up the spawn budget before a run (more threat) in exchange for richer loot. The director becomes a wager. anchor-rimworld and Risk-of-Rain-style difficulty dials live here; ties into system-loot-tables and system-reward-schedules.

How it wires to Hayao

• Determinism is non-negotiable. Every spawn roll, budget draw, and director decision goes through the world’s deterministic RNG — never wall-clock time or ambient randomness. A director that can’t be replayed can’t be proven fair or debugged. Same seed, same run.
• Model the director as pure state. Keep tension, budget, and cooldown timers as explicit fields advanced by a pure step function — the same discipline as a turn-based rules module. The scene tree only reflects what the director decided; view stays cosmetic.
• Read player state through a stable probe. The director consumes a snapshot (health, position, recent damage, pace) — not live view nodes. Order every iteration (enemy lists, spawn candidates) so decisions are reproducible.
• Learn the parts in isolation before wiring a director: procedural placement and seeded rolls belong in a procgen sandbox; enemy behavior once spawned belongs with system-enemy-ai. Compose, don’t conflate.

Fails when…

• It feels unfair: threat teleports in, spawns in your blind spot, or peaks with no telegraph → antipattern-guess-the-designer, and see pattern-fairness-and-trust.
• It feels arbitrary: pure randomness with no curve reads as noise, not authorship. A bad seed ruins a run → antipattern-rng-frustration.
• No valleys: relentless pressure with no lull numbs the player and erases the peaks it’s trying to sell → antipattern-fail-loop-tax.
• Rubber-band too visible: punishing success so hard that playing worse is optimal → antipattern-boring-optimal. Keep reactivity gain low; see system-difficulty-and-dda.
• Content desert: three enemy types on infinite shuffle — a director needs a deep archetype roster to compose from, or every peak feels the same → antipattern-content-desert.

Verify

Determinism and fairness are the whole game here — prove them, don’t assert them by feel. A seeded run must replay identically; the intensity curve must show peaks and valleys, not a flat or monotone line; no spawn may violate the distance/sightline fairness rule. Hold it to docs/VERIFICATION.md.