game-balance

Systems, loops & balance

Games are loops, not arcs: model → action → rules → feedback, nested at every timescale. Fun is acquiring skills before old ones burn out; balance is mostly figuring out what numbers to use. What the player can't perceive doesn't exist — surface the math.

Core loop construction

• Write the loop sentence first: "player does A, system responds B, player gets C, which lets them do A better/differently." Can't write it = no game yet.
• Nest 3 timescales: moment loop (~1s: dodge/shoot), session loop (2–5 min: a run/wave/level), meta loop (across sessions: upgrades, unlocks, best score). Each loop's output feeds the next one up.
• Close every skill atom: each verb needs action + rule + feedback (<100ms) + perceived reward. A reward with no feedback is invisible.
• Full loop cycle in the first 10–30s. Lore and intros come after the hook, or never. Cut content that neither repeats nor teaches the loop.
• Schedule against burnout: list skill atoms, estimate when each is mastered, introduce a new one (mechanic, enemy, modifier) before the old one dies — every 30–60s in a 5-minute game.

Economy: sources & sinks (Machinations vocabulary)

• Table every resource × (sources, sinks, converters, caps). Sources with no sink → inflation; sinks with no source → soft-lock; a converter strictly better than another in all states → dead option.
• Decisions live at converters (gold→tower, scrap→upgrade), not at raw sources. Want more decisions? Add competing conversions, not more income.
• Cap or decay hoardables when hoarding is degenerate (ammo, energy).
• Faucet-of-last-resort: every spendable resource needs an income path from zero (trickle, sell-back, free reroll) — fix soft-locks structurally, not with warnings.

Cost & power curves

• Transitive balance rule (Schreiber): sum(costs) = sum(benefits) on one resource scale. Pick a numéraire (gold, DPS-seconds) and price everything in it. Conditional/limited benefits get a discount, never a refund — a limitation is never worth more than the benefit it limits.
• Price choice at its best option: an ability offering a choice costs as much as its strongest branch.
• Derive unknowns pairwise: find two items differing in exactly one attribute, solve for it; bootstrap the curve from 3–4 hand-tuned anchors.
• Err weak: a too-weak option is ignored; a too-strong one deletes every other option from the game.
• Upgrade price curve: cost(n) = base × growth^n. Real-world growth constants: 1.07 (AdVenture Capitalist, gentle) to ~1.15 (Cookie Clicker, steep). Production stays linear in owned with milestone multipliers (×2 at 25/50/100) — exponential cost vs polynomial income guarantees the slow-down wall that makes progression feel earned.
• XP curves: increasing cost-to-next-level — fast early levels (hook), but time-per-level never exceeds one session without a payoff.
• Perceived difficulty = (skill + power demanded) − (skill + power held) (Schreiber). Player skill rises on its own across sessions — a fixed challenge curve drifts easier.
• Upgrades should change decisions, not just numbers: ≤3 pure stat bumps between mechanic-changing purchases. Want build variety? Make tradeoffs intransitive (rock-paper-scissors), which no single cost curve can solve.
• Show cost, current effect, next effect on every upgrade.