iteration-loop

Iteration Loop

The loop every gamestack content type runs to close the gap between an intended target and what's actually implemented — whether the target is concept art, a balance curve, or lore canon. One loop shape, instantiated per modality.

Tier: process (tier-spanning, lives in gamestack-core). Not a knowledge skill — a recurring procedure other skills plug into.

The loop

Reference → Diff → Prioritize → Generate/Implement → Verify, run per system, repeated until the gap closes or is explicitly deferred.

1. Reference — pull the intended target for this system: a concept-art subset, a stated balance target, a lore-canon fact set. If references aren't organized per-system, that's the first fix — an unindexed reference pile can't be diffed against.
2. Diff — compare the current implementation against the reference and name the single biggest delta (not a laundry list). This step stays human/agent judgment, not a mechanical pixel-diff or string-diff — see LOOP.md for why and for the modality-specific diff method.
3. Prioritize — pick the highest-benefit delta to close next. One delta at a time keeps each cycle verifiable.
4. Generate/Implement — close it: generate an asset (asset-gen pipeline + your project's optimize/import pass), tune a system (cost curves, drop tables), or author content (within established canon) — then wire it in data-driven, never as a hardcoded one-off, and file it into its categorized asset pack (LOOP.md §7). For visual/aesthetic deltas, generate against one instance first, not the whole system; for heavyweight pipelines (3D model gen, tileset gen), stage that behind a few cheap 2D mockups first — see LOOP.md §6 for why and the batching pattern. When generation is the wrong tool for this delta (animations, trees, a hand-authored hero asset, …), route to the external-request escalation (§5) instead of forcing a worse generated result.
5. Verify — re-check against the reference. Did the named delta close? If not, the cycle repeats; if a new bigger delta surfaced, that's the next Prioritize. For visual/aesthetic deltas, the generating agent's own judgment of "does this look right" isn't reliable verification — that routes through a human checkpoint (LOOP.md §6) before the change scales beyond the single instance.