engagement

engagement 016 · 2026

Narrative and art, growing in one direction

Two project teams, five disciplines, and a hidden coordination cost of 60–80 hours a week. We found two invariants, encoded them, and shipped a harness that let narrative and art converge instead of drift.

client
A Shanghai game studio (50 people)
sector
Games · production
duration
12 weeks
85%
weekly-active AI use (was ~15%)
81%
narrative first-pass approval (was 52%)
4.4/5
art-style consistency (was 3.2)
2
invariants encoded

The studio had two project teams and five disciplines working in parallel. The coordination cost was not visible on any schedule: narrative would ship a character, art would interpret it independently, and three sprints later someone would notice they had grown apart. The rework estimate, once we mapped it, was 60–80 hours a week — not in any single team, but distributed across the gaps between them.

We started by reading the artifacts, not the postmortems. Rejected assets, revision threads, the quest items that cycled through review two and three times. The failures had a shape. Narrative had no required structure for character intent. Art had no binding rule for style coherence. Both teams were making judgment calls that should have been invariants.

Two invariants, one coordination surface

By week four we could name two things that held across every failure: every character needed a recorded motivation–conflict–turn three-part structure before any dependent asset could be created; every art asset needed to bind to a named style anchor and an explicit list of forbidden elements. Those two rules could be encoded. The actual creative judgment — how a character arc resolves, where a style anchor sits — stays human. The job was to stop letting coordination failures masquerade as creative ones.

“We didn’t do less — for the first time, narrative and art grew in the same direction.”

The harness

We did not deliver a framework document. We delivered a running system: a narrative-design chain that enforces the three-part structure at creation time, and a style-constraint panel that binds every asset at submission. Over 12 weeks, narrative first-pass approval across a sample of 96 quest and narrative items rose from 52% to 81%. Art-style consistency, scored against the same anchors across 140 assets in internal review and small playtests, moved from 3.2 to 4.4 out of 5. Weekly-active AI use across all role families rose from roughly 15% to 85%.

The team’s capacity also changed. The studio moved from 50 to approximately 34 people over the same period — not because the work diminished, but because the coordination overhead that had occupied the difference was no longer there.

how it ran

  1. 01

    Discovery

    Mapped coordination failures across design, narrative, art, and QA. Identified rework patterns and the two rules nobody had written down. 4 weeks.

  2. 02

    Specification

    Wrote the invariants — character motivation–conflict–turn structure; art assets bound to a style anchor with a forbidden-elements list. 4 weeks.

  3. 03

    Harness

    Shipped the running system: narrative-design chain with structure enforcement, style-constraint panel with asset binding. 4 weeks.