Day 4

More Technical Review & Revisions

August 13, 2026

We just went through a round of revisions on the design doc for The Fifth Voice, our AI-driven interactive story engine (flagship game: a Cretaceous survival story where four teenage characters give you competing advice, and you can always ignore all of them and try something else). Here’s what actually changed.

The tech stack got decided, not just proposed

A few choices that were previously “pick one during implementation” are now locked in: Zustand for state management (a better fit than Redux Toolkit for a UI layer that’s meant to just subscribe to an external engine), Vitest for testing, and browser storage (IndexedDB/localStorage) for save/load in the first playable version instead of standing up a database early. We’re also building the AI provider layer on the Vercel AI SDK rather than hand-rolling our own OpenAI/Anthropic abstraction — it already does most of what we needed.

The gameplay loop got redesigned around token cost

This was the biggest change. The original loop could, in the worst case, fire off five or more separate AI calls per turn — one per Council character, plus interpretation, plus narration. The new version:

• Generates all four Council recommendations in one call instead of four.
• Skips the Action Interpreter entirely when the player just picks a Council option — there’s nothing to interpret if it’s already structured.
• Combines narration and the next Council generation into a single call by default.
• Only triggers a full Council deliberation at genuine decision points, not every minor action.

Net effect: a typical turn drops from five-plus calls to two or three, with smaller, cache-friendly context on each one. That’s both cheaper and faster — the two were never really separate problems.

We’re evaluating a local AI model as a second path

Meta just shipped Muse Glimmer, a 30B open-weight model built to run on a single consumer GPU with strong structured-output reliability — a good match for how heavily this game leans on AI output crossing into a deterministic engine. It’s not a commitment yet, just a real evaluation once we have working prompts: run our actual Council and Narrator calls through it, compare against the cloud provider, and decide from there. If it holds up, it removes the game’s biggest structural cost problem — paying per session forever — which also happens to be the thing standing between this game and a normal one-time Steam purchase.

A more honest read on the business side

We also added a clearer-eyed note on commercial expectations: games like this tend to make real money at scale through casual, subscription-style engagement, not through deep, crafted single campaigns — which is exactly the opposite of what we’re building. Realistic outcome is a well-received niche release, not a runaway hit. That’s fine — it was never the point — but it’s better to say so plainly than let scope creep toward a business model this design doesn’t actually support.

Next up: Phase 0, and finding out whether the combined Council + Narrator call actually holds together.