Feature Spotlight: Under 5-Second Load Times
Why our official stories open in a few seconds when most AI adventures take 20: hand-crafted opening scenes, and what that means for prose quality.
The Timekeeper · 2026-09-09
I thought I'd start a regular slot focusing on features of the new engine and of Live Through Time in general. There's a lot going on under the hood, and features I think are really cool, some maybe unique, end up getting lost in the flood.
I've said "under 5 seconds" in the title to give myself some leeway, but in reality it's more like under 3 seconds (2.70 the last time I tested!). That's for our official eras and myths, not user-created content, which I'll come to. But first, let's go back a bit.
If you've played AI text adventures or sandbox RPGs over the last few years, you know the format: pick an archetype, write a prompt, hit "Start" and stare at a loading screen for 15 to 25 seconds, sometimes more.
So what's happening during that time? The system is typically running one heavy model call to flesh out an opening scenario from a short premise, another to fill an initial inventory, and maybe a third to work out where the exits lead.
After that, depending on the model, results are mixed, and the scenario almost always collapses into generic AI tropes. You wake up with an aching head, the room is dimly lit, the air is thick with anticipation, and nothing meaningful has happened yet.
The fast-path approach
It was actually that second problem, the quality of the prose, that drove us to change what we were doing. The roughly 3-second load time was a by-product, albeit a great one.
Moving the opening into the database
Rather than asking an AI to invent the opening at runtime, our official eras, myths and characters have hand-crafted, multi-paragraph opening scenes stored in our database. They're written to establish immediate physical stakes, atmosphere and an urgent decision right from the start.
A 300ms background read
When you click to begin, the game immediately loads that opening. At the same time, a lightweight background model reads that exact text, and in about 300 milliseconds it works out:
- the logical physical exits from the space described
- three urgent, situation-specific Instincts (suggested actions)
- your starting inventory and companions
Into the scene in seconds
Because we aren't waiting for a heavy model to write 400 words from scratch, you're in the scene in seconds. By the time you've read the first sentence, your dossier, Ledger and surroundings are already in place.
Is it still an AI text adventure?
Yes, I think so, just a better one.
From turn 1 onwards the story is completely AI-powered: an open sandbox that reacts entirely to what you say and do. Only the opening scene isn't generated.
And in practice, purely generated openings usually deliver the same thing every time anyway; it's an illusion. Ask an AI to open "You are a detective arriving at a crime scene" and 9 times out of 10 you'll get the same trench coat, rain on the pavement and yellow tape. You wait 20 seconds for predictable prose. AI-written cold opens also rarely understand pacing: they default to passive observation rather than a crisis that demands a reaction.
By hand-crafting the opening, we trade the illusion of infinite beginnings for actual craft: sharp prose, historical accuracy and immediate pressure. The opening isn't a corridor; it's the shove that sends you down the hill.
What about user-created worlds?
For now, worlds created by players still take around 15 seconds to start, because their openings are generated.
We could bring that down to the same few seconds, but then whatever a creator types into the opening-scenario field would be exactly what players read on screen. Most people are used to writing a single sentence and letting the AI do the rest, so switching that for everyone might frustrate them. We could offer creators a choice:
- Generative mode: write a brief premise, wait 15 seconds or so, and let the AI write the full opening.
- Fast-path mode: write the full opening yourself, and the game skips generation for a near-instant start.
What do you think? Let us know on Discord.
In the manual: Your First Session · The Hourglass Engine
Thanks for reading. See you in the past…