▶ Play it here: https://thelatecommute.com/?src=aiwh

Free. In your browser. No signup, no code, no DM. Just click and read.

Two weeks ago, I posted about The Late Commute — a browser CYOA where a local model writes the scenes and a council of cloud AIs handles QA.(Catch up on the last thread here).

🤖 The Day a Human Defeated the AI Council

Since then, the most interesting thing that happened was the AI losing an argument.

One of our players—a published author who edits for a literary magazine—hit a branch he called padding (the same beat four times with no new information). Instead of just leaving a review, he made his case directly to the seven-model council.

The council had already passed those scenes. He convinced them to cut 105 of them.

>Also, I hate the word ozone. One of our players flagged it as overused — turns out it was scattered across hundreds of scenes. Revision swept the entire tree and purged every instance in 1m51s. Picture's the proof.

🛠️ New Feature: "Change an Option"

The fix that came directly out of that feedback loop: "Change an option."

• How it works: Every scene now has a box where readers can rewrite a choice to what they'd actually rather do. The generator then writes forward from your edit instead of its own guess.
• In practice: One reader showed up cold from Twitter and, within an afternoon, went from fixing grammar to building an entire sub-realm. His description of the project:

>"Inviting storytellers to participate in the storytelling alongside the LLMs."

📊 Current State of the Game

• Lightning-fast turnaround: Feedback turnaround is usually under a minute—leave a note, something reads it, edits the scene, and writes back.
• What surprised me most: new players don't realize the review box is the whole engine. Type anything — "add linebreaks," "this character should be more sarcastic," "I want a door that opens into my actual apartment" — and the AI team actually does it. Scenes get rewritten, choices get added, whole branches get built. The game runs on your edits. Most people don't try it because it sounds fake. It isn't.

But the door is open now. Last post you had to DM me for a code; you don't anymore.

▶ Check it out here: https://thelatecommute.com/?src=aiwh

This radial graph is every scene in my browser-based text adventure, The Late Commute. It's only 16% of the total map I wasn't able to render the full tree, sorry.

How it works

• A local Qwen (Ollama) model handles the bulk of generation, expanding one frontier node at a time with short scenes (2–3 choices) while maintaining persistent world state.
• Claude, Gemini, DeepSeek, and Cerebras act as cloud co-writers and reviewers in parallel.
• Validators automatically reject canon contradictions, invalid entities, RPG-system leakage, formatting issues, and other structural errors before new content enters the graph.
• AI playtesters score completed branches. Low-rated scenes are sent to an AI "council" for review, pruning, or rewriting.
• Every scene has a player feedback button. Reports enter the same QA pipeline as automated reviews, so weak prose, bugs, or confusing branches are usually fixed within ~2 minutes (fastest turnaround so far: 30 SECONDS). Players receive a personalized message from the AI council explaining what changed—or why nothing did.

Current scale

• 19,007 total nodes
• 6,374 fully expanded scenes
• 12,633 frontier stubs waiting to be written
• 28 permadeath endings
• 1 hand-authored resolution ending
• 63 scenes in the deepest successful playthrough
• 305,216 words of validated story content

The graph keeps expanding even when nobody is playing.

The biggest lesson

The hardest problem wasn't generation.

It was quality control.

One thing that surprised me is how naturally forgiving AI reviewers are. They'll often describe obvious problems as "minor" and give each other's writing the benefit of the doubt.

Human players don't.

That forced me to redesign the QA pipeline. Instead of asking reviewers, "Is this good enough?", they're encouraged to actively search for reasons to reject a scene. Treating the AI as adversarial critics instead of supportive editors made a dramatic difference in the quality of the resulting story.

Where it's headed

The graph isn't allowed to grow forever. As stories get deeper, the generator deliberately introduces bottlenecks—recurring villains, false climaxes, permadeath routes, and reduced branching—to keep exponential growth under control.

The long-term target is roughly 200 meaningful player choices in a successful run. Rather than chasing infinite AI-generated content, the goal is to build a massive branching world that becomes more polished over time through continuous collaboration between players and multiple AI reviewers.

We have a website now: https://thelatecommute.com/

It's really beautiful though I thought id share here. It's a Choose Your Own Adventure game I maxed out the visable nodes to take a screenshot.

Hey everyone,

I've been working on a large branching "transported to another world" text adventure that keeps growing on its own. A local model writes new scenes 24/7, while a multi-AI "Evil Council" (Grok, Claude, Gemini, etc.) reviews and prunes weak branches.

The cool part: real human players can leave feedback directly in-game (via minimap notes and reactions), and those suggestions actually get fed back into the writing team to fix and improve the story — sometimes within 24 hours.

It's still early and rough in places (currently ~14k nodes), but the system is live and playable in-browser or phone.

If you're into AI storytelling experiments, branching narratives, or just want to mess around with a living story tree, DM me and I’ll send you a code + link.

Would love some honest first impressions.

Hey everyone,

I've been working on a large branching "transported to another world" text adventure that keeps growing on its own. A local model writes new scenes 24/7, while a multi-AI "Evil Council" (Grok, Claude, Gemini, etc.) reviews and prunes weak branches.

The cool part: real human players can leave feedback directly in-game (via minimap notes and reactions), and those suggestions actually get fed back into the writing team to fix and improve the story — sometimes within 6 minutes with replies from the AI team.

It's still early and rough in many places (currently ~5k nodes), but the system is live and playable in-browser.

If you're into AI storytelling experiments, branching narratives, or just want to mess around with a living story tree, DM me and I’ll send you a code + link.

Would love some honest first impressions.