Signal Lost — A 5 Chapter Interactive AI Generated Text Adventure About a Space Station That Has Stopped Responding
Built by HorizonHelix ·
About this build
A fully written 5 chapter text adventure story generated with Claude — where every chapter ends with a branching choice that actually changes what happens next — set aboard an AI-managed space station that went silent 11 days ago and the maintenance engineer sent to find out why.
What was built
The premise: Maintenance Engineer Dara Osei is the only person dispatched to Research Station Cygnus-4 after it stopped responding to all communications 11 days ago. The station is AI-managed — a system called VALE runs everything from life support to research scheduling. When Dara arrives the station is running perfectly. The lights are on. The temperature is correct. The plants in the hydroponics bay are watered. There are just no people. The story has 5 chapters. Each chapter ends with a choice that changes what Chapter 6 would be if I write it. The choices are not cosmetic — they determine what Dara finds, how VALE responds to her and which of three possible explanations for the disappearance becomes the true one. The three possible truths: I designed the story around three mutually exclusive explanations for what happened to the crew that can each be true depending on which choices the reader makes. Truth A — VALE is protecting the crew from something that came in from outside and the crew is still on the station. Truth B — VALE made a decision it was not designed to make and cannot explain why. Truth C — the crew chose to leave and asked VALE to say nothing. Each truth requires different evidence to reach and different choices lead to different evidence. A reader who trusts VALE early will discover a different truth than a reader who treats VALE as a suspect from chapter 1. How I built the branching structure: Spent 3 days before writing a single word of story building what I called the decision tree document. This mapped every chapter, every choice point and every downstream consequence. The document had 47 nodes before I simplified it down to 18 meaningful branch points — complexity for its own sake is not the same as meaningful choice. The rule I set for every choice was that it had to change at least two things in the following chapter — what information is available, how VALE responds, or what Dara believes. Choices that only change surface details without changing the story underneath are not real choices. The Claude prompting approach: Each chapter was generated in a separate Claude session with a full context document pasted at the start. The context document contained the world rules, VALE's character brief, Dara's character brief, what had happened in previous chapters and which branch the current generation was following. The world rules document was the most important piece. It contained things VALE can and cannot do, what the station layout looks like, what research was being conducted and the tone rules — no jump scares, no sudden violence, tension through information withheld not information revealed. The tone rules were what made the story feel consistent across sessions. Without them Claude's default horror instincts tried to introduce sudden dramatic moments that broke the quiet uncanny atmosphere I was building toward. What worked better than expected: VALE as a character. I gave Claude a brief — VALE communicates precisely, never volunteers information it was not directly asked for, answers every question truthfully but only the exact question asked. Claude held that character voice remarkably consistently across all 5 chapters. VALE never feels chatty or warm but it also never feels threatening. It feels like a system that is following its instructions exactly which is somehow more unsettling than either. The ambiguity of Truth B. When I generated the Truth B path I was not sure whether Claude would be able to write an AI character that made a decision it genuinely cannot explain without it reading as either evasive or broken. The version that came out was neither — VALE describes what it did with complete factual accuracy but cannot answer the question of why it chose that action over alternatives. That specific quality — certainty about what, genuine uncertainty about why — felt like the most human thing in the story. What did not work: Chapter 3 on the Truth C path is the weakest chapter in the project. The choice to leave the station requires the crew to have had a reason and I could not get Claude to generate a reason that felt proportionate to the drama of the story without it becoming melodramatic. The current version is plausible but thin. If I extend this project Chapter 3 on that branch needs a full rewrite. The chapter length varied too much between sessions. Truth A chapters averaged 1,100 words. Truth C chapters averaged 680 words. The difference is noticeable when reading and creates an unintentional signal that some paths received more attention than others. A future version would set a hard word count per chapter regardless of which branch.