Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
REVIEW: 50 Years of Text Games, by Aaron Reed - By John Psmith
TL;DR
Text games mattered because words became the game world, not just the interface — John Psmith draws a sharp line between terminal-based simulations like Oregon Trail and Rocket versus true interactive fiction born with Will Crowther’s 1975 Adventure, where hand-authored prose, parser input, and “Your feet are now wet” created the illusion of limitless agency.
The genre’s core problem has always been fake freedom — interactive fiction can feel magical until the parser fails, at which point the dream collapses back into Choose Your Own Adventure-style rails, a limitation Psmith calls the fundamental reason the form never escaped subculture status.
Aaron Reed’s history becomes unexpectedly relevant in 2026 because text is AI’s native habitat — despite multimodality, Psmith argues LLMs still think in internet-text tokens, making worlds made of words unusually legible to both humans and models.
AI Dungeon didn’t fail because language models can’t understand players; it failed because LLMs are bad game masters — Psmith says current models are too accommodating, too average-seeking, and too weak at maintaining world models, so they struggle to generate the hard constraints, novelty, and simulation consistency that make games fun.
The better lineage for AI games is Dwarf Fortress, not chatbot improv — from The Hobbit’s stateful NPCs to Dwarf Fortress’s 750,000-line simulation stack, Psmith argues fun comes from agency plus emergence: systems with enough ground truth that bizarre plans like “infect a cat with vampirism and then set it on fire” produce real consequences.
Psmith ends by revealing he’s building his own AI-native text game — delayed from January because he spent months coding it, he describes a ruined world of broken concepts where LLMs handle exception cases and industrial-scale AI playtesting helps surface missing responses, in hopes that the next 50 years of text games may be their real golden age.
The Breakdown
The review opens as a parser joke, then pivots into a real thesis
Psmith starts by framing the whole review like an old-school text adventure, with commands like “Click Scott Alexander post” and the parser replying, “I’m sorry, I don’t know the verb.” It’s funny, but also the point: text games are a lost art form with their own rhythms, frustrations, and charm, and Aaron Reed’s 2023 book arrives just as AI makes text unexpectedly relevant again.
Why early computer games were graphical before text took over
He makes the counterintuitive hardware argument that in the late ’60s and early ’70s, graphics were cheaper than prose. A few numbers could place pixels or move missiles in Spacewar, but storing lots of ASCII sentences was brutally expensive — this review alone, he jokes, would overflow the Apollo guidance computer.
Before interactive fiction, text often just wrapped a simulation
Psmith distinguishes between games that merely used text terminals and games fundamentally made of language. Oregon Trail, Rocket, Hunt the Wumpus, and even The Sumerian Game are treated as simulations rendered through words — “A person is not a phone” — with text acting as an instrument panel, not the actual substance of the world.
Adventure happened because a lonely caver wanted to share his world with his kids
The emotional heart of the history is Will Crowther: caver, programmer, newly divorced father. After he lost both his marriage and his access to caving, he built Adventure in 1975 for his daughters, turning Mammoth Cave into a parser world full of dwarves, treasures, and those unforgettable “maze of twisty little passages, all alike” descriptions.
Interactive fiction became huge, brilliant, and doomed
From Adventure came Zork, Hitchhiker’s Guide to the Galaxy, Suspended, A Mind Forever Voyaging, and eventually games that really deserved the label “interactive fiction.” But just as the medium matured, graphics got good enough to crush it; Psmith quotes Reed’s killer analogy that this was like movies arriving within a decade of the printing press.
Infocom accidentally created the conditions for an immortal subculture
The Z-machine mattered twice: it preserved Infocom’s games across future hardware, and it made authorship semi-accessible by isolating hard engineering inside the VM. That opened the door to a post-commercial IF scene — Galatea, Photopia, Violet — with archives, awards, conventions, and lots of creativity, even if the parser problem never stopped haunting the form.
Why old tricks, MUDs, and AI Dungeon still don’t solve the real problem
Psmith tours the genre’s escape attempts: MUDs brought in other humans, wizard-users authored content, Eliza-style smoke and mirrors faked understanding, and games like His Majesty’s Ship Impetuous pretended your input mattered when it mostly didn’t. He says modern LLMs improve comprehension but still fail as game engines because they’re too sycophantic, too mediocre, and too unreliable about world state — the “dogs that haven’t barked” line about AI-driven games is one of the review’s sharpest observations.
The real answer might be simulation plus AI, and Psmith is trying it himself
The review crescendos through The Hobbit and Dwarf Fortress, where systemic depth creates real agency and genuine surprise. Then Psmith admits the review was delayed because he’s been building his own text game: a cyclical ruined world of broken concepts where simulation provides ground truth, LLMs bridge edge cases, and robot playtesters hammer the game at scale to find what human players might try next.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
