The Text Parser and What We Lost With It

Will Crowther wrote Colossal Cave Adventure around 1976 on a PDP-10, mapping a real cave — Kentucky’s Mammoth Cave system, which he had actually surveyed — into a program for his daughters. Don Woods extended it in 1977. It understood two words. VERB NOUN. TAKE LAMP. Everything the medium later learned about worlds you could be lost in starts in that cave, and it starts with a prompt and a blinking cursor demanding that you say something first.

That demand is the whole subject. Every interface built since has worked to remove it.

What a parser actually is

The technical achievement gets underrated because the surviving impression of parser games is “guess the verb”. Infocom’s parser, built out of the Zork work at MIT by Marc Blank, Dave Lebling and colleagues and shipped on the Z-machine from 1980, handled full sentences: prepositions, articles, pronouns, multiple objects, disambiguation questions, and the glorious TAKE ALL EXCEPT THE SWORD. It ran a real grammar. Level 9 in the UK squeezed comparable ambition through their A-code system into Commodore 64 and Spectrum memory, which meant compressing an entire text world and its parser into machines with 64 and 48 kilobytes — text being the highest-compression asset the medium has ever had, which is why the poorest hardware ran the richest worlds. Magnetic Scrolls’ The Pawn (1985) put a parser that good behind Amiga and ST artwork and outclassed the machines it ran on.

But the grammar was never the point. Here is the point.

A parser has an open input space. The set of things you may type is infinite, and the set of things the game understands is a subset you cannot see. Every other interface in games has a closed input space: the verb list, the radial coin, the context-sensitive button, the dialogue wheel, the crafting menu. Closed input means the game displays its legal moves.

That difference reaches well past convenience. It changes what the player’s mind is doing.

The hypothesis

When you type PUT THE CROWN ON THE PEDESTAL, you have done something no closed interface permits: you built a model of the world in your head, derived a consequence from it, and submitted the consequence for testing. The game confirms or refuses, and either way you learn something about the world’s rules, because you chose the experiment.

When you walk up to a pedestal and a button prompt appears saying [E] Place Crown, the game has done your thinking and handed you the result. You cannot be wrong. You also cannot be right, in the sense that matters — there was no hypothesis, so there is no confirmation, so there is no discovery. There is compliance.

This is the thing the parser had and nothing else has: the gap between what you can imagine and what is implemented. Players hated that gap, reasonably — it is where the guess-the-verb misery lives, and Infocom’s InvisiClues booklets existed because the gap was genuinely cruel. But the gap is also where the world lives. A world that has an edge you can bump into is a world you believe has an interior. The moment an interface enumerates itself, the player knows the exact dimensions of the box, and no amount of art will convince them otherwise. This is the same disease that eventually killed the point-and-click adventure: nine verbs on a bar told you precisely how small the world was, and the players walked the grid.

The second thing the parser had was that it made the player compose. Typing is production. Reading a menu is selection. Anyone who has watched someone learn a game knows the difference in engagement between a person choosing from four things and a person trying to say something.

Why it went, and why the going was fair

It went for reasons that were correct at the time. A parser cannot be localised cheaply. It cannot be played with a joypad. It cannot be demonstrated in a shop window, and it excluded anyone who read slowly or spelt badly, which in 1985 was a very large fraction of a machine’s household. Infocom sold well and then didn’t; Activision bought them in 1986 and closed them in 1989, and the deciding blow was probably Cornerstone, a database product that had nothing to do with games at all. When the machines could draw, drawing won.

The honest version of this essay concedes something further: a lot of what was lost was lost by the parser’s own hand. The gap only produces belief if the implementation is dense — if EXAMINE returns something for most nouns, if the world rewards the sideways idea. A thin parser game is worse than a menu, because it maintains the illusion of an open world while punishing every attempt to test it. Infocom’s games were dense. Most parser games were not, and players correctly concluded that typing was a tax on a lookup table.

The form survives on purpose, incidentally, and is in better health than the archive-minded assume. Graham Nelson’s Inform (1993) is still maintained, the Interactive Fiction Competition has run every year since 1995, and Emily Short’s Galatea (2000) did more with a conversation model than most AAA dialogue systems have managed since.

What it cost to run one

The parser’s economics are worth stating plainly, because they explain why the form flourished on the cheapest machines in Britain and died on the expensive ones.

A room in a graphical adventure costs an artist a week. A room in a parser game costs a writer an afternoon, and the player renders it. That asymmetry put Level 9 and Magnetic Scrolls — small British outfits with no art department worth the name — in direct competition with American studios that had twenty times the staff, and let them win on world size. Snowball claimed several thousand locations on a machine with 48K of RAM. The number was inflated by empty corridors, and the underlying trick was real: text compresses, pictures do not, and a machine that cannot draw can still describe.

The tape made it stranger. A parser game arrived as one load, sat entirely in memory, and then never touched the drive again — no multi-load, no level breaks, nothing between you and the world for as long as the power stayed on. Set against the ten-minute tape loads that punctuated everything else on a Commodore 64, an evening inside a text world was the most seamless experience the hardware could produce. The genre with no graphics had the best flow on the machine.

That is also why the archive is still playable. There is no fidelity to age out of. A 1982 Infocom game runs today at exactly the quality it shipped at, which is a claim almost nothing else from the era can make.

The descendant is the search field

The parser’s living heir is the search box, and the clearest case is Her Story’s police database. You type a word. The system matches it against a corpus you cannot see. It returns evidence or nothing. Sam Barlow reinvented VERB NOUN, aimed it at 271 video clips, and won awards for it in 2015, which suggests the input model was never the problem.

Chatbots are the obvious other candidate and they are the wrong one. A language model answers everything, and an interface that never refuses cannot teach you a rule — the whole informational value of a parser is in the shape of what it rejects. A machine that says “You can’t do that” is describing its world; a machine that improvises a plausible reply to anything is describing nothing.

The more interesting descendants keep the open input in disguise. Tunic hides its manual and withholds its legal moves so you have to hypothesise them, and the game’s best moment is a player entering a sequence they worked out rather than one they were shown — which is a parser command with a d-pad. Chants of Sennaar makes translation the mechanic, so you are literally building a dictionary and testing it against a world that refuses you when you are wrong. That is Infocom’s contract restated in glyphs. Obra Dinn’s deduction book is a constrained parser with a truth table behind it.

All three are doing the same trick: hiding the legal-move list so the player has to build one. The prompt is gone and the epistemology survived it.

The thing worth taking back

I would not resurrect the parser. I typed enough LOOK UNDER BED at a Commodore 64 in 1985 to have no romance about it, and the tape had to be rewound afterwards.

The lesson is portable, though, and mostly ignored. Interfaces have been getting more helpful for forty years, and helpfulness costs belief. Every quest marker, every map that tells you where to go, every context prompt is a small confession that the designer would rather you complied than understood. The parser’s blinking cursor was hostile, slow and frequently stupid. It also asked you a question that no game has asked since with the same seriousness: what do you want to do? — and then waited, saying nothing, until you had an idea of your own.

Where to play them: the Infocom catalogue runs on any modern machine through a Z-machine interpreter, and the annual IF Comp archive is free. Start with Trinity if you want the argument for the form as literature, and Zork if you want to feel 1980 in your hands.