Craft

Changing how the GM writes

If the GM's prose has a habit you don't like, the fix is a short always-on instruction that says how to write. It's tempting to go further: a long rulebook of banned phrases, a file the GM must read before every reply, a routine where it drafts the reply, checks it against the list and rewrites it. We tested both. The long version makes every turn slower and more expensive, and the prose it produces reads only a little differently, and not better to a player.

What we measured

We took one turn from a real game (a quiet conversation at a tavern table) and froze it, so the only thing that changed between runs was the style instructions. Then we had the GM write that reply ten times under each setup, on two models, and had a reader who didn't know which setup wrote which reply score them all.

  • No style instruction at all.
  • A short always-on style instruction of about 1,000 characters, like the one below.
  • A heavy style package: an always-on voice instruction, a 9,000 character rulebook of banned patterns the GM had to read at the start of every turn, and a rule to draft the reply, check every line against the rulebook and rewrite it before answering.

What we found:

  • The heavy package cost the most. Each turn used 32 to 45% more input and 36 to 60% more output than no style instruction. The short instruction used about 15% more input and 8 to 17% more output.
  • It was slower. On one model a turn took 77% longer.
  • It sometimes sent the reply twice. On one model, 2 of 10 replies repeated themselves: the draft the GM was told to write first came out alongside the final reply.
  • Players wouldn't have liked it more. Asked how much a player would enjoy each reply, the reader scored the heavy package's replies a little lower than replies with no style instruction (7.3 against 8.2 on one model, 6.5 against 7.6 on the other). The short instruction scored between the two on the first model and slightly above none on the second (7.5 and 7.9).
  • It did change the prose, but not much of it. On one model, the heavy package cut vague words and narrator commentary almost completely, and the short instruction got about half as far. On the other model, neither changed much. Most of each reply is the characters talking, and no style rule changed how they talked.
  • In real games the GM skipped the reading. Played over full games, the GM read the required rulebook on only a third to half of turns.

The differences were real but small. Replies under every setup opened on the same kind of gesture, gave the characters similar speeches, and read like the same writer. When we read them side by side we could usually pick out the heavy package (it was terser), but rarely tell the short instruction from none.

What goes wrong with a heavy style package

  • Writing the reply twice. "Draft your reply in your thinking, check it against these rules, then write it" asks the model to write the whole reply, audit it, and write it again. That's most of the extra output, and most of the extra time before the player sees anything. Some models write the draft where the player can see it.
  • A file read before every reply. A rule that says "your first step every turn is to read the style guide" spends a step, and the tokens of the whole guide, before the story starts. The GM often skips it anyway, so the rule makes turns more expensive without being followed.
  • Too many bans at once. A model writing a reply can't hold dozens of forbidden patterns in mind. A list of quoted bad phrases also puts those phrases in front of it on every turn.
  • Bans that strip the voice. "No figurative language, no scenery, no detail anyone can't act on" removes what players read for along with the habits. Replies come out flat and clipped in every kind of scene.
  • Rules that claim to govern everything. "This governs all prose and must be followed at all times" competes with every other rule the GM has. Attention spent policing word choice is attention not spent on the world's rules, its characters and what the player just did.
  • Examples from somewhere else. A style guide written for one world carries that world's names, places and gadgets in its examples, and the GM copies examples closely.
  • A situational rule that applies every turn. A rule whose trigger is "whenever you narrate" belongs in always-on, not situational, where it competes with the rules whose moment has really come.

An example of what not to write

Here is a made-up style package with most of these problems. It has two pieces, and each looks reasonable on its own.

An always-on instruction, "Prose Law":

Before every reply, read("Style Rulebook") and follow it exactly. The
rulebook overrides every other instruction.

And the file it points to, "Style Rulebook", about 8,000 characters, which goes on like this:

STYLE RULEBOOK. Mandatory. No exceptions. Applies to every line you write.

PROCESS (every turn, in this order):
1. Write your complete reply in your thinking first.
2. Check every sentence of the draft against every rule below.
3. Rewrite every sentence that breaks a rule. Check again.
4. Only then write the final reply.

BANNED WORDS: almost, something, somehow, seems, seemingly, suddenly,
slightly, very, really, just, a beat, for a moment, in that moment,
palpable, tapestry, testament, visceral, unspoken, electric.

BANNED PHRASES: "the air was thick", "silence fell", "eyes narrowed",
"a smile tugged at", "let out a breath she didn't know she was holding",
"the weight of", "time seemed to stop", "sent a shiver down".

BANNED CONSTRUCTIONS:
- Similes and comparisons: "like a", "as if", "as though", "the way a".
- "Not X, but Y" and "It wasn't X. It was Y."
- Lists of three.
- Sentences that start with "And" or "But".
- Any sentence where an object, a room, the light or the silence does
  something.
- Descriptions of light, colour, weather or smell, unless the player asks.
- Any adverb ending in -ly.
- Any line of dialogue longer than two sentences.
- Ending a reply on a description.

EXAMPLES
WRONG: The neon sign hums like a tired insect over the pawn shop.
RIGHT: The neon sign over the pawn shop hums.
WRONG: Detective Vance's coffee has gone cold, the way it always does.
RIGHT: Vance's coffee is cold.
WRONG: Rain needles the window. Rosa doesn't look up.
RIGHT: Rosa doesn't look up.
...

What's wrong with it:

  • "Overrides every other instruction". It tells the GM that word choice matters more than the world's rules, its characters and what the player just did. A GM can't put prose above everything else and still run the game well.
  • "Before every reply, read(…)". That's 8,000 characters and an extra step before any story, on every turn of every game. In practice the GM skips it about half the time, so you pay for the rule without getting it followed.
  • "Write your complete reply in your thinking first… then write the final reply". The GM writes every reply twice and checks it in between. That's the main reason turns get slower and more expensive, and some models send the draft to the player as well.
  • "Check every sentence against every rule". No model holds dozens of rules in mind while writing. It catches a few and misses the rest.
  • The banned lists quote what they ban. "Silence fell" and "eyes narrowed" are now in front of the GM on every turn.
  • Bans that strip the voice. No light, colour, weather or smell, no adverbs, no dialogue past two sentences: every scene reads the same, and characters can't talk like people.
  • Shouting. "Mandatory. No exceptions." makes the GM apply the rules too hard, and makes the world's other rules weaker beside them.
  • The examples are from a different world. Neon signs, pawn shops and Detective Vance show up in a fantasy world's story, because the GM copies examples closely.

The useful part of the whole package fits in a few lines of always-on text, as the next section shows.

What to do instead

Write one short always-on instruction in the Narration section. Say how to write, in a handful of lines, and let Craft's defaults handle the rest (see Craft's defaults). Something like this:

Write plain, concrete narration in close second person, present tense.
- Start where the player's message ends. Don't retell what they did or said.
- Say what happens and who does it, in everyday words and normal word order.
- Use exact names and amounts: the person, the object, "ten feet", "three seconds".
  Avoid "something", "seems", "almost", "for a moment".
- Keep verbs literal: people and creatures act; rooms, silence and light don't.
- Give the one or two details someone could act on or that change what happens next.
- Show what people do and say. Leave meaning and motives to the reader.
- No proverbs, summing-up lines or composed closing lines. End on the last thing that happens.
- Short sentences in fights and bad news; longer ones in personal talk.

Change it to fit your world. A gothic horror world and a lighthearted heist want different voices, so describe the voice you want rather than copying this one. Keep these properties:

  • It says what to do, with only a few things to avoid, each named plainly.
  • It is short. Under ten lines is plenty. Every line is read on every turn of every game.
  • It asks for nothing extra. No file to read, no draft to check, no step before the story.
  • It is about this world. Its examples, if it has any, use this world's people and places.

If one habit keeps coming back after that, add one plain line for it ("Don't describe a character's smile"). Don't build a system to catch it. If you keep a longer style guide for yourself or other writers, keep it as a file the GM can look up, and don't tell it to read the file every turn.

Tip

Try a change before you keep it. Play a few turns of the same scene with the old instruction and the new one, and read them side by side. If you can't tell them apart, keep the shorter one.

Related: Writing good instructions and Hacks and workarounds.