game-narrative-craft

name

game-narrative-craft

description

Sentence- and passage-level prose craft for short-form game narrative: choice text, interstitials, vignettes, event text, quest copy, epilogues, story bibles, and any in-game writing the player actually reads. Use when writing or reviewing narrative prose for a game, when AI-drafted game text reads flat or generic, when a passage told to be 'brief' came back as a summary instead of a scene, when narrative doesn't land because the player lacks context the writer had, or when branch content is doubling without the story feeling more responsive. Covers the compression rule, the information-state ledger, word-count bands per structural slot, choice-as-dialogue, branch-the-frame, failure-as-different-content, callback discipline, entropy seeding, perspective, and an AI-tell blacklist. Genre-neutral — pair with a register skill such as cosmic-horror-register for tone. Sits below narrative-design, which covers architecture and branch budget rather than prose.

Game Narrative Craft

Purpose: Make short-form game prose land on a human player who has partial context, limited attention, and no interest in your design document.

Scope: The passage, the paragraph, the sentence. For quest architecture, branch budgets, and the agency spectrum, see narrative-design. For tone and diction, see a register skill such as cosmic-horror-register.

Sources: Rules here are derived from two measured corpora, deliberately chosen for opposite registers so that shared technique separates from house style. See references/source-analysis.md for the measurements and the divergences.

• Arkham Horror LCG: Edge of the Earth campaign guide (FFG, 2022) — cold, earnest, second-person-party, printed and read aloud between sessions.
• Fallen London writer guidelines (Failbetter Games, published) — arch, comic, Victorian, web storylets read in 90-second sittings.

A rule appearing in both is craft. A rule appearing in one is style, and is marked as such.

---

The compression rule

This is the spine. Most bad AI game prose fails here and nowhere else.

Told to be brief, a model compresses by abstraction. It writes the summary of a scene instead of a short scene. The word count drops, the effect goes to zero, and the instruction that caused it was "keep it brief."

Given a beat where the player finds a dead scholar's sketchbook, the published text (≈200 words) does this: a physical action, a fruitless search, a discovery, one image rendered in real detail, an emotional beat, a hook. The AI version does this:

You examine Dyer's sketches and find disturbing images of an archway and something terrible beneath it.

Same information. One sentence. Nothing happens to the reader.

Rule: compress by scope reduction, never by abstraction. One moment, one room, one exchange — rendered fully. Never summarize three moments.

Test (falsifiable): if the passage could be a bullet in a design document without losing anything, it is a summary, not a scene. Rewrite it by picking the single most concrete thing in it and staying there.

Failbetter states the positive form: results should describe one action or closely-related actions and their consequences, not strings of sequential actions. Both corpora converge on the same move — narrow the aperture, keep the detail.

The rendered-detail floor

A scene that survives compression contains at least one thing a camera could photograph. Not an atmosphere, not a realization — an object, a gesture, a sound, a specific line of dialogue. If you cut the passage down and no photographable thing survives, you cut the wrong things.

---

Word-count bands

Both corpora impose hard limits per structural slot. The numbers differ because the delivery differs; the practice of a per-slot ceiling is the craft. Pick the row matching your delivery, or measure your own corpus.

slot	printed / read-aloud (AHLCG, measured n=58)	web storylet (Failbetter, published)
choice / branch label	one line of spoken dialogue	≤ 20 words
situation / root text	350–430 words	≤ 30 words
outcome / vignette	150–350 words (p25 165, median 203, p75 329)	≤ 100 words
closing beat	180–280 words	—

Working default when you have no corpus: 150–350 words for a vignette, and a hard ceiling you write down before drafting. The ceiling is the instrument. Without one, "keep it tight" is a weak ask and will be complied with by abstraction — see the compression rule.

Ranges are targets, not laws. A passage 40% over the band is a signal to look for two moments that should be one, not an automatic cut.

---

Teaching copy is a third slot with its own rules

A tip, a tutorial notice, a HUD hint — a line under about twenty words whose job is to make a mechanic usable right now — is neither a vignette nor a branch label, and most of this skill does not apply to it. The compression rule and the rendered-detail floor are about scenes. Teaching copy is not a scene. It is an instruction competing with a threat for the player's attention.

Six rules, all checkable.

1. One concept per line. If the line teaches two things, the player retains the first and the second is spent. Split it or drop one.

2. No forward references. A line that names something the player has not met is a riddle, and a riddle in the teaching channel is worse than silence — it consumes the slot and teaches nothing. Ominous forward references are excellent and belong in the flavour channel, where a high miss rate costs nothing.

3. Canonical referents only. One word per game noun, every time. See the referent ledger in generated-content-quality; this is where its findings land.

4. No metaphor above the player's current model. "Light is ground you do not have to fight for" requires the player to already hold darkness is contested space. If the game has not taught that, the sentence is decoration wearing an instruction's clothes.

5. Do not negate a misconception the player may not have. "Gear comes out of your allowance, not out of the load you logged" is answering an unasked question, and it introduces the wrong answer in the same breath as the right one.

6. Actionable within ten seconds, using only what is on screen. This is the test:

Test: could the player act on this line in the next ten seconds, with only what is currently visible? If not, it is flavour. Move it to the flavour channel and free the slot.

The dependency order check

The failure this catches is mechanical, invisible to per-line review, and I have now found it in two separate games on the first pass.

Build a two-column table: every teaching line, the concepts it introduces, and the concepts it uses. Sort by fire order. Any line that uses a concept a later line introduces is a bug.

Worked example. A logging game fires haul first —

WOOD IN YOUR ARMS IS NOT WOOD ON THE TRUCK. ONLY THE TRUCK COUNTS.

— and quota second, which is the line that introduces the idea of counting toward anything. For the length of one notice the player holds a scoring rule for a score that does not exist yet. Neither line is wrong; the order is.

The same bug appears in prose. A dungeon crawler says "THE COURT EXPECTED SOMEONE" on floor three and "THE PROCESSION ENDS AT A COURT" on floor four — definite article before the referent exists. Grep your corpus for THE <NOUN> and check each first occurrence against the fire order.

Density and pressure

Teaching lands or fails on when it arrives, not only on what it says.

• Under threat, teaching is not received. A line that fires during combat is competing with survival and will lose. Either fire it before the threat or accept that it is flavour.
• Count the lines per location and per minute. Three panels for one building, seven seconds apart, is not three lessons — it is one lesson and two interruptions.
• Measure hit rates, and read the zeroes both ways. A line that never fires may have a broken trigger or may be describing a state the game never produces. Those need opposite fixes, and the counter cannot tell them apart.

---

The information-state ledger

The crux of AI-written game narrative. The model holds the whole design document; the player holds forty fragmented minutes from three days ago. Every recap-bloat and every unearned reveal traces to that asymmetry.

Before writing any passage, fill four columns. Two minutes, in a scratch file or in your head, but actually fill them:

column	question
Certain	What has the player been shown directly and unambiguously?
Seen, not understood	What have they observed without an explanation?
Told, unreliably	What has an NPC, document, or rumor asserted?
Assumed by this passage	What does the draft require the reader to already hold?

Anything in column 4 that is not in columns 1–3 is the miss. Either seed it earlier, move it to column 3 by having someone say it, or cut the sentence that depends on it.

Reorientation is a structural slot, not a prose duty

Both corpora refuse to recap in prose, and both solve re-entry structurally instead:

• AHLCG references state by name and trusts the player to hold it — "If William Dyer is crossed out:" — because the campaign log carries continuity as a physical artifact the player maintains.
• Failbetter builds re-entry surfaces on purpose: quality-linked descriptions, dedicated reminder branches, and explicit [advisory text in square brackets] for players who engage sporadically.

Rule: never write "as you recall" or "having previously discovered." If the player needs reorientation, give it a slot of its own — a log line, a journal entry, a bracketed note, a reminder branch — and keep the prose clean.

This is also what makes the reference-by-name style possible. The prose can say "If Dyer is crossed out" with no explanation precisely because the log is doing the remembering.

---

Player agency in the sentence

Both corpora forbid the same thing and draw the line in different places. The prohibition is craft; the line is a dial.

Shared rule: never assign the player a decision, a conviction, or an opinion in narration. Choice is exposed as choice. "You realize you have to stop him" is the single most common failure in AI game prose.

The dial — how much interiority may narration assign?

position	permits	corpus
Strict	Nothing internal. Only images and incident. "Wary of putting words in the player's mouth, thoughts in their head, or feelings in their heart."	Failbetter
Sensory	Physical reaction and uncertainty. "chills you to your core", "You're still unsure what to make of all this." Never conviction.	AHLCG

Pick one per project and write it down. Strict suits games where the player authors a character; sensory suits games with a defined protagonist or a party. Mixing them mid-project is what makes a script feel like several people wrote it, because it means several people did.

Failbetter's corollary, worth adopting in either position: don't assume player gender, clothing, or skin colour.

Choices are dialogue, not menu items

Write the option as a first-person line the character says, in quotes:

"I believe you… but if what you say is true, should we not investigate these findings further?"

Not Option A: Investigate further. The choice becomes an act of characterization instead of a selection. Costs nothing, changes the entire texture.

Where the choice is an action rather than speech, Failbetter's rule applies: the branch must make unambiguously clear what the character will do. A mysterious branch label is not intrigue, it is a broken control.

---

Branch the frame, not the plot

Divergent paths that reconverge should share their content and differ in framing. In the AHLCG prologue, the two paths — the player believed the warning, or dismissed it — share three verbatim-identical paragraphs, reach the same destination, and apply the identical mechanical outcome. Only the frame around them differs.

AI writing will not do this unprompted. It writes two genuinely divergent branches, doubles the content debt, and buys no additional felt agency.

Rule: when paths reconverge, write the shared middle once and vary only the frame. Budget divergence for the moments the player will remember.

See narrative-design for the architecture this serves (hub/diamond, branch budget, reactive narrative).

Failure branches are different content, not lesser content

The highest-value technique in either corpus, and the one AI most reliably inverts.

When the scholar is alive, you get a conversation with him. When he is dead, you do not get a diminished version of that conversation — you get his sketchbook, an archway with five glyphs, and a drawing whose lines "disintegrate into a tangle of light strokes and splotches of ink." The dead branch is better material than the live branch, and it grants a card the live branch does not.

Rule: draft the worst outcome first. Draft the success path first and every other branch becomes a degradation of it — which is exactly how you get this backwards. Start with the branch where the character died, the mission failed, the resource ran out. Ask what that state makes available that success does not.

Test: if every failure branch is the success branch minus something, no failure branch has been designed.

---

Callbacks

Plant a specific, unexplained detail; return to it hours later without explanation; return once more at the end. In the AHLCG corpus a student's half-remembered Poe recitation appears in the prologue, again in a mid-campaign interlude, and again in the epilogue. It is never glossed.

Rule: three appearances, hours apart, zero explanation. Explaining the callback on its second appearance destroys it. If you are worried the player missed it, that worry is what the log line is for.

---

The state schema is the narrative's API

Both corpora make the same structural claim from opposite directions.

• Failbetter calls it quality parsimony: use as few unique state variables as possible; two is the magic minimum for content with an engaging interactive structure; excess qualities complicate the economy, cause bugs, and clutter the UI.
• AHLCG achieves rich branching on a campaign log holding almost nothing: who is alive, a handful of flags, the contents of a bag.

Rich prose over a small typed state. If the state sprawls, the branching becomes unwritable and the writer will paper over it with recap — which reintroduces the information-state failure above.

Rule: design the log before the prose. Write down every state variable the narrative may branch on. If the list exceeds what you can hold in your head, cut it before you write a word.

---

Entropy

Model output for creative content converges hard. Two seeding moves, both cheap:

1. Seed from real entropy, not imagined randomness. "Pick something random" has a basin of its own. Take entropy from the environment — a commit SHA, openssl rand -hex 6, a die rolled in the analysis tool, the low bits of a file hash — and spend it on a constraint: which of five registers, which sense leads, which object is present, which branch gets drafted first. Never ship the seed as output.
2. Roll the draft order. Combined with the failure-first rule above: rolling which branch you write first is the difference between designed branches and degraded ones.

For character, faction, and place names specifically, this is a deeper problem with its own procedure — models converge on the same names across unrelated projects. Name the two or three candidates you would have defaulted to and rule them out before generating, and apply the portability test: if the name would work equally well for a different project in a different genre, it is a mood, not a name.

---

Perspective

Three questions before any passage. They take fifteen seconds and they are the difference between a scene and a report.

1. Whose eyes? Not "third person limited" — which person, in this room, right now.
2. What do they not know? A perspective is defined by its gaps.
3. What does the player know that they don't — or the reverse?

Question 3 is where the content usually is. The scene of a broken man reciting Poe to himself works because the player recognizes the quotation's weight and he is past recognizing anything. That gap is the scene. Without it there is a man muttering.

---

AI tells

Fix these before anything else; they are what makes drafted prose read as drafted.

Structural

• The summary-instead-of-scene (see the compression rule) — the big one.
• Recap of what the player already did.
• Every paragraph ending on a portentous short sentence. And then it moved.
• Perfectly symmetrical branch outcomes.
• Explaining the mystery, the callback, or the joke.
• Every NPC speaking in the same register.

Sentence-level

• Asserting the effect instead of causing it: the weight of the moment settled over you, a chill ran down your spine, you can't shake the feeling that…
• Adjective stacks doing the work of an image: strange, unsettling, eerie.
• Tricolon on every beat: the cold, the dark, the silence.
• Passive voice as a retreat. Failbetter: passive "drains energy from prose" and signals a writer who has got distant from the text. Treat it as a reminder to get back down there.
• Single-sentence paragraphs used only to control pace. Failbetter calls this "precious," and it is.
• Em-dashes at AI density. They are the clearest tell in published prose.

Dialogue

Failbetter's test, adopted verbatim because nothing else is as fast: dialogue must pass the say-this-shit test. Read the line aloud. If no human would say it in that situation, it is exposition wearing quotation marks.

---

Story bible vs. in-game text

One source, two products, and the compression between them is a deliberate step, not a hope.

• The bible may be long. It holds the state schema, the voice sheet, the callback plants, the frame for each branch, and everything the player will never read.
• The in-game text is the band. Draft in the bible, then compress by scope reduction into the band, with the word ceiling written down before you start.

Do not draft in-game text directly from the bible in one pass. That is the step where abstraction gets in.

The voice sheet

Three lines per speaking character, written once, kept beside the draft:

1. What they want in this scene, in six words.
2. How they talk — sentence length, formality, one verbal habit.
3. What they never say — the subject they route around.

Line 3 does the most work and is the one AI omits. It is also where phonetic dialect belongs, if at all: one character rendered phonetically is characterization, three is noise.

---

Review checklist

Run against any drafted passage.

• Could this be a design-doc bullet without loss? → summary, not a scene.
• Is there one photographable thing in it?
• Word count inside the declared band?
• Does it assume anything not in the ledger's first three columns?
• Does narration assign the player a decision, conviction, or opinion?
• Is interiority consistent with the project's declared dial position?
• Are choices written as dialogue or as actions the player can predict?
• For reconverging paths: is shared content written once?
• Is the failure branch different content, or the success branch minus?
• Any callback explained on this appearance? Unexplain it.
• Any recap that belongs in a log line, journal entry, or reminder branch?
• Read all dialogue aloud — does it pass the say-this-shit test?
• Scan the AI-tell list.

---

Cross-references

• narrative-design — quest architecture, branch budget, agency spectrum, pacing. The layer above this one.
• cosmic-horror-register — tone, diction, and sensory palette for cosmic horror. Instance one of the register template; the same six slots build folk-horror, noir, or comic registers.
• experience-design — session-level pacing and emotional arc.
• player-ux — where information lives on screen, which is the delivery half of the information-state problem.

---

Cross-References

Three tight links. Everything else routes through the map, so adding a skill touches one file rather than twenty: references/routing-map.md (in game-design).

• narrative-design — Architecture and branch budget: the layer above
• cosmic-horror-register — Tone, and the template for other registers
• generated-content-quality — When the writing happens at volume