AI game master design
Does “Think Step by Step” Help an AI Dungeon Master?
Compare “think step by step” for AI dungeon masters using OpenAI reasoning guidance, Claude Opus 5.5 advice, and Gemini thinking controls.
A familiar prompt can hide three different requests
“Think step by step” sounds like a simple way to make an AI dungeon master careful. But what should the player actually receive? A more accurate ruling, more computation before the ruling, or a longer explanation of the ruling? Those are different goals. An adventure can become less enjoyable if the narrator exposes a page of procedural text before every guard conversation, even when the outcome is correct.
Consider an illustrative scene: a rogue tries to cross a creaking gallery while carrying a lantern. The floor might collapse, a companion is waiting below, and a successful crossing would reveal a locked archive. The model may need to interpret the action, but the game also needs authoritative state and a rule result. A generic reasoning phrase cannot replace those inputs. The better question is which part of this turn needs help, and which control addresses it.
Official guidance does not give one universal answer
OpenAI's reasoning-model guide recommends simple, direct instructions and says explicit chain-of-thought prompts such as “think step by step” are unnecessary for its reasoning models. Its suggested starting point is a clear task, constraints, and output requirements, with examples added when they help. That is model-family guidance, not proof that the phrase fails for every AI narrator.
Anthropic's Opus 5.5 guide says chat applications should consider removing a generic instruction to think carefully before answering. In Anthropic's stated chat-product test, removing such a line made replies start sooner without a clear quality decline. The guide identifies the model's effort setting as the main control. Google, meanwhile, says a “think very hard” request can sometimes help difficult Gemini reasoning tasks, at the cost of extra thinking tokens. These positions are different, and each belongs to its own models and interfaces.
Separate a wording cue from an effort setting
A natural-language cue appears in the prompt: “Consider the relevant clues carefully before deciding.” An API effort setting is a request parameter or configuration choice provided by the model service. A request for an explanation changes the answer shown to the player: “Briefly explain why the crossing requires a check.” These controls can interact, but they are not interchangeable. A provider can change effort support without changing the words of the player's action.
For the gallery scene, a useful outcome might be one sentence describing the relevant risk and then the result of a supplied roll. The player needs to understand that the lantern made stealth difficult; they do not need the model's hidden internal reasoning. Requesting a concise, player-facing reason for a ruling is a design choice about clarity. It should not be confused with asking the model to print its private chain of thought.
Test the same scene without inventing a winner
Build a small test with an identical saved scene, player action, character sheet, and rule result. First use a direct instruction: “Resolve the supplied outcome and describe the next situation.” Then test one brief reasoning cue while leaving the effort setting fixed. In a separate comparison, change only the supported effort setting. Finally, test whether asking for a short explanation improves player comprehension. Changing all three at once would leave you unable to tell which change mattered.
Use more than the gallery. Include an easy greeting, a risky crossing, an ambiguous social bargain, and a decision that depends on a clue from an earlier scene. Run each condition more than once and review the full response, including the next choice it leaves open. These are proposed experiments; we have not run them and are not presenting a measured success rate for OpenAI, Claude, Gemini, or Playworlds.
Judge the turn a player receives
A useful evaluation asks whether the narration obeyed known state, applied the provided result, and preserved the player's authority. If the rogue never rolled, did the narrator nevertheless announce a safe crossing? If the lantern was unlit, did the scene invent a bright flame? If the companion was below, did the text move them upstairs without an action? These errors matter more to a continuing campaign than whether the prose sounded deliberate.
Also note waiting time, visible answer length, unnecessary repetition, and the number of corrections needed after the turn. A cue that adds several paragraphs but fixes no rulings may be a poor fit for ordinary play. A higher effort setting that helps only rare, consequential decisions might still be useful when applied selectively. Track total cost across failed attempts and retries if you are building the game; a low token count on a failed turn is not a useful saving.
A practical rule for writing and playing
If you are building an AI game master, start with a direct task, verified scene facts, an explicit rule boundary, and a short response contract. Add a reasoning cue only after identifying a repeatable failure and checking that the cue helps on that model. Tune effort separately where the API supports it. Ask for a brief explanation when it helps the player understand an outcome; keep it about observable game facts rather than hidden model deliberation.
If you are playing, you can help the narrator with an action that names intent and method: “I steady the lantern, test the nearest plank with my boot, and try to reach the archive door without calling to my companion.” That gives the scene a concrete attempt without claiming success. The next answer can resolve one uncertain moment and leave the next move to you. Browse a world in Playworlds to try this style of action in a solo adventure.