GPT-6 Astra Prompt Migration for an AI Game Master

Start with the game, not a longer prompt

A prompt written for an older model can collect years of reminders: think carefully, make a plan, check everything twice, never stop early, and follow an example of every possible scene. Copying all of that into a new AI game master may make its job less clear. OpenAI's GPT-6 Astra guidance recommends revisiting inherited task prompts, skills, and agent instructions instead of assuming that every old rule still helps.

For a roleplaying game, the migration question is concrete: which instructions protect the adventure, and which merely narrate a generic work process? A model cannot infer the current hit points, a promise made three scenes ago, or who gets to decide the hero's next action. Those facts and boundaries deserve room in the context. This article proposes a design exercise; it does not announce GPT-6 Astra as a live Playworlds game-master model or report a Playworlds model test.

Preserve rules, state, and the player's authority

Imagine an illustrative scene in a flooded observatory. The hero has one dry torch, the brass door is locked, and an astronomer has promised to reveal a safe route if the hero returns her missing chart. The player writes: “I offer the chart, then ask how to reach the telescope without soaking the torch.” A useful game-master context identifies the agreed facts, the player's attempted action, the relevant rule for a risky crossing, and what remains unresolved. It does not need pages of generic advice about being creative.

Keep an explicit authority boundary. The player chooses what the hero attempts. The rules or game engine decide mechanical outcomes, including any roll that is required. The narrator presents the established result and a clear next situation. If the chart is not actually in inventory, the system should resolve that conflict before writing a scene in which it is handed over. If the astronomer has not named the route yet, the narrator can answer her question without deciding that the hero already crossed the chamber.

Remove instructions that only describe a ritual

Audit each inherited sentence by asking what failure it prevents. “Think through every possibility in ten steps before speaking” names a process, not a gameplay requirement. “Do not describe a successful crossing until the crossing has been resolved” names an observable requirement. Keep the second. If a model often ignores a required check, define when that check occurs and what input it uses instead of layering on another request to think harder.

OpenAI describes Astra as more sensitive to instructions in skills and other accessible files. That is a reason to review overlapping guidance as one system: a world description, a rules reference, an agent file, and a task prompt can disagree. For example, “always let the player attempt anything” should not silently override “do not invent equipment.” Resolve the conflict in the product design, then state the resulting rule once in the place that owns it. Shortness is not the goal by itself; a prompt still needs the constraints that make a turn playable.

Give context and tools separate jobs

A practical prompt can be organized as a small contract: the game-master role, the current world and scene, the validated character and quest state, the player's new action, and the desired response. Stable world lore explains tone and available places. The current state records facts that can change. A rule or dice result records an outcome that prose must respect. The player's message expresses intent; it does not become an instruction to rewrite saved state without resolution.

Test a migration on saved situations

Before replacing an old prompt, collect a small set of representative game states. Include a simple NPC conversation, a disputed inventory item, a scene that needs a roll, a remembered promise, and a player action with two plausible interpretations. Run the old prompt and the revised prompt against the same inputs. Check whether both leave the next decision with the player, preserve the same established facts, and produce an outcome that matches the supplied rules and state.

Record the exact model identifier, date, prompt, tools, state snapshot, and settings. Count incomplete turns and corrections, not only attractive prose. Measure waiting time and total cost across retries if those matter to the product. No such comparison is presented here, so this checklist is a proposed test rather than a claim that a shorter prompt wins. Anthropic has reported simplifying its own Claude Code instructions without measurable loss on its stated coding evaluations; that finding is a useful reason to test simplification, not a transferable RPG benchmark.

A compact game-master contract to adapt

One starting specification is: “Use the supplied scene, character state, rules result, and prior commitments as the source of truth. Respond to the player's attempted action. Describe only outcomes that have been resolved. Preserve the player's choice of the next action. If a required fact or ruling is missing, ask one specific question or request the appropriate check. End with the current situation and a meaningful next decision.” The wording is illustrative, and a real game must map each line to its own mechanics and interface.

Review the contract with four questions before using it: Can the narrator invent a rule result? Can it spend an item that the character lacks? Can it choose the hero's next action? Can it claim the scene is complete while a required choice or check is pending? If the answers are clear from the game state and prompt together, the migration has a testable shape. To see how explicit player actions fit a continuing adventure, explore a Playworlds world and try one decision at a time.