Craft

Writing GM Instructions

Your GM instructions are the single most important factor in how your game feels. They shape tone, pacing, rules enforcement, and narrative style. But more instructions don't mean better results. The best instruction sets are short, specific, and structured for how the GM actually reads them.

Less Is More

Every instruction you add competes with every other instruction for the GM's attention. A tight set of 10 clear rules will be followed more reliably than 50 vague ones. When you write an instruction, you're not just telling the GM what to do — you're implicitly telling it to pay less attention to everything else.

Before adding a new instruction, ask yourself: is this important enough to dilute everything else I've written? If it's situational or only matters occasionally, it might belong in a file with Partial visibility instead.

Preview Your Context

If you are trying to troubleshoot your instructions, open the context preview. Check for:

  • Conflicting instructions. If your GM instructions say "keep combat fast and loose" but a pinned rules file has a detailed turn order system, the GM has to pick one. It might not pick the one you want.
  • Missing context Your instruction might not be working because it is relying on something that is missing from context. Check the preview to see if all of the pieces are there.
  • Total volume. If you are having trouble keeping all of your different instructions straight in your head, then it's likely the GM will have trouble with it too! When writing instructions aim for clarity and brevity.

Write Conditionals, Not Absolutes

A common mistake is writing instructions like "ALWAYS read the character sheet before responding." or "NPCs should ALWAYS do x" Ask yourself: is it really every turn? If a player asks an out-of-character question, do you want the GM to update its notes? If someone cracks a joke, should it re-read a file?

Absolute instructions burn tokens on turns where they don't apply and can make the GM feel rigid or robotic. Instead, write conditionals that give the GM judgment about when to act:

  • Instead of: "Always describe the environment in detail."

  • Try: "When players enter a new location, describe it in detail."

  • Instead of: "NPCs should always write interesting dialog"

  • Try: "When it makes sense, have NPCs bring up interesting details in conversation"

If you have some basic process that you do expect to have the GM do on every turn, it is okay to have a bit of absolute wording in your prompts. Just don't overdo it!

Don't Ask the GM to Re-Read Files

You might be tempted to add instructions like "Re-read the world bible every turn." Don't. The GM already manages its own context. When it uses the read tool, that content stays in the conversation history for multiple turns. It will read files again when they fall out of context or when it needs to reference them.

You can control how much conversation history is included in context through the conversation history length setting. If the GM has the last 10 messages in context, a file it read 3 turns ago is still visible. Telling it to re-read every turn just wastes tokens on duplicate content.

If the GM seems to be forgetting something from a specific file, the fix is usually to pin a short summary of the key points, not to force re-reading the whole file.

Split Your Instructions by Frequency

Not every instruction is relevant on every turn. Combat rules don't matter during a shopping scene. Character creation guidelines are useless mid-session. Loading all of this into every single response is wasteful.

Use context visibility levels to match instructions to how often they're needed:

Pin instructions the GM needs on 80-90% or more of turns:

  • Core tone and personality ("You are a dark, gritty narrator who favors consequence over heroism")
  • Universal rules ("Players can attempt anything, but actions have proportional consequences")
  • Session-level directives ("This session focuses on the trial of Lord Ashwick")

Set to Partial for instructions needed in specific situations:

  • Combat rules and mechanics
  • Magic system details
  • Specific NPC behavior guides
  • Travel and exploration procedures

Set to Searchable for deep reference material:

  • Full lore documents
  • Historical timelines
  • Detailed location descriptions
  • Item catalogs

The GM will pull in Partial and Searchable content when it's relevant. You don't need to force it all into every turn.

Understand Model Differences

Different models have different strengths when it comes to instruction following. A larger model will generally adhere to more of your instructions more consistently, while a smaller model might miss subtler directives or deprioritize instructions it considers less important in the moment.

This isn't a bug — it's how language models work. Instructions are guidance, not code. The model interprets them in context and may decide that a particular instruction doesn't apply to a specific moment. For example, if you say "always end with a question" but a scene reaches a dramatic climax, the model might choose to let the moment land instead.

If you're not sure whether the GM considered your instruction:

  1. Open the action details. You can view the model's reasoning to see if it weighed your instruction and chose to act differently, or if it missed it entirely.
  2. Try simplifying. If you have many instructions, the model may be deprioritizing some. Cut the less important ones and see if adherence improves.
  3. Try a bigger model. If instruction following is critical for your game, a larger model will handle a more complex instruction set.

Don't expect 100% adherence from any model. Write your instructions knowing that some will occasionally be deprioritized, and make sure the ones that matter most are clear, prominent, and few.

If an instruction isn't being followed and you're not sure why, try switching to a larger model temporarily. If the larger model follows it, the instruction is fine — your current model just isn't strong enough for that level of nuance. If the larger model also ignores it, the instruction itself probably needs to be rewritten.

Correct Bad Patterns Early

Language models are pattern followers. If the GM has been doing something wrong for the last several turns, it's likely to keep doing it — not because it's ignoring your instructions, but because its own recent output reinforces the habit.

When you notice the GM falling into a bad pattern (wrong tone, skipping a rule, narrating a character incorrectly), send a message telling it what to adjust. Think of it like course-correcting — the model learns from the conversation flow, so a direct correction resets the pattern. Left uncorrected, the GM will keep repeating the mistake because that's what its recent context tells it to do.

Quick Reference

Do Don't
Use conditionals ("when X, do Y") Use absolutes ("ALWAYS", "EVERY TURN")
Preview your full context before playing Assume your instructions exist in isolation
Pin only what's needed every turn Pin everything "just in case"
Use Partial for situational rules Keep combat rules pinned during exploration
Let the GM manage its own file reading Force re-reads of files every turn
Correct bad patterns with a direct message Hope the GM self-corrects over time
Test with a bigger model to diagnose issues Assume your instructions are wrong

See Also

PreviousProgressive Disclosure