The research notes in here concern the Sigurd model. Do not apply anything in here to newer models (e.g. Euterpe or Krake).

For Euterpe and Krake, there is a guide on the UKB by placebomancer that describes current best practices for directing the narrative.

(pioneered by Monky, GruntBull, Cass in v2; refined by Cass for v3)

You can use [] in your input to give "instructions" to the AI. It might be most useful to think of these as "stage instructions" like you would see them written out in a play just before a scene. If the instructions are ignored in the first two sentences, it's almost always worth it to retry (or reword the instructions) as the AI tends to snowball. You may need to retry a few times depending on the settings.

[ Describe Peter guiding rats out of the city. ]

[ Describe Tom ]

If in doubt, use [ Describe... ] for stage instructions. This has worked well for many different types of stage instructions, from letting the AI describe persons or objects to describing non-immediate, longer sequences of plot events (for sudden, immediate events see event:). For describing non-immediate events, usually write your instruction in concise prose as a proper sentence, like the Peter example above.

An alternative method for describing actions seems to be [ Describe now: Peter guides rats out of the city ]. This seems to sometimes work in instances where a normal [ Describe ... ] fails. More testing is needed to determine the best method for describing actions, however. (Shincore)

[ event: sudden weather change storm ]

[ event: door kicked in 3 men ]
[ event: they abduct Peter ]
Suddenly

event: seems useful to get specific immediate things to happen. Use these for things that are more short interjections rather than long, drawn-out sequences of events.

In general, stage instructions for event seem to work more reliably in condensed, caveman-like form. Write them similar to a Caveman or Featherlite entry: Keep it short keep, ideally below 10 tokens per line. As for lore entries try to get the topic word(s) to the front, details to the back. Multi-line event might work, but try not to overdo it. event will work more reliably if the event actually fits with the scene. The more of a stretch the event is, the more the AI is likely to ignore the instruction or to describe it as if it already happened. However, starting the description with a "sudden", like in the weather example above, can help in these cases. Another way of reinforcing "unusual" events is adding a "Suddenly" after the stage instructions like in the abduction example. (TravelingRobot)

[ Scene: An Ancient Temple on a Hilltop ]

[ scene: abandoned city ]

scene allows you to specify the scenario in which the following part should take place. Very useful to set the location for a scene, less reliable for actions. Something like [ scene: visceral combat ] can work, but might not be as effective. [ Setting: seems to work in a similar way, but further testing is needed if one can be recommended over the other.

(placebomancer, Cass)

[ define: Romanticism ]

define can be useful to have AI give you "knowledge" about certain things, terms, etc. Also very useful for testing to see what the AI associates with certain keywords.

[ POV: Metzger ]

Can be used to switch the perspective to a different character in the middle of a story.

(Rando)
If you want an AID-like experience with inputs like [ say: "Hello how are you doing"] or [ do: Smash in the door] you can try the following ideas. However, usage of say and do is no longer generally recommended. In most cases, describe will be more powerful. So use the following only if you determined to get back some AID nostalgia...

You can use stage instructions to mimic AID's CYOA style with [say/do: <Your text>] - no awful AID fine-tune necessary! It is recommended to reinforce this with the following in memory.

[ Rewrite do text in prose and the story continues. ]
[ Rewrite say text in dialogue and the story continues. ]