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 by pume_ that describes current best practices for memory (an important note is that you want to use the ideas discovered here in Memory, not in Author's Note).

First of all, also take a look at this page on the NAI knowledge base about A/N categories.

• Note on the list of keywords (TravellingRobot): Just because some keywords (like keywords for Authors, Genres, etc.) have good looking associations when using discovery methods, that does not necessarily mean that they do influence the output in the expected way. From some outputs I have seen it seems like Siggy gets genres, authors, tags or similar from Amazon reviews etc. So keywords discovered this way might not automatically translate into having an impact on Sigurd's writing style. Therefore I urge all researchers to not only explore if the AI has good looking associations for a specific keyword, but to also test if this keyword actually influences output in the desired way. If we ever want to produce useful lists of keywords for A/N instructions we must make sure these keywords actually do something meaningful with the output.
• Unless otherwise noted, all listed keywords have only seen very limited testing (mostly empty prompt or almost empty prompt). So behaviour in actual play is unknown. I'd be happy to hear back about play reports for these in the community research channel!

The following attributes are the ones actually used in the fine-tune (Zaltys)

[ Author: <author>; Tags: <tag list>; Genre: <genre list> ]

It is currently recommended to write your own A/N in this format as well. Note that the fine-tune used an empty whitespace before the closing brackets. Limited testing suggests that it might be beneficial to imitate this for the A/N format.

Example:

[ Author: H. P. Lovecraft; Tags: Cthulhu Mythos, horror; Genre: Horror ]

If you do not use all of the Big 3 attributes, it is recommended to still type out empty attributes. This seems to reinforce the effect of the remaining attributes.

Example:

[ Author: ; Tags: Cthulhu Mythos; Genre: ]

The lists of keywords for the big 3 have been moved to their own pages find them here:

• List of keywords for Author
• List of keywords for Tags
• List of keywords for Genre

The following attributes have been confirmed to have been used in the fine-tune, but only infrequently so: Title:, Setting:, Location:, Biome:. In theory, these might make them more useful than emergent categories (see bellow). However, except for Setting which has been confirmed to work fairly well, their effectiveness in actual play has yet to be determined.
In general, these seem to have been listed after Tags: and before Genre: in the fine-tune so for now it is recommended to place them there.

Example for Setting:

[ Author: ; Tags: ; Setting: Mass Effect; Genre: ]

The exception to this rule is Title which should be placed after Author:

[ Author: ; Title: Clifford the Big Red Dog; Tags: ; Genre: ]

Note that it is unclear if all of the big three should be included empty or not when only using a supporting attribute. For now, it is safer to list out all Big 3 even if you do not use them.

• The Two Towers (Cass)
• The Melancholy of Haruhi Suzumiya (Cass)
• Twilight (Cass)
• The Prelude (Cass)
• Clifford the Big Red Dog (Cass)
• Jaws (Cass)
• IT (Cass)
• Artemis Fowl works slightly better in Title than in Setting. Confirmed with recursive testing. (Cass, TravelingRobot)
• Wizard of Oz works just as well in Setting. Confirmed with recursive testing. (Cass, TravelingRobot)

Originally used to tag franchises, but seems to also work with movie titles or countries:

• The Big Sleep (1939-40) seems to give you Noir crime. Other movie titles might also work? (Cass)
• Toy Story (Cass)
• Mass Effect works quite well. Pulls knowledge about Shepard, Reapers, etc. (Covalent, Cass)
• Stargate (Shincore)
• Dragon Ball (Cass, TravelingRobot)
• The Baby-Sitters Club (Cass)
• Harry Potter (Cass, TravelingRobot)
• JoJo's Bizarre Adventure (Cass, TravelingRobot)
• My Little Pony (Cass, TravelingRobot)
• South Park (Cass, TravelingRobot)
• The Simpsons (Cass, TravelingRobot)
• SCP Foundation works especially well if you want to do SCP-style classified reports. Sigurd will even occasionally black out parts of the text... (TravelingRobot)
• various countries like France, Japan etc. Initial testing suggests setting might work slightly better than using the attribute Location for this, but needs more testing. (TravelingRobot, Cass)
• Wizard of Oz works just as well in Title. Confirmed with recursive testing. (Cass, TravelingRobot)

Many of the keywords listed here also seem to work fine in Setting. Right now it is unclear which one performs better for locatons or general environments. For now these keywords are listed here, until further testing can be conducted on 'Location' versus 'Setting' (Cass, TravelingRobot)

• Miskatonic University (placebomancer, TravelingRobot)
• can also be used to indicate the general type of environment the story takes place. Keywords that have been confirmed to work:: airplane, under the sea, beachside inn (Cass, placebomancer)

The following categories were not in the meta-data of the fine-tune. Nonetheless, they have been tested and seem to be useful for steering the writing of Sigurd. Thus, these categories can be seen as emergent properties of the model. In general these are weaker than the Big 3 described above (Author, Tag, Genre). In contrast to the Big 3, these should not be capitalized. If you are not using any of the Big 3 Attributes, it is recommended to still prepend an empty [ Genre:].
Example:

[ Genre: ]
[ period: Edo ]

• poetic (OccultSage)
• mythological seems to combine nicely with Genre: mythological if you want a text written like ancient mythology (djhato)
• mysterious on its own this mostly seems to pull standard fantasy/magic themes, but in combination with other keywords this can lead to a more brooding tone. Combining this with Genre: religious can give you descriptions of ominous religious rituals for example. (djhato, TravelingRobot)
• high detail what it says on the tin. Sigurd will give you more elaborate descriptions. (various)
• clinical (Potato)
• dark also works in tone, unclear which achieves better results. Tends towards demon story in initial testing, but this might be heavily influenced by the type of context (various)
• play, stage directions Can give you output in the format of a play. Probably best to use these two in combination. In initial testing play alone was too weak and stage directions produces text with only stage directions and nothing else. (oliacym, TravelingRobot)
• playwriting seems more focussed on writing about plays? Although this probably could work for plays it looks to be weaker than the keywords above (Cass, TravelingRobot)
• epic seems to work slightly better in style than in tone (Basileus, TravelingRobot)
• explanatory (Cass)
• classified report for generating classified action reports, files from secret services etc. (TravelingRobot)
• depressing (Potato)
• jaded seems like it can go into a depressing tone, but can also get you "seen it all" cynical characters - might depend on the context it is used in (Potatao, Cass, TravelingRobot)

• gritty seems good for harsh low-life stories (various)
• dark also works in style, unclear which achieves better results. Tends towards demon story in initial testing, but this might be heavily influenced by the type of context (various)
• whimsical in initial testing this mostly generates stories with young protagonists (children) in a style appropriate for children. Seems to work slightly better in tone than in style (placebomancer, Cass, TravelingRobot)
• wacky seems to give you people acting strangely or strange (or scary!) things happening (placebomancer, TravelingRobot)
• zany quite weak in initial testing, but might influence the output towards "strange" (magical?) things happening? Needs more testing (placebomancer, TravelingRobot).

• Edo (OccultSage)
• modern times (Fuzzy)
• Middle Ages or Medieval (Vegetables are not Geese, TravelingRobot)

Both site: and source: seem to work similarly well, more testing is needded here.

• https://powerlisting.fandom.com/wiki/Superpower_Wiki seems to pull info from this wiki. If you enter a name into an empty prompt it will assign these characters various super-powers. Might be useful for generators? Other fandom wikis might be worth trying as well. (Shincore)
• https://www.4chan.org/ start your prompt with something like Anonymous 06/27/11(Mon)01:24:16 No.15396072 and you got yourself a 4chan generator. The grammar might still be too good though... (Shincore, TravelingRobot)

Can be used to introduce well-known characters into the story. However, the characters are not necessarily introduced in the story together.
The format seems to use / so for example pairings: Santa/Devil will introduce both Santa as well as Satan into the story.
In principle any character should work, but Sigurd might only know about some famous characters and not about others. For example, Sigurd does not seem to know much about Starsky/ Hutch without extra Lorebook entries (it usually makes Starsky female), but it knows enough about Marvel characters that Ironman/ Hulk works out of the box. Reinforcing with Lorebook entries might be a good idea in some cases, though. Also, some weirder parings might confuse Sigurd enough for him to on only introduce one character. (Cass, Shincore, TravelingRobot)
The following is a list of pairingsthat seem to work well:

• Santa/ Devil (Cass)
• Ironman/ Hulk (Shincore)
• Harry/ Draco (Cass)

• slow has been reported to slow down the tonal pacing of the story (Cass)