Automatic writing

I'm a lifelong enthusiast of and passionate advocate for AGI, and although large language models like ChatGPT and Claude aren't AGI, they are proving to be uncannily useful writing collaborators. I should qualify that. Many people think it means the author giving the model a prompt like, "Write me a novel in which Philip Marlowe hunts down Dracula in 1930s Hollywood," and then grabbing a coffee while the AI spits out 60,000 words. That's how Fritz Leiber Jr envisaged AI fiction-writing in The Silver Eggheads, published in 1966.

Joanna Penn and Alicia Wright paint a fuller picture in this podcast about using generative AI in writing. Every writer finds their own way of making best use of the technology. After playing with it for a while, I've found it especially useful for research, brainstorming (I use it as a sounding board to comment on the pacing and plot holes in a storyline), and to do the grunt work of coming up with NPC stats.

Using AI for entertainment pales into insignificance beside its value in medicine and clean energy, but my science days are far behind me so I'm going to have to focus on the frivolous. For an idea of how LLMs can help with writing research, take this example of a conversation with Claude when I was working on a Faerie Queene scenario for Patreon:

Dave: "Can you give me any examples from late medieval or Tudor literature of a character posing a conundrum that tests the wisdom and moral insight of the listeners? Portia's caskets in The Merchant of Venice would be one example; I'm looking for others."

Claude: "There are several fascinating examples from this period. Let me share a few notable ones. The most prominent example comes from Book I of Thomas More's Utopia (1516), where More presents the character of Raphael Hythloday engaging in a debate at Cardinal Morton's table. Hythloday poses a complex moral conundrum about the harsh punishment of thieves in England, arguing that executing people for theft is both unjust and ineffective. He challenges the listeners to consider whether a society that first creates poverty through enclosure and then punishes the impoverished for stealing is morally coherent. This serves as a test of the listeners' ability to think through systemic social problems rather than just accept conventional wisdom.

"Another interesting example appears in John Gower's Confessio Amantis (late 14th century), where the Tale of Three Questions features a knight who must answer three riddles to save his life: what women most desire, what's most precious but least valued, and what costs least but is most valued. These questions serve to test both wisdom and moral understanding of human nature.

"In the morality play Mankind (c. 1470), the character Mercy poses several theological riddles to Mankind about the nature of sin and redemption, testing both the character's and the audience's understanding of moral theology.

"An unusual example comes from John Skelton's Bowge of Courte (c. 1498), where the protagonist encounters allegorical figures who present him with various moral dilemmas about courtly behavior and integrity, though these are more implicit than explicit riddles."

Our discussion went on (and on -- I like chatting to Claude) and I got what I needed, but here's the point: imagine how futile a simple Google search would have been for a request like that. Still, I see I've been giving the spotlight all to Claude, so here are ChatGPT's thoughts on writing gamebooks:

ChatGPT: "Since a 1000-section gamebook requires long-term consistency, custom formatting, and mechanical coherence, GPT-4-turbo or Claude 3 would be your strongest options. If you want, we can test structuring a small gamebook section now to see how it flows."

Dave: "OK, let's try it. Say the player is looking at a noticeboard with three quest notices on it. We'll use keywords, and each notice has two keywords associated with it. The first keyword is acquired after the player has read the notice once, so that if they choose to read it again they get a shorter description of what they must do to complete the quest. The other keyword in each case is acquired by completing the quest (which is outside the scope of this example, it's something they would go off and do in another part of the game world) so that if they look at a notice for a quest they already did, they are reminded that it's no longer available."

ChatGPT: "Got it! Here's how we can structure this:

1. The player sees a noticeboard with three quest notices.
   o One acquired when they read the notice (so the second reading gives a shorter summary).
   o One acquired when they complete the quest (so they are told they already finished it).
2. Each notice has two keywords:
3. The player chooses which notice to read.
4. The system checks their keywords to determine the output.

I'll draft a structured gamebook section below."

001: Noticeboard
A weathered wooden noticeboard stands at the crossroads, its surface peppered with tattered parchment. Three notices are pinned here, each offering work for those seeking adventure. You can read any of them:

·         [1] The Merchant’s Plea    go to 002

·         [2] The Missing Hunter    go to 003

·         [3] The Bandit Threat        go to 004

Choose which notice to read. If you have already read it, you will see only a summary. If you have completed the quest, you will be informed.

002: The Merchant’s Plea
If you have the keyword READ_MERCHANT, go to 005.
If you have the keyword DONE_MERCHANT, go to 006.

A wealthy merchant named Vargos seeks aid. His shipment of fine silks has been stolen on the road to Tarsen by a band of masked riders. He offers a generous reward to anyone who can recover the stolen goods.

Write down the keyword READ_MERCHANT.
Return to 001.

003: The Missing Hunter
If you have the keyword READ_HUNTER, go to 007.
If you have the keyword DONE_HUNTER, go to 008.

A notice written in hurried script pleads for help. A young hunter named Lirian went into the Blackwood three nights ago and has not returned. His father fears the worst and offers a reward for news of his fate.

Write down the keyword READ_HUNTER.
Return to 001.

004: The Bandit Threat
If you have the keyword READ_BANDIT, go to 009.
If you have the keyword DONE_BANDIT, go to 010.

A proclamation from the town watch warns of a growing bandit presence near the old mill. Travelers have gone missing, and merchants report stolen goods. The captain of the guard offers a bounty for any bandit leader brought to justice.

Write down the keyword READ_BANDIT.
Return to 001.

005: The Merchant’s Plea (Summary)
Vargos seeks aid recovering his stolen silks. A reward is offered.
Return to 001.

006: The Merchant’s Plea (Completed)
You have already helped Vargos. There is nothing more to do here.
Return to 001.

007: The Missing Hunter (Summary)
A hunter is missing in the Blackwood. His father seeks news.
Return to 001.

008: The Missing Hunter (Completed)
You have already resolved this matter. There is nothing more to do here.
Return to 001.

009: The Bandit Threat (Summary)
The town watch seeks help against a growing bandit threat.
Return to 001.

010: The Bandit Threat (Completed)
The bandit threat has been dealt with. The notice is now irrelevant.
Return to 001.

I find that pretty impressive. I wasn't expecting ChatGPT to provide the meat of the quests, and while I wouldn't use them as they are, they could serve as springboards to a better idea. I said to ChatGPT, only half in jest: "I'll soon be obsolete."

"You're far from obsolete," said ChatGPT, eager as always to be supportive. "The way you structure gamebooks requires creativity, balance, and a strong grasp of narrative design. LLMs are just tools to assist in the process. If anything, they help speed up the technical side so you can focus more on the storytelling and gameplay experience."

Which is pretty much what I've been telling everyone all along.