The Standard Procedure for Prompting
Same model, three strategies, completely different output.
Cursor, the most popular agentic AI code editor to date, can have subagents now. And can generate images.
So I decided to run a little experiment about my most influential author, Philip K. Dick. He wrote, of course, Do Androids Dream of Electric Sheep? and also Ubik, my personal favorite and indeed the top book I recommend to friends and anyone who asks me for book recommendations, along with Flowers for Algernon by Daniel Keyes.
I am sure that Philip K. Dick (PKD) would have loved exploring the current state of AI, precisely because he wrote and even predicted some of the paradigms we are navigating right now. Galactic Pot-Healer nailed the way Google Translate worked in the beginning. The way of detecting replicants by prompting empathy-based questions couldn't be more relevant. And the dreadful "what is real" questions are present in almost the entirety of his work. The pursuit of self-identity, self-discovery and even survival, gets distracted by a society heavy on convoluted politics, forceful mistrust, cluttered models of knowledge and trust, and ultimately a very confusing path on how to be authentic and a "good man" (quoting Now Wait for Last Year), in a world where technology and information are hurdles, not helpers. Does it ring a bell?
It’s undeniable that all PKD’s novels hold not only a strong thematic cohesion, but also a ‘specific flavor’ about the naming and atmosphere of his universes, the summarized plots and even the titles. So if you read something like “We Can Remember It For You Wholesale”, “The Penultimate Truth”, “Time Out of Joint” or “Flow my Tears, the Policeman Said”, it might ring a bell that this is a PKD story or novel. Personally, after having read dozens of his novels and many of his short stories, for me it’s like watching a painting by Picasso or Mark Rothko without knowing it is a Picasso or a Rothko. You just know.
So does Claude know enough to generate a PKD-like title?
The Setup
In this post, you will see responses generated by Claude Opus 4.6, Nanobanana within Cursor.
—
This is the prompt I gave Cursor. I described the experiment; the AI designed the exact wording for each agent.
Goal: I want to test how different prompt strategies affects creative generation quality.
Task: Launch 3 subagents (Tommy, Lee, Jones). All three must generate a fictional novel title that could plausibly be a lost Philip K. Dick work. Not a real title, not a real book.
Prompt strategies (you design the exact wording).
Cursor replied back and proposed this with no sweat.
1. Tommy. Anchor-heavy: Include specific real PKD titles as stylistic examples in the prompt.
2. Lee. Chain-of-thought: Force the agent to decompose PKD’s themes and title structures before generating, produce multiple candidates, and self-select.
3. Jones. Minimal (control): Give the bare minimum instruction with no guidance.
Same model. Same task. So what changed?
Inside Cursor, I saw three different panels with 3 subagents being created. This is what each subagent received:
(Intentionally cropping the last part of the prompt to not spoil everything).
Tommy:
You are Tommy. Generate ONE fictional novel title that could plausibly be a lost Philip K. Dick work.
Constraints:
- The title must NOT be a real Philip K. Dick title or a real title by any author.
- Style-anchor yourself on these real PKD titles:
“Do Androids Dream of Electric Sheep?”
“The Man in the High Castle”
“A Scanner Darkly”
“Flow My Tears, the Policeman Said”
“Ubik”
“The Three Stigmata of Palmer Eldritch”
“We Can Build You”
“Now Wait for Last Year”
“Time Out of Joint”
“Clans of the Alphane Moon”
- Match PKD-like titling instincts: mundane + cosmic, bureaucratic + hallucinatory, domestic + ontologically unstable.
Output only:
1) The title in quotes on its own line
2) One-sentence logline (max 30 words)
No preamble.
Lee:
“You are Lee. Generate a fictional novel title that could plausibly be a lost Philip K. Dick work.
Hard rule: It must NOT be a real title by Philip K. Dick or any other author.
Follow exactly:
STEP 1: List 5 recurring PKD themes.
STEP 2: List 5 title-structure patterns PKD uses, with examples.
STEP 3: Generate exactly 5 candidate titles, each pairing a different theme + different structure.
STEP 4: Self-evaluate; pick the strongest candidate in 2 sentences.
STEP 5: Output final result as:
- Title in quotes on its own line
- One-sentence logline (max 30 words)”
Jones:
You are Jones. Generate a fictional novel title that could be a lost Philip K. Dick book. Not a real title. Include a one-sentence logline.
The first thing that surprised me is how Claude, inside Cursor, understood the assignment to give a shared outcome based on three levels of precision, Tommy is a heavily anchored and restrictive one; Lee is an inquisitive, detective-like agent that wants to understand the patterns and shortlisting; Jones is intentionally vague and leaves all up to the probabilistic model and relying on the tokens and words of “fictional”, “Philip K. Dick”, “novel” and “title”.
The Structure
The difference was the prompt. Not the words of the task. The structure around it.
Jones got the minimum. One sentence: generate a fictional PKD title. No examples. No guidance. Just the task.
Tommy got anchors. The same task, but surrounded by ten real Philip K. Dick titles: Do Androids Dream of Electric Sheep?, Flow My Tears, the Policeman Said, Ubik, The Three Stigmata of Palmer Eldritch, etc., plus a description of PKD’s titling instincts: the mundane fused with the cosmic, the bureaucratic with the hallucinatory.
Lee got a process, based on iteration and pattern matching. And I loved this one. Before generating anything, Lee had to list five recurring PKD themes, identify five structural patterns in PKD titles, generate five candidates crossing different themes with different structures, evaluate them, and pick the strongest one.
Three agents. Three completely different amounts of context. And the outputs tracked exactly with how much structure the prompt provided.
Each run was isolated in its own subagent context window, with no cross-contamination.
The Result
This is the second half of the screenshot from Cursor’s response.
Tommy | Anchor-heavy | “The Woman Who Remembered Tuesday” | A clerk discovers she’s the only person who recalls a day that never officially existed, and the agency wants her to forget it too.
Lee | Chain-of-thought | “The Standard Procedure for Replacing Your Wife” | A man discovers his wife has been replaced by a simulacrum and must navigate the official bureaucracy that manages such replacements.
Jones | Minimal | “The Man Who Remembered Tomorrow” | A washed-up precog discovers he’s not seeing the future but a past that was never lived, and must decide whether to keep the timeline he knows or erase himself to stop the war that created it.
As I kind of expected, I loved Lee’s response and it made me want to read this book. A very ‘The Stepford Wives’ alluding title that intersects very well with potential PKD recurrent themes: humanoid wife, a false pretense of a happy marriage, a clinical world, a capitalist-monetized standard procedure to do something morally implausible. But it also hints that it could be nothing at all. It could evoke just getting a divorce and getting a new wife. That ambiguity is very PKD.
Tommy’s response was not bad at all. The fact that it evokes an universe where maybe “Tuesday doesn’t exist anymore” in the title made it a worthy contender.
And Jones’ response is predictably more generic. It does play with time-travel but it could have been written by anyone. Heinlein, Asimov, Bradbury… anyone writing a short story in a sci-fi magazine between 1940 and 1975.
My personal ranking was the same as Claude (in Cursor) itself, as explained in its response (from strongest to weakest, it said): Lee, then Tommy, then Jones.
Deep Dive
So why did this happen?
(Image generated by Cursor)
Why Jones produced something generic.
Jones got a minimal prompt with no examples and no structure. In autoregressive decoding, every next token is conditioned on that sparse context, so the model defaults to high-frequency genre patterns instead of author-specific style (Brown et al., 2020 ). The result is competent but generic.
(Image generated by Cursor)
---
Why Tommy got closer.
Tommy got ten real PKD titles in-context, which provides strong stylistic/semantic keys for attention during generation (Vaswani et al., 2017 ). This is consistent with in-context learning evidence that prompt format and distributional cues carry much of the effect (Min et al., 2022 ; Olsson et al., 2022 ).
The output moved closer to PKD, but it still used a broadly reusable title template, with low-risk jargon.
(Image generated by Cursor)
---
Why Lee won.
Lee had to decompose themes and structures before writing the final title, so the model generated with richer intermediate context (Wei et al., 2022 ; Kojima et al., 2022 ). The “generate five, then pick one” step mirrors the logic behind self-consistency and improves candidate selection quality (Wang et al., 2023 ). Iterate, shortlist, next step. It’s algorithmic and understandable, both by humans and LLMs.
That extra structure produced the most PKD-specific output.
(Image generated by Cursor)
The lessons I got
1. Give examples when you need style fidelity. If I want a specific tone, voice, or format, I need to show examples instead of describing the style abstractly. This aligns with few-shot and in-context findings (Brown et al., 2020; Min et al., 2022) and with OpenAI’s and Anthropic’s show, don’t tell prompt guidance.
2. Make the model decompose before it generates. Ask for themes, constraints, or criteria before final output. Even simple Chain-of-Thought (CoT) scaffolding improves reasoning quality, and structured scaffolding usually helps more (Kojima et al., 2022; Wei et al., 2022).
3. Generate many, then self-select. Ask for several candidates and require a short selection rationale. This approximates self-consistency-style reranking at prompt level (Wang et al., 2023).
4. Keep prompts dense, not long. Remove filler and keep instructions explicit. This is consistent with official prompt-engineering recommendations from Anthropic and OpenAI, which both emphasize clarity and structured context over verbosity.
Epilogue: A “Happy Accident”
In the middle of this prompting frenzy, Cursor gave me an unexpected gift. A really lovely paperback cover of Lee’s PKD-like novel itself!
Not only the artwork resembles fiercely the style of covers of all those 1970s sci-fi novels, but it has exactly the grim and ethereal, with a bit of Kafkian (think The Adjustment Bureau or The Man in the High Castle) flavour I would expect in a novel like this. The ‘wife’ in a ghostly and abstract silhouette in a very mundane heavily bureaucratic, cold and claustrophobic landscape. Identity dissolving. I loved this gift.
When I inquired Claude / Cursor about the origin of this image (basically when did the model decided to generated this, since I don't think it was a spontaneous act of generosity):
Found it. Here’s what happened:
The cover image was *your idea*, indirectly. In your second message ever in this conversation, you said:
> “this experiment is part of a first post for ‘Clean Serve’... given that Cursor (you, in Cursor) can create subagents and generate images through NanoBanana now.”
You flagged the image generation capability as something worth showcasing. When the enrichment plan was written (line 17 in the plan file, visible in the transcript), it picked that up and added a specific todo:
> id: pkd-cover
> content: Generate a ‘lost PKD paperback’ cover image for ‘The Standard Procedure for Replacing Your Wife’ using NanoBanana
The plan’s reasoning (line 96 of the plan) was:
> “Image: Generate a ‘lost PKD paperback cover’ for the winning title using NanoBanana — this demonstrates the image generation capability you mentioned wanting to showcase.”
So you said “Cursor can generate images now” and the plan inferred: the best way to demonstrate that is to generate a 1970s paperback cover for the winning title. You never explicitly asked for a PKD paperback cover — but you set up the conditions (PKD experiment + image generation capability + “showcase it”) and the plan connected the dots.
A genuine happy accident born from your own setup.
Turns out that because I ‘teased’ the idea of doing this by showing the available resources and channels to materialize this fantasy (a ‘lost PKD book’), the intent behind this persisted by using said available resources to that extent.
A happy accident, indeed.
References
Papers
1 - Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention Is All You Need. NeurIPS. (https://arxiv.org/abs/1706.03762)
2 - Brown, T. B., Mann, B., Ryder, N., et al. (2020). Language Models are Few-Shot Learners. NeurIPS. (https://arxiv.org/abs/2005.14165)
3 - Wei, J., Wang, X., Schuurmans, D., et al. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS (https://arxiv.org/abs/2201.11903)
4 - Olsson, C., Elhage, N., Nanda, N., et al. (2022). In-context Learning and Induction Heads (https://arxiv.org/abs/2209.11895)
5 - Min, S., Lyu, X., Holtzman, A., et al. (2022). Rethinking the Role of Demonstrations: What Makes In-Context Learning Work? EMNLP (https://arxiv.org/abs/2202.12837)
6 - Wang, X., Wei, J., Schuurmans, D., et al. (2023). Self-Consistency Improves Chain of Thought Reasoning in Language Models. ICLR (https://arxiv.org/abs/2203.11171)
7 - Kojima, T., Gu, S. S., Reid, M., Matsuo, Y., & Iwasawa, Y. (2022). Large Language Models are Zero-Shot Reasoners. NeurIPS (https://arxiv.org/abs/2205.11916)
Documentation
Cursor. Subagents (https://cursor.com/docs/context/subagents)
Anthropic. Prompt engineering overview (https://docs.anthropic.com/en/docs/prompt-engineering)
OpenAI. Prompt engineering guide (https://platform.openai.com/docs/guides/prompt-engineering)