What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
A bare request usually returns a bare answer. Ask an AI model for “a social media post about my business” and you invite the most typical version of that post. Stephan Miller’s article “You’re Using AI Like a Vending Machine,” updated September 16, 2026, argues that the fix is not a longer list of instructions but a different kind of move: change the shape of the problem so the model has a less obvious path to follow. These moves are ways to get unstuck. They are not a proven formula for originality, and deciding what is good still falls to you.
Why a one-shot request gets a conventional answer
Miller’s core complaint is that most people use a chatbot like a vending machine. They insert a request, take whatever comes out, and leave. When the request carries little context, direction, or constraint, the answer tends to gravitate toward the common version of the task. This is Miller’s editorial argument drawn from his own working sessions. It is not a measured rate at which one-shot prompts fail.
![Claude AI for Beginners Bible: [5 in 1] The Ultimate Guide to Automate Your Work, Save Hours Every Week, and Use AI for Real-World Results](https://m.media-amazon.com/images/I/415+fSJacsL._SL160_.jpg)
As an Amazon Associate I earn from qualifying purchases.
A question asks for an answer; a move changes the problem
The useful distinction in Miller’s piece is between a question and a move. A question asks the model to produce the expected output. A move alters the task itself, so that the expected output is no longer the obvious one. Miller’s examples fall into six types, and each one changes the request in a different way.
Free tools Windows power users keep installed
One-click scans. No signup required.
1. Force two unrelated ideas into one frame
Pair the subject with something it has no natural relationship to, and ask the model to work within that pairing. Illustrative prompt for the bakery example below: “Write the post as a lighthouse keeper’s logbook entry about the morning’s bread.” The mechanism is unexpected association: the model must find a path between two domains instead of retrieving a bakery post.
#1 Best Overall
2. Add a constraint that rules out the obvious response
Name something the answer may not do. Illustrative prompt: “No exclamation marks, no emojis, and do not use the word ‘delicious’.” Constraints remove the first phrasings that come to mind, which are often the generic ones.
3. Break the problem into parameters
Decompose the task before asking for output. Illustrative prompt: “First list the one feeling a customer should have after reading, then the single fact they must remember, then draft the post around those two items.” Decomposition makes the choices visible, so you can change one parameter without rewriting everything.
4. Request perspectives from different positions
Ask for the same thing from several viewpoints. Illustrative prompt: “Write three versions: one for a skeptical regular customer, one for a new parent in the neighborhood, one for a rival shop’s marketing lead who is judging the post.” The mechanism is a perspective shift, which surfaces different angles on one subject.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute5. Use a random input to choose a direction
Introduce an arbitrary choice and build from it. Illustrative prompt: “Pick one word from this list at random, then build the post around it: harbor, ledger, orchard, anvil.” The randomness is supplied by a list you choose, which keeps the selection outside the model’s default preference.
6. Invert the objective
Reverse the goal and work backward. Illustrative prompt: “Write the post that would make someone least likely to visit my shop, then tell me which lines to avoid and which to keep.” Inversion often exposes assumptions that a direct request hides.
Miller also traces several of these methods to older idea-generation practice: forced connections, morphological analysis, deliberately bad ideas, arbitrary constraints, brainwriting, and defamiliarization. He notes that each can be tried on paper without any AI. The article is the source for this historical framing. The origin dates of each method and their effectiveness are not independently verified here.
Six moves at a glance
| Move | What it changes | Starting question to adapt |
|---|---|---|
| Forced connection | Pairs the subject with an unrelated domain | What would this look like in a domain I would never associate with it? |
| Arbitrary constraint | Rules out the obvious phrasing or structure | What is the first answer I would expect, and how do I forbid it? |
| Decomposition | Splits the task into parameters chosen before drafting | Which two or three choices control the result? |
| Perspective shift | Asks for the same output from different positions | Who else has a stake in this, and what would they notice? |
| Random input | Replaces the model’s default direction with a list you supply | What three unrelated words could set the direction? |
| Inversion | Reverses the goal and works backward | What would guarantee failure, and what does that reveal? |
The article does not rank these moves, and no comparative evidence in the reporting here shows that any one of them reliably beats the others. Choose by the mechanism your task needs: a stale tagline usually calls for a constraint or inversion, while a product with several audiences calls for perspectives.
A workflow for when the first answer is too ordinary
- Define the work you want and what a useful result looks like. Write this down in one sentence before you prompt.
- Add the context, constraints, or examples the task actually depends on, rather than sending a bare request.
- If the first answer is conventional, change the task with one move from the table above. Change one thing at a time so you can tell what helped.
- Use follow-up turns to critique, revise, or explore alternatives. Treat the first answer as a draft.
- Check factual claims and judge the result against the goal you wrote in step 1. Output varies between runs, and model versions can behave differently. OpenAI’s official documentation, as of October 2026, recommends evaluating prompt behavior as models change. Anthropic’s guidance starts with defining success criteria and running empirical tests before refining prompts.
These steps are practical advice drawn from Miller’s argument and current provider guidance. They do not guarantee quality, and longer prompts do not always produce better results.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What the 2026 diversity study does and does not show
Miller’s argument is often read alongside a 2026 academic paper on output diversity. Constantinos Karouzos, Xingwei Tan, and Nikolaos Aletras submitted it to arXiv on April 17, 2026 (arXiv:2604.16027). The authors analyzed three post-training lineages of the OLMo 3 model family, called Think, Instruct, and RL-Zero, across 15 tasks and four text-diversity metrics. Their abstract reports that where diversity is lost varies with the training data mix and the lineage, and that diversity collapse is determined during training by data composition and cannot be fixed at inference time alone.
That finding is specific. It concerns particular OLMo 3 lineages and the diversity metrics the authors chose. It does not show that every language model gives average answers, and it does not show that a user’s prompting technique can overcome training effects. Its most useful implication for a prompter is modest: some sameness in output may originate in the model itself, so the moves above help most when you are working within the range a model can produce.
Where the human stays in charge
Miller does not present these methods as a way to hand over judgment. He recounts model predictions that were confidently wrong, and the reader still has to verify claims and decide what fits the business, the audience, and the facts. He also describes one working session in which the gain was not speed: “The win wasn’t speed. There was no speed.” That is his account of a single session, not a general productivity measurement, and it should be read that way.
- Verify every factual claim the model makes about your business, prices, or dates before publishing.
- Keep your own taste as the final filter. A surprising draft is useful only if you can explain why it works.
- Note which move produced the best version, so you can reuse it on similar tasks.
Miller’s article is current as of its September 16, 2026 update, and its examples are his own. The moves are a way to ask better questions, and the answer still needs your check.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
