Why the Same Prompt Gives You a Different Answer
Run the same prompt twice and you get two different texts. That is not a glitch — it is how the machine is built, and it quietly undercuts the claim that the prompt is the author.
5 cited sources
Prompting
Most prompt advice is folklore. Some of it isn't.
There is real published work on what changes model output and why. This pillar keeps the techniques that survive contact with evidence and drops the incantations.
Guides
Everything in this pillar
Briefs, follow-ups, formats, and the prompts that produce something you can actually edit.
-
-
Prompts Do Not Transfer
The 'perfect prompt' is the least portable thing you own. Change the task, the model, or even the formatting and its gains vanish. The method is what carries.
5 cited sources
-
Show, Don't Tell: Examples Beat Instructions
The copywriter's most common question — how do I teach AI my style? — has a boring, well-tested answer: paste examples of your writing, not a list of adjectives.
7 cited sources
-
Chain-of-Thought for Writers, Not Mathematicians
Chain-of-thought was invented and measured on math problems, then sold as a universal prompting trick. Here is the part that transfers to writing — and the larger part that doesn't.
5 cited sources
-
The Self-Critique Pass
Asking a model to review its own draft is the cheapest quality upgrade in prompting — and the only one with a hard number on facts: 7.5% to 77.5% accurate citations.
7 cited sources
-
The Brief Is the Prompt
A generic draft is not a flaw in the model — it is a faithful echo of generic input. Give a model no brief and it returns the statistical average of everything it has read, which is the definition of generic.
9 cited sources
-
Prompt Engineering Techniques That Measurably Improve AI Writing — and the Ones That Don't
A sourced field guide to the prompt techniques with measured effects on writing — examples, structure, iterative feedback — and the popular ones, like personas, that don't.
10 cited sources
What the research shows
Techniques with something behind them
Four prompting moves that are documented rather than merely popular.
Self-critique follow-ups
Explicitly asking the model to re-evaluate and verify its own output converts partially correct material into accurate material at a measurable rate. It is the single cheapest quality win available.
Chain-of-thought
Reasoning prompts activate a broader subset of feed-forward neurons, letting the model draw on more of what it learned during training — with the caveat that performance degrades under shifts in task, length or format.
Role and format constraints
Stating the professional context and enforcing a strict output format improves the usefulness of responses, and makes the output far easier to verify afterwards.
Automatic prompt optimization exists
Hand-tuning prompts is tedious enough that an entire research line automates it. Worth knowing before you spend an evening rewriting one instruction.
Common questions
Questions people actually ask
Why does my first draft always sound generic?
Because the brief was generic. Audience, goal, constraints and tone have to exist somewhere before the model can reflect them — if you do not supply them, the model supplies the statistical average of everything it has seen. That average is precisely what generic means.
Is there one prompt that works for everything?
No, and the research explains why: prompting performance degrades sharply under shifts in task, length and format. A prompt tuned for one job is not portable to another. What transfers is the method — brief, draft in layers, then critique.
Does telling the model it is an expert actually do anything?
Specifying professional context does show up in the literature as improving output usefulness, alongside enforcing a specific output format. It is a modest, real effect — not the magic switch the prompt-pack industry sells.