Because the prompt was generic. Models fill gaps with the most average answer available, so a vague request produces an average website. The fix is not a longer prompt or a secret phrase. It is being specific about the output you want and why you want it.
Anthropic’s prompting documentation puts this better than we can. It says to think of the model as a brilliant but new employee who lacks context on your norms and workflows, and that the more precisely you explain what you want, the better the result. That framing changes how you write almost immediately.
This is what the vendors themselves recommend, applied to the kind of work we do every day: writing copy, generating layouts, auditing pages, and reviewing code. None of it is clever. All of it is specific.
Clarity and constraints. Anthropic’s guidance is to be specific about the desired output format and constraints, and to give instructions as sequential steps when the order or completeness of those steps matters. Most weak prompts fail because they describe a topic rather than an outcome.
There is a test in that same documentation worth stealing outright. It calls it the golden rule: show your prompt to a colleague with minimal context on the task and ask them to follow it, and if they would be confused, the model will be too. We now run new prompts past someone else before blaming the model.
The example Anthropic gives maps directly onto web work. Asking for an analytics dashboard produces something forgettable. Asking for a dashboard and adding that it should include as many relevant features and interactions as possible, going beyond the basics to a fully featured implementation, produces something you can actually use.
Because the model can generalise from a reason but not from a rule. Anthropic’s documentation notes that providing context or motivation behind your instructions, such as explaining why the behaviour matters, helps the model understand your goals and give more targeted responses.
Their example is small and clarifying. Telling a model never to use ellipses works less well than telling it the response will be read aloud by a text-to-speech engine, so it should never use ellipses because the engine will not know how to pronounce them. The second version lets the model handle the cases you did not think to list.
We apply this constantly on content work. Instead of saying keep sentences short, we say this is for business owners reading on a phone between meetings, so keep sentences short and skippable. The model then makes sensible calls about vocabulary and paragraph length that we never specified.
A handful, and they should be varied. Anthropic’s guidance is to include 3 to 5 examples for best results, and to make them relevant to your actual use case, diverse enough to cover edge cases, and structured so the model can tell examples apart from instructions.
The diversity point is the one people miss. If all five of your examples are short product pages, the model learns that everything should be a short product page. Anthropic warns about exactly this, noting examples should vary enough that the model does not pick up unintended patterns.
OpenAI’s prompt engineering guide describes the same mechanism from the other side, saying the model implicitly picks up the pattern from those examples and applies it to a prompt. That is a useful reminder that examples are instructions, whether you meant them that way or not.
Separate the parts clearly. Anthropic recommends XML tags for this, noting they help the model parse complex prompts unambiguously when a prompt mixes instructions, context, examples, and variable inputs. Wrapping each type of content in its own tag reduces misinterpretation.
OpenAI’s guidance points the same direction with different syntax, suggesting markdown headers and lists to mark distinct sections of a prompt, and organising prompts into identity, instructions, examples, and context. The specific markup matters less than the fact that boundaries exist.
This is the single biggest improvement available to most people writing long prompts. A wall of text where the brand guidelines, the page brief, and the example copy all run together gives the model no way to tell which part is a rule and which part is reference material.
Because the model defaults to the middle of its training data unless you push it away. Anthropic’s documentation names this directly, noting that without guidance models can default to generic patterns that create what users call the AI slop aesthetic.
The specific tells it lists will be familiar to anyone who has looked at a lot of generated interfaces lately. Overused font families such as Inter, Roboto and Arial. Clichéd colour schemes, particularly purple gradients on white backgrounds. Predictable layouts and cookie-cutter component patterns that lack context-specific character.
The remedy in that same guidance is to be prescriptive about aesthetics rather than hoping for taste. Ask for distinctive typography instead of safe defaults. Ask for a committed colour direction with dominant colours and sharp accents rather than a timid, evenly distributed palette. Ask for atmosphere in backgrounds instead of flat fills.
Our own view is that this is why AI is a good accelerator and a poor art director. It will execute a direction well and it will not choose one for you. We wrote about where we do and do not let it near the work in our piece on using AI in a web design workflow.
Tell it to look before it answers. Anthropic’s guidance for reducing hallucinations in coding work is a prompt that says never speculate about code you have not opened, and that if the user references a specific file, the model must read the file before answering.
That instruction generalises far beyond code. The same shape works for content: never state a statistic you have not retrieved, and if you cannot find a source, say so instead of writing something plausible. Models are agreeable by default, and an explicit permission to say I do not know is worth adding.
None of this makes verification optional. It reduces the rate of invented detail, it does not eliminate it, and anything going in front of customers still needs checking by a person. That is the argument we make at length in our guide to fact-checking AI-generated content before you publish it.
Yes, and it costs you one sentence. Anthropic’s documentation says setting a role in the system prompt focuses the model’s behaviour and tone for your use case, and that even a single sentence makes a difference. It is the cheapest quality improvement on this list.
OpenAI structures this through message roles instead, describing developer messages as providing the system rules and business logic, like a function definition, separate from what the user asks. Different mechanism, same idea: standing instructions should live apart from the request of the moment.
The practical version for a web team is to keep your role and brand instructions in one reusable block and paste it above every task. Rewriting your tone of voice from memory each time is how a site ends up with eight pages that sound like eight different companies. Our take on where that goes wrong is in our article on whether you should let AI write your website copy.
More often than prompt guides admit. Anthropic’s own overview says not every success criterion or failing evaluation is best solved by prompt engineering, and gives the example that latency and cost can sometimes be improved more easily by selecting a different model.
The other common case is missing information. If the model does not have your pricing, your case studies, or your brand guidelines, no amount of rewording will conjure them. OpenAI’s guidance covers this as supplying contextual information, giving the model access to proprietary data or other data outside its training set.
Complexity is the third case. When a task keeps failing in different ways, the usual answer is that it is really several tasks. OpenAI recommends decomposing a request into all required sub-requests and confirming each is completed, which in practice means running three focused prompts instead of one ambitious one.
Pick the task you run most often and fix that one prompt properly. Anthropic’s documentation suggests starting with a clear definition of success criteria, a way to test against them, and a first draft prompt to improve. Without those, you are guessing about whether changes helped.
Then keep it. The mistake we see in teams is treating prompts as disposable messages rather than as working documents. The prompt that reliably produces good meta descriptions for your site is an asset, and it belongs somewhere your team can find it, not in one person’s chat history.
If you want help working out where AI genuinely speeds up your web work and where it quietly adds review time, we’re happy to talk it through. Reach out at phoenix.studio and we’ll give you an honest read on what is worth automating.
Tell us where you want to go. We'll tell you how we'd get you there.