Prompt engineering

Prompt engineering is the practice of structuring the text sent to a language model, instructions, context, examples and formatting, so that its output meets a success criterion you defined in advance. It is an empirical activity, not a stylistic one: the same prompt run twice does not return the same text, so a change can only be judged against a test.

Anthropic's prompt engineering overview makes that precondition explicit. It assumes you have a clear definition of success, a way to test against it, and a first draft to improve. It also says plainly that not every failing evaluation is best solved by prompting; latency and cost are often fixed more cheaply by choosing a different model.

Which prompt engineering techniques are documented?

The published guidance from the two largest providers converges, which suggests these are properties of the systems rather than folklore.

• Be explicit and sequential. Anthropic recommends numbered lists or bullets when the order or completeness of steps matters.
• Use a few good examples. Anthropic calls examples one of the most reliable ways to steer format and tone, and recommends 3-5 that are relevant, diverse and structured. OpenAI's prompting guide gives the same advice: a handful of input and output pairs covering a range of inputs.
• Delimit the parts. Anthropic recommends XML tags such as <instructions>, <context> and <input>; OpenAI recommends Markdown headers or XML tags.
• Put long inputs first. For inputs of 20k+ tokens, Anthropic advises placing documents near the top, above the query and instructions.
• Order for caching. OpenAI advises putting content you reuse at the start of the prompt, for cost and latency.
• Separate roles. OpenAI describes developer, user and assistant messages as carrying different levels of authority.
• Chain when you need to inspect the middle. Draft, review against criteria, refine, when you need to log or branch on an intermediate result.

How much does prompt wording change an AI visibility number?

Completely, and our own data shows it. Attensira tracks questions for attensira.com on ChatGPT and Perplexity every day. Over the 30 days to 4 October 2026, prompts about the same brand gave opposite pictures.

"what is Attensira" named us in 66 of 66 answers. "best GEO tools" named us in 0 of 114. The table in the band below adds two more questions of each kind, and the pattern does not move: the questions that contain our name were answered with our name in 197 of 198 answers, and no category question named us once.

Neither number is wrong. The first kind measures whether engines know us when asked by name; the second kind measures whether they reach for us when a buyer asks about the category. Report only the first kind and we look universally known; report only the second and we look invisible. That is why an AI visibility number without its prompt list is not a number, and why two tools can disagree about one brand in the same week without either being wrong.

The practical rule follows: treat a prompt set like a schema. Write it down, version it, change it deliberately, and never compare two windows across an edit, because the denominator changed meaning, not just size.

How should you write prompts for tracking AI answers?

The way a buyer types them, not the way your marketing team talks. Attensira's prompt documentation puts it as a contrast: "What's the best expense tool for a 20-person agency?" is a prompt; "expense tool agency" is a keyword, and no one asks it. It also advises leaving your brand out of most prompts, because an unbranded question is the honest test of whether a model reaches for you unprompted.

Language matters too. When a workspace reads a market in another language, Attensira translates each prompt once into that language and keeps your brand, domain and competitors' names exactly as written, because a buyer in Germany asks in German and gets a different answer. A translation that fails waits; it is never asked in the original language under the German label.

What goes wrong with prompt engineering?

• Tuning without an eval. One better-looking output is one draw, not evidence.
• Prompting around a retrieval failure. If the document never entered the context window, rewording cannot surface it. That is a retrieval problem.
• Overstuffing the window. Accuracy and recall degrade as a window fills.
• Confusing compliance with correctness. A well-structured prompt makes output well-formed, not true.
• Treating retrieved text as instructions. Content pulled in from the web can carry instructions of its own; see prompt injection.
• Silently changing a tracked prompt set. The most common measurement error in this field.

How does Attensira handle prompts?

It keeps the instrument stable and visible. You paste your buyers' questions, the crew files each one under a topic and you confirm before anything is tracked; every prompt then runs on each engine you track once a day, with the same wording, and its history starts the day it was added rather than being backfilled. Every rate shows the prompts behind it and the sample size, so a number can always be traced to what was asked. See prompt tracking.