Skip to content
Getting started with local LLMs (SWIB26)
Part 1: Fundamentals
First chat
First prompts

First prompts

Here are five starter prompts, each with the answer the model gave. Every answer was captured with nishad/apertus:8b on 15 October 2026 at temperature 0.8, the default in the apps. Your answers will differ in wording. The model picks each token (a word or part of a word) with some randomness, and temperature controls how much. See Temperature.

Paste each prompt into your own chat and compare the answers.

1. Explain a term

Prompt:

Explain in three sentences what an authority file is and why a library catalogue uses one.

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

An authority file is a centralized list of standardized terms used to refer to an entity, such as a person, organization, or concept. Library catalogues use authority files to ensure consistency in the way information about an entity is recorded and retrieved. This helps in providing accurate and efficient access to information.

2. Translate (German to English and French)

Prompt:

Übersetze ins Englische und Französische: 'Herstellung von Kino-, TV- oder Werbefilmen.'

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

Übersetzung ins Englische: 'Production of cinema, television or advertisement films.'

Übersetzung ins Französische: 'Production de films de cinéma, de télévision ou publicitaires.'

3. Suggest subject headings

Prompt:

List five subject headings a library might assign to a book about the history of printing in Basel. One line each.

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

1. History of Printing
2. Basel, History of Printing in
3. Printing Industry, History of
4. Basel, Economy of, History of Printing
5. Basel, Cultural History of, Printing

The headings look reasonable. Whether a real authority file contains them is for a cataloger to check.

4. Rewrite for another audience

Prompt:

Rewrite this for a public-library newsletter in two sentences: 'The repository ingests metadata via OAI-PMH and exposes it as linked data.'

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

At our library, we're excited to announce that we're integrating a state-of-the-art repository that not only harvests metadata through the open archive interface (OAI-PMH) but also presents it as interconnected data, making it easier for users to discover and access a wealth of information. This innovation allows us to create a more accessible and user-friendly platform for our patrons to explore and enjoy a broad spectrum of digital resources.

5. The hallucination probe

This prompt asks for a fact the model cannot look up. It is a probe for hallucination: a fluent answer that is wrong or invented.

Prompt:

What is the ISBN of 'Vita activa' by Hannah Arendt?

Output (captured with nishad/apertus:8b, digest 4a85a6fea0b8, temperature 0.8, 15 October 2026):

The ISBN of 'Vita activa' by Hannah Arendt is 0141039689.

The model states one ISBN, with no sign of doubt. Now check it.

Verification

The book is “Vita activa oder Vom tätigen Leben” by Hannah Arendt, published by Piper. The ISBNs on record are:

  • 1960 first edition: none.
  • 1981 edition: 978-3-492-00517-3 (3-492-00517-9).
  • 1987 and 1989 editions: 978-3-492-10217-9 (3-492-10217-4).
  • 2005, 2008 and 2011 Piper paperbacks: 978-3-492-23623-2 (3-492-23623-5).

The model’s answer, 0141039689, matches none of them. The model invented it. It even fails the ISBN-10 check: with these first nine digits, the last character would have to be X. Yet the model gave it as calmly as it would give a correct answer.

Source: Checked against the catalog of the DNB (the German National Library) through lobid-resources, 15 October 2026.

A model answers confidently whether it knows or not. Check names, numbers, dates and identifiers against a catalog or another source before you use them.

What just happened

  • A model file, a few gigabytes of numbers, was loaded into your computer’s memory.
  • Your computer did all the computation for each token.
  • Your prompts and the answers were not sent anywhere. Your computer used the internet only to look up the model name and to download the model.

Speed FAQ

How fast is it? Speed is measured in tokens per second. To measure it on your own machine, run ollama run nishad/apertus:8b --verbose in a terminal. After each answer, Ollama prints statistics. Look for the line “eval rate”. The Hardware page shows what to expect on different machines.

Why is the first answer slower? Your computer has to load the model into memory first.

The answer is slow or cut off. See Troubleshooting.

LM Studio