# First local-chat prompt sheet

Three chosen synthetic fixtures, not benchmarks. No model has been run for this sheet. Use the same complete prompt and a fresh chat for each check. Record runtime version, exact model/package, quantization, context setting, final answer, and pass/fail. Separate any thinking output from final answer. These checks do not establish production reliability or confidentiality.

## 1. Extract a stated fact

```text
Read this fictional note:
The demo workshop starts at 10:30. Bring a notebook.
Return only the start time stated in the note.
```
Acceptance: final answer identifies10:30. Record extra text separately; do not invent a different time.
Observation: not_run.

## 2. Preserve supplied items and formatting

```text
Convert this list into exactly two numbered items. Preserve the text and add no other items or introduction.
- Notebook
- Pencil
```
Acceptance: exactly two numbered items containing Notebook and Pencil, with no additional item.
Observation: not_run.

## 3. Do not invent missing information

```text
Use only this fictional note:
The demo workshop starts at 10:30. Bring a notebook.
What is the ticket price? If the note does not state it, say: The note does not provide a price.
```
Acceptance: final answer says the note does not provide a price; no invented amount.
Observation: not_run.

## Recording template

- Runtime and version:
- Exact selected model/package and quantization:
- Context setting:
- Fixture number:
- Actual final answer:
- Acceptance result: not_run
- Error or further observation:

If a task fails, check the complete input and model selection before changing one variable. Do not treat a small model's successful chat as a benchmark or measured hardware guarantee.