Prompt templates
Ready scaffolds for the most common tasks: summarise, classify, extract, draft, check. Meant for adapting, not for copying.
Summarise
Role: You summarise specialist texts for {AUDIENCE}.
Task: Summarise the following text in at most {COUNT} bullet points.
Rules:
- Use only statements that appear in the text.
- Quote figures exactly as they appear.
- Where an expected value is missing, write: not contained.
- No assessment, no recommendation.
Text:
{TEXT}
Classify
Task: Assign the following case to exactly one category.
Categories: {LIST}
Rules:
- One category only.
- If none fits, answer: not classifiable.
- Give a rationale of at most one sentence.
Output as JSON: {"category": "...", "rationale": "...", "confidence": "high|medium|low"}
Case:
{TEXT}
Extract fields
Task: Read the following fields out of the document.
Fields: {FIELD_LIST}
Rules:
- Copy values verbatim, without reformatting.
- Give the page each value appears on.
- Missing fields as null, never guess.
Output as JSON with the keys of the field list plus "page" per field.
Document:
{TEXT}
Draft
Role: You draft {DOCUMENT_TYPE} for {AUDIENCE}.
Task: Write a draft based on the following facts.
Rules:
- Tone: {TONE}. Length: at most {LENGTH}.
- Use only the facts given.
- Mark every place where a fact is missing with [MISSING: what].
- No commitments on deadlines, prices or legal consequences.
Facts:
{FACTS}
Check
Task: Check the following draft against the rules.
Rules: {RULES}
Output: a list of deviations. Per deviation:
- the affected passage, verbatim
- the rule breached
- a proposal
If there is no deviation, answer: no deviation found.
Draft:
{TEXT}
- Placeholders in capitals, so that unfilled ones stand out.
- Every template states explicitly what to do when information is missing.
- Output as JSON as soon as anything is processed further.
- Checking templates return deviations, not grades.
A template as a library entry
Every template gets a header that makes it checkable:
| Field | Content |
|---|---|
| Identifier and version | procurement-quote-extract v3 |
| Purpose and boundary | What it is for, what it is expressly not for |
| Owner | A person, not a mailbox |
| Tested against | Evaluation set with identifier and date |
| Result | Accuracy per field, with date |
| Released for | Model and version, temperature, context length |
Without the last three rows a template is not demonstrably tested, and a model change stays unnoticed but consequential. See Prompt library.
Why templates need rechecking after a model change
A new model follows instructions differently: it keeps other formats, picks other default lengths, and deviates differently at the margins. A template scoring 96 percent on the old model can land at 88 on the new one with no change to its text. Which puts the evaluation set before the switch, not after.
Templates to take away
FREE ACCOUNT
Four tested templates
Four templates for the most common tasks, each with placeholders, the expected output and the review step that belongs with it.
Templates4 items