Pro RSS Prompts: build reusable instructions you can evaluate

Give every instruction a clear role

Start with the task and audience, then describe the supplied material. Set the evidence rule separately: use only these entries, identify the supporting source labels, and mark missing information. Finish with the output contract. A digest might require three short sections; a comparison might need a table with one row per source.

Keep variable material outside the reusable instructions. Put article text in a clearly labeled source section and preserve its title, URL, and date as evidence fields. This makes it easier to update a template without accidentally changing the material being evaluated.

Write one small, testable template

Audience: developers maintaining a feed publisher.
Task: identify changes requiring action.
Evidence: the labeled entries supplied below.
For each finding return:
- the change and affected component
- supporting source labels
- an action, or "no action established"
- information still needed
Never infer an omitted version or deadline.

The OpenAI prompt engineering guide provides official guidance on giving models instructions and relevant context. Adapt those principles to your own task instead of collecting elaborate wording with no measured benefit. A shorter prompt that consistently preserves attribution may be easier to maintain than a long catalogue of exceptions.

Evaluate changes against real difficulties

Keep a small fixture set containing a straightforward announcement, a duplicate entry, conflicting dates, an excerpt that omits the conclusion, and an item that requires no action. Write the expected behavior before running a revised prompt. You need not prescribe exact sentences: require support, correct attribution, completeness, and a visible acknowledgment of uncertainty.

Change one important instruction at a time. Record the template version, fixture set, resulting drafts, and reason for accepting the change. If a revision improves clarity but introduces an unsupported deadline, the tradeoff should be visible rather than hidden by a fluent example.

Make review part of reuse

Recheck templates when the source fields, task, or model configuration changes. A prompt written for full articles may mislead when a feed begins supplying excerpts. The LLM evidence workflow helps define that boundary. Retire templates whose purpose has disappeared, and keep one named owner responsible for the examples and acceptance criteria.