The Working Set Prompt: Keep LLM Outputs Consistent Across Multi-Step Work

Intro

Large context windows are tempting; the Working Set Prompt keeps only the relevant pieces so outputs remain stable and focused across steps.

How it works

1. Define a small set of facts (5-10 items).
2. Include only those facts in every prompt as a concise reference.
3. Update the working set explicitly when facts change.

Example

Working set for a feature: product_goal, user_story, constraints, current_state, next_steps.

Why it helps

Reduces drift, makes outputs reproducible, and shrinks token usage.

Wrap-up

Choose one recurring workflow and introduce a working set next time you prompt the model.