Run one-change loops

Before round one, write the parts that should not move:

• the user’s actual goal;
• the source material and known facts;
• the intended audience;
• non-negotiable safety or privacy boundaries;
• the required decision or action;
• the success criteria.

Copy this core forward or explicitly say, “Keep the facts, audience, and constraints from the previous version.” Do not assume the assistant will perfectly preserve every detail. Review the result against the core after each revision.

Useful dials include scope, structure, length, reading level, tone, evidence handling, uncertainty, and level of detail. Change one dial with an observable instruction:

• “Keep all facts, but put the decision and owner in the first two lines.”
• “Keep the structure, but replace specialist terms with language a new employee can understand.”
• “Keep the content, but distinguish confirmed facts from assumptions in separate lists.”
• “Keep the recommendation, but add one credible alternative and its tradeoff.”

These are stronger than “be clearer” because you can inspect whether the requested change occurred.

Run five steps:

1. Observe: Identify the highest-impact gap.
2. Predict: State what one prompt change should improve.
3. Revise: Ask for that change while preserving the stable core.
4. Compare: Put the old and new outputs against the same criteria.
5. Decide: Keep, reject, or modify the change.

Write down a short result such as “Moving the decision first improved scanability but removed the reason; keep the new order and restore one evidence sentence.” Your next round now has a focused job.

Long chats accumulate instructions, examples, and corrections. Later directions may conflict with earlier ones. After several rounds, restate the current specification in one clean prompt or begin a fresh conversation with the selected source material. This reduces accidental dependence on obsolete feedback.

Version labels help:

• V1: original draft;
• V2: decision first;
• V3: decision first plus evidence sentence;
• V4: shortened to 120 words.

Do not overwrite the only copy. A later version can sound smoother while losing a crucial fact from V2.

The same prompt can produce different wording across runs. Do not conclude that a prompt edit caused improvement from one lucky output. For an important reusable prompt, test it on several representative inputs. A support-summary prompt that works only on neat tickets is not robust.

Hold the test cases stable and include difficult examples: missing dates, contradictory notes, emotional language, or unusually long input. The goal is not identical prose. The goal is consistently satisfying the requirements.

One-change loops are a learning method, not a rigid law. If an answer leaks personal information or fundamentally misunderstands the task, stop and rewrite the specification. Safety and correctness take priority over preserving experimental purity. Record that you changed several things because the baseline was unusable.

For routine low-risk work, speed may matter more than diagnosis. You can make a combined edit when the repairs are independent and obvious. The discipline still helps: name what changed so you can inspect it.