Controlling LLM Prose

Tutorial · Lesson 1

Your First Controlled Rewrite

From a competent draft to a controlled one

Who this is for: Anyone who has asked an LLM to “make it better” and gotten something vaguer.

What you’ll build: You will take one flabby paragraph and apply three precise controls in sequence, watching each one change a measurable property of the text.

Language models write competently by default, but competence is not control. In this lesson you will stop asking for vague improvement and start naming exact moves. We work on a single paragraph so you can see each change in isolation.

  1. 1

    Start with a baseline

    Here is our draft. It is grammatical, wordy, and blurry—typical default output. Read it once and notice how little sticks.

    Baseline draftThe implementation of the caching layer was undertaken by the team, and it was found that a reduction in latency was achieved, which is not uncommon in these kinds of situations.
  2. 2

    Apply de-nominalization

    First, revive the buried verbs. “Implementation of” becomes “implemented”; “reduction in latency was achieved” becomes “latency fell.” Agents reappear and the sentence shortens.

    Reference: De-nominalization

    After de-nominalizationThe team implemented the caching layer, and latency fell—which is not uncommon in these situations.
  3. 3

    Apply negation reduction

    “Not uncommon” is a double negative that costs the reader a beat. State it positively.

    Reference: Negation Reduction

    After negation reductionThe team implemented the caching layer, and latency fell—a common result.
  4. 4

    Apply BLUF ordering

    Finally, lead with the outcome. The reader should get the point before the backstory. This is bottom-line-up-front ordering.

    Reference: Bottom Line Up Front (BLUF)

    Final: controlledLatency fell after the team added a caching layer—a common, expected result.
Recap

Three named moves—de-nominalization, negation reduction, BLUF—turned a vague sentence into a crisp one. Notice you never said “make it better.” You named the mechanism each time, and the mechanism is what an LLM can reliably execute.