Set the voice and the reading level

“Simple” is not a specification

Ask for something “simple” or “for beginners” or “in plain language” and you get the model’s idea of simple, which is a slightly shorter version of its normal voice. The normal voice is the average of the internet: sentences around 20 to 25 words, a fondness for “however” and “additionally”, abstract nouns, a formal register that makes a Class 6 reader stop halfway down the page.

Simplicity has to be specified as things that can be counted or checked.

• Sentence length. “No sentence longer than 15 words.” Measurable, and the single most effective instruction.
• Word frequency. “Use only words a 10-year-old would hear at home or in school.” Or, more precisely: “Avoid any word with more than three syllables unless it is a subject term, and define every subject term the first time.”
• Idiom. “No idioms, no metaphors unless you explain them.” Idioms are the thing second-language readers trip on hardest, and the model uses them without noticing.
• Concrete over abstract. “Every abstract idea gets a concrete example in the next sentence.”
• Person. “Address the reader as ‘you’.” Second person is easier to read than the impersonal passive.

Put four of those in a prompt and the output changes visibly. Put “write simply” and it does not.

Setting a reading age

Readability formulas (Flesch-Kincaid is the common one) estimate a reading level from average sentence length and syllables per word. They are crude, they were built for English, and they do not measure whether the ideas are hard. But they give you a number, and a number is something the model and you can both aim at.

“Write this at a Flesch-Kincaid grade level of 5” works reasonably well in English. Then check it yourself: Microsoft Word shows readability statistics under the spelling options; free web tools such as the Hemingway Editor show sentence length and grade level as you paste. If the model says grade 5 and the tool says grade 9, believe the tool.

For languages other than English the formulas do not apply, and the model’s sense of “simple Hindi” or “simple Kiswahili” is weaker than its sense of simple English, because it has read less graded material in those languages. There, the sentence-length and idiom instructions still work; the numeric target does not.

Why it drifts

The check question describes the most common failure. Paragraph one obeys; paragraph four does not. The mechanism is the one from the first block: each sentence is predicted from the sentences before it. Your instruction is at the top; the last three sentences are immediately before. Once one sentence runs slightly long or slightly formal, the next is predicted from a long formal sentence, and the register slides back toward the default. The instruction has not been forgotten. It has been outweighed.

Two fixes:

1. Ask for shorter pieces. Two hundred words at a time drifts less than eight hundred. Generate the explanation in sections and check each.
2. Restate the constraint inside the request. “Every sentence under 15 words, including in the last paragraph.” Naming the place where it usually fails measurably reduces the failure.

Register is separate from level

Two different things get confused here. Reading level is about how hard the text is to decode. Register is about who it sounds like. You can want a low reading level and a formal register (an exam instruction), or a high level and an informal one (a note to a strong Class 12 student).

Describe register with a person, not an adjective. “Write as a firm, kind teacher would speak to a nervous student before an exam” produces something usable. “Write in a supportive tone” produces the generic warmth of a greetings card. The person carries a whole bundle of choices (sentence rhythm, directness, what to leave unsaid) that the adjective does not.

Check it as the weakest reader

Before printing, read the text as the student in your class who reads slowest. Where would she stop? Which word would she not know? Which sentence has a clause in the middle that loses the subject? Mark those places and send them back: “Sentences 3, 7 and 12 have a clause in the middle. Split each into two.”

This is the second-message habit applied to voice. It takes three minutes and it is the difference between a text that reaches the whole class and one that reaches the top third.

The honest limit

The model can make language simpler than the ideas it carries. A concept that needs a prerequisite the student does not have will read easily and land as nothing. A reading age of 8 on the explanation of photosynthesis does not make photosynthesis an 8-year-old’s concept. Level of language is yours to set; level of idea is a teaching decision, and no readability score sees it.