Edit for clarity and structure

“Make this clearer” gives an AI system broad permission. It may simplify vocabulary, but it may also remove a qualification, strengthen a claim, combine distinct ideas, or invent a bridge. Define clarity as observable changes and state what meaning must survive.

Edit in levels:

1. Document: Does the overall order serve the purpose?
2. Section: Does each section fulfill one promise?
3. Paragraph: Does each paragraph have a visible job and coherent sequence?
4. Sentence: Are actor, action, relationship, and qualification clear?
5. Word: Is each term precise, necessary, and appropriate for the audience?

Working from larger to smaller prevents polishing sentences that should be moved or deleted.

Before rewriting, ask for a map of the current draft:

Do not rewrite. For each section and paragraph, state its job in one sentence.
Identify repeated jobs, missing steps, unsupported leaps, and content that
does not serve the brief. Quote only short phrases as evidence.

Compare the map with your outline. A paragraph may contain three jobs because drafting mixed context, evidence, and recommendation. Split it if the reader must shift questions. Conversely, two short paragraphs may repeat the same claim and can be combined.

Inspect beginnings and endings. A section opening should orient the reader to its purpose. Its ending should complete the promise or create an accurate transition. Avoid a “conclusion” that merely repeats headings without answering the reader’s question.

A clear paragraph usually makes its controlling idea visible early. This does not mean every paragraph must use the same formula. It means the reader should not have to reach the final line to learn why the detail matters.

Check reference words such as “this,” “it,” “they,” and “the issue.” Replace them when more than one noun could fit. Keep related information together: place a condition beside the claim it limits and a source beside the fact it supports.

Use lists when readers need to scan parallel items, compare options, or perform steps. Do not turn every paragraph into bullets. Narrative explanation is better when relationships and reasoning matter more than retrieval.

Ask whether each example earns its space. Examples should illuminate a concept, reveal a consequence, or let readers apply a rule. Decorative examples can distract, and generated examples can quietly introduce unverified facts.

Give the model explicit invariants:

Suggest sentence-level clarity edits.
Preserve exactly: factual meaning, uncertainty, conditions, numbers, names,
technical terms, quoted text, ownership, and causal relationships.

Allowed: shorten sentences, replace unnecessary jargon, make actors explicit,
remove repetition, and reorder clauses.
Forbidden: add facts, strengthen certainty, remove exceptions, change blame,
or infer motives.

Return the revision plus a change log and any meaning-sensitive edits.

Compare source and revision side by side. A change log is useful, but the model can miss its own drift.

Prefer concrete actors and actions. “Implementation was delayed” hides who or what delayed it, although sometimes the actor is unknown or irrelevant. Do not invent one for grammatical neatness. Write “The implementation was delayed; the cause is still under review” when that is the verified state.

Shorter is not always clearer. “The committee may approve the proposal after legal review” must not become “The committee will approve the proposal.” Removing three words changed uncertainty and sequence. Editing quality is measured by reduced effort without lost meaning.

Build a terminology check. Define unfamiliar terms at first use, use one label for one concept, and avoid switching synonyms merely for variety. If “customer,” “account holder,” and “user” have different meanings, preserve those distinctions. If they mean the same person, choose one term.

Plain language favors familiar words, direct relationships, useful headings, and manageable sentences. It is not childish language or automatic shortening. Keep necessary technical precision and explain it.

Consider accessibility beyond reading level. Descriptive link text, meaningful headings, expanded acronyms, non-color cues, and text alternatives support more readers. If the document will be translated, avoid unexplained idioms and ambiguous references.

Readability scores, sentence length, passive-voice counts, and jargon flags can reveal patterns, but they are not quality verdicts. A safety instruction may need repeated exact wording. A legal condition may be long because its relationships matter. Optimize for the audience’s task, not a universal score.

After edits, run a reverse outline and read aloud again. Verify that every claim, condition, number, and source marker survived. Search for placeholders rather than assuming they disappeared because the prose looks smooth.

The author decides which suggestions to accept. AI can reveal possible friction, but it cannot experience the reader’s context. When possible, test the draft with a representative reader and ask what they understood, expected next, and had to reread.