Protect Narrative Voice Before You Polish

Protect Narrative Voice Before You Polish

A revision can fix every grammatical rough edge and still leave you with a weaker narrator. The facts survive; the person telling them disappears.

Before asking an AI tool—or a human editor—to improve a passage, decide which choices the revision must preserve. Then name the reader problem you actually want solved.

The news prompt, and the limit of the claim

In Camila Osorio's October 18, 2026 interview for El País, Julián Delgado Lopera discusses Spanglish, grammar, and a writing practice shaped by sound and rhythm. That interview appears in today's World Literature Daily.

It raises an editorial question for AI-assisted revision: what happens when an instruction to make prose more fluent also removes deliberate texture?

The method below is our proposal, not Delgado Lopera's advice. We have not tested an AI model on his fiction. Our examples are independently invented, entirely in English, and do not imitate his prose or represent a linguistic community.

1. Separate an obstacle from a choice

Start with the intended experience. A reader may need to understand who is leaving a room while remaining uncertain about why. A repeated phrase may slow a paragraph because the narrator cannot bear to finish the thought. Those are different revision problems.

Use two questions:

• Obstacle: What is the reader failing to understand that I intended them to understand?
• Choice: What is the prose doing that I would miss if it became more conventional?

Do not declare every irregularity expressive. If you cannot explain what a feature contributes, mark it for testing rather than automatically protecting or deleting it.

2. Write a short voice-preservation brief

Choose three to five observable features from the passage itself. Avoid vague instructions such as “keep it authentic.” Give the editor something it can point to.

These are diagnostic prompts, not universal rules. A narrator may deliberately obscure the actor; a relationship may deliberately shift. State that intention before requesting a repair.

3. Repair one obstacle with the smallest useful change

Consider this original miniature:

I kept the chair. Not for him. For the scrape it made when he came home.

A generic smoothing pass might produce:

I retained the chair because the sound it made reminded me of his return.

The second version is easy to parse. It also removes the denial, collapses three beats into one explanation, and makes the speaker sound more detached. That could be right for another narrator. It is a poor match if the broken rhythm is the point.

Now suppose the actual scene contains two chairs and readers cannot identify this one. A narrower repair is:

I kept the kitchen chair. Not for him. For the scrape it made when he came home.

One identifying word resolves the practical uncertainty without deciding that the fragments must go. This is an editorial comparison written for this article, not a recorded model result.

4. Ask for a diagnosis before a rewrite

Use this prompt with a passage you own or have permission to process:

Intended reader and scene context: [describe briefly]
Reader problem to investigate: [one specific obstacle]
Choices to preserve: [three observable features]

Identify at most three places where the stated obstacle occurs.
Quote the shortest relevant phrase from the supplied passage.
Explain what a reader might misunderstand and why.
Do not rewrite the whole passage or infer errors from unusual grammar alone.
Offer a minimal repair and a leave-as-is option for each finding.
If the evidence is weak, say so. I will decide which changes to keep.

An AI response is a set of suggestions. Check that each quoted phrase actually exists, that the diagnosis matches the intended reader, and that the proposed repair does not change the event or emotional stance. Reject invented problems.

For multilingual work, do not ask a model to “correct” unfamiliar usage without knowing its role. Where cultural or linguistic judgments matter, seek a reader or editor with relevant knowledge. A general-purpose model's confidence is not that expertise.

5. Compare the result aloud and in context

Read the original, the minimal repair, and the surrounding paragraph aloud. Listen for where the speaker breathes, hesitates, insists, or retreats. Then check the physical action on the page. Sound does not excuse a repair that changes who did what.

For each change, record one outcome:

• Keep: It solves the named obstacle and preserves the choices you marked.
• Reject: It solves a different problem or erases a necessary effect.
• Test: You cannot yet tell; ask a fresh reader a specific question.

Instead of asking whether the passage is better, ask: “Which chair do you picture?” and “What do you think the speaker is refusing to admit?” A useful edit can improve the first answer while leaving the second open.

A 15-minute exercise

Take 120–180 words from your own draft. Spend three minutes marking three voice features and one obstacle. Spend four minutes producing a minimal repair, with or without AI. Spend four minutes comparing versions aloud and in their paragraph context. Use the remaining four minutes to record what you kept, rejected, or need a reader to test.

The deliverable is one justified revision and a reusable brief. Speed, fluency, and a higher automated score are not substitutes for that judgment. This exercise has not been evaluated as a performance claim.

Try the workflow in DeepCrafter

If you use DeepCrafter, consult the user guide to choose an editing workflow that fits the passage. Keep your preservation brief beside the text, review suggestions individually, and retain the final editorial decision. This article does not claim that DeepCrafter automatically detects or preserves narrative voice.

Next action: Write three choices you want the next edit to preserve before requesting any rewrite.

Source and editorial note

• Camila Osorio's interview with Julián Delgado Lopera, El País, October 18, 2026. Interview used as the news prompt; accessed October 20, 2026. The workflow, miniature, prompt, and exercise above are DeepCrafter editorial material, not attributed to the interviewee.
• DeepCrafter user guide, for choosing a product workflow rather than treating this article as a feature guarantee.

Prepared with AI assistance and editorial source checking for DeepCrafter Insights. Published October 20, 2026. No model benchmark or reader study is reported.