The AI Editorial Bots: Copy Editors That Won't Cross the Line

Most AI writing tools have the same problem: they want to write for you. Ask for feedback and they rewrite your sentence. Ask about a scene and they pitch a different one. For a novelist, that's not help — it's erosion. The voice is the whole job, and a tool that quietly replaces it is worse than no tool at all.

UnboundProse takes the opposite stance. Its first AI features are editorial bots — opt-in, paid, and deliberately constrained to the kind of work a meticulous copy editor does: catching objective, technical issues and staying entirely out of the creative decisions. The design principle is one sentence: AI as a copy editor who only speaks when it catches a genuinely valuable technical problem.

What they catch (and only this)

The first bot shipped is the POV Consistency bot. Open a chapter, click the bot, and it checks for:

• Interiority slips — prose that accesses the thoughts or sensations of a character the point of view doesn't have access to ("Marcus felt a flash of doubt" when the chapter is third-limited on Sarah).
• Tense violations against the chapter's declared tense.
• Voice-rule violations when the author has set voice rules in the manuscript metadata.

The second bot, Chapter Opening & Ending, scores the first and last paragraph of a chapter against a small craft rubric — immediate engagement, character motion, forward momentum for openings; consequence, resonance, and earned meaning for endings — on a 1–3 scale. For every score it quotes the actual prose, names the craft principle behind the score, and shows what would strengthen it. It's pedagogical: the author should walk away having learned something, not just been graded.

What they are forbidden from doing

The constraint is as important as the capability. The bots never:

• Suggest plot or story direction
• Recommend character development
• Critique prose style or "engagement"
• Rewrite the author's sentences

That line — technical and objective on one side, creative and subjective on the other — is the product. It's what lets a nervous author turn the bots on without feeling like they've invited a co-writer they didn't ask for.

The two-phase pattern: say what you can't do first

The most important interaction-design decision was making the bots declare their own limits before they act. Every bot runs in two phases, driven entirely by its system prompt — no state-machine plumbing, because a capable model follows "echo first, then wait" reliably.

Phase one is a goal echo. On the first turn the bot does not analyze anything. It states plainly what it will check, and — critically — what it cannot. The POV bot says, in effect: "I'll check this chapter for interiority slips, tense breaks, and voice violations. But I only see this one chapter. If a character knows something they 'shouldn't,' I can't actually tell — they may have learned it in a chapter I can't see." Then it stops and asks the author to confirm or refine.

Phase two is the scan. Once the author says go, the bot quotes exact passages, explains in one line why each might be an issue, and suggests a fix — while skipping prose that's merely unusual but valid.

Stating the limitation up front is a trust feature. An author who's been told exactly what the tool can't see trusts the things it does flag. A tool that over-claims gets switched off the first time it's confidently wrong.

Confidence over volume

The bots are tuned to under-report. A finding only surfaces when the bot is genuinely confident it's a real, objective issue — a false positive costs more trust than a missed nit. The author can dismiss any finding, ask the bot to ignore a scene, or have it re-scan after an edit. The goal is a tool the author reaches for, not one they tolerate.

Why this matters beyond fiction

The editorial bots are a small, concrete answer to a large question every product touching AI now faces: where is the line between assisting and replacing? Drawing that line explicitly — encoding it in the system prompt, surfacing it in the UI, and designing the interaction so the AI announces its own boundaries — is a transferable pattern. It's the same instinct behind the platform's snapshot firewall and its bring-your-own-key model: the author stays in control, and the system is honest about what it is.