Name the editing layer
Most bad AI editing starts with an underspecified command: make this better. Claude then optimizes for a blended average of clarity, smoothness, politeness, and generic correctness. That may destroy the thing you care about.
Instead, name the layer. Developmental editing asks whether the argument, story, or document architecture works. Line editing asks about rhythm and emphasis. Copy editing asks about grammar, consistency, and house style. Proofreading asks for last-mile errors.
Never ask for all four unless the document is low stakes. For serious work, the layers should be separate because they produce different kinds of judgment.
- Developmental: structure, stakes, missing sections, order, reader promise.
- Line: cadence, emphasis, transitions, sentence force, needless repetition.
- Copy: consistency, grammar, spelling, terminology, style-guide adherence.
- Proof: typos, punctuation, formatting, broken links, final defects.
Use tracked changes or a change table for accountable edits
For documents in Microsoft Word, Anthropic documents Claude for Word as a beta add-in that can edit selected text, preserve surrounding formatting, and use tracked changes mode. The same page also warns that human review is required for client-facing and audit-critical work.
If you are not using the Word add-in, ask Claude for a change table: original, revised, reason, risk. This is slower than a silent rewrite but makes editorial judgment visible.
The reason column is the most important part. It lets you reject edits that merely polish away voice.
Always review tracked changes.
Give Claude a rubric before you ask for judgment
Claude can critique a draft against a rubric with more discipline than against taste words alone. A good rubric has five to seven dimensions and describes what excellence looks like in your context.
For a reported essay, the rubric might include thesis, source quality, claim support, structure, compression, quote integration, and reader value. For a sales page, it might include promise, proof, objection handling, specificity, compliance, and conversion path.
Ask Claude to score first, then explain the top three fixes, then rewrite only the highest-leverage section. This avoids a full rewrite that conceals the editorial diagnosis.
- Define the rubric.
- Ask Claude to score the draft against it.
- Ask for the three most important fixes.
- Approve a focused rewrite.
- Run a human final read.
Keep final judgment human
Claude can make language cleaner, but clean is not always better. Some sentences need friction. Some essays need a strange turn. Some copy needs legal restraint rather than persuasive heat.
The final pass should check whether the draft still sounds like the intended author and still makes only claims you can defend. If Claude made the piece more generic, restore specificity before publishing.
For high-stakes legal, financial, medical, or client-facing documents, treat Claude as an assistant, not an authority. Human expertise and verification remain the gate.
FAQ
Is Claude good at proofreading?
Claude can catch many errors, but proofreading still benefits from a second tool or human pass because language models can miss small defects or introduce changes.
Can Claude preserve my style?
It can preserve style better when you provide examples, project instructions, and explicit do-not-change rules.
What is the safest edit prompt?
Ask for diagnosis and a change table before accepting a rewrite.
Primary sources
Official beta documentation for Word add-in editing, tracked changes, limitations, and human review guidance.
Claude Help Center Create and edit files with Claude 2026Official examples for document creation, reports, file conversion, and analysis workflows.
Anthropic Docs Prompting best practices Reviewed 2026-07-06Official guidance on clarity, context, examples, XML tags, roles, and long-context prompt structure.
Claude Help Center What are artifacts and how do I use them? Updated 2026Official description of artifacts as a separate surface for substantial standalone content.