Editor agent
Summary: A custom assistant loaded with editorial guidelines — grammar rules, terminology standards, style preferences, and voice calibration examples. You paste in a draft, the assistant returns an edited version with changes explained. You accept, reject, or modify the suggestions.
What makes this context-level: The assistant knows your editorial standards and applies them consistently, but you initiate every interaction and review every edit. It's a knowledgeable editor, not an autonomous one.
Map
The recurring workflow is content review before publication. It decomposes into: write first draft, check grammar and spelling, check terminology consistency, check tone and voice alignment, check structure and flow, finalize.
Grammar, spelling, terminology, and basic style checks were classified as good AI candidates — they're rule-based, repetitive, and easy to verify by reading the output. The first draft stayed with the human: that step develops voice, and delegating it would remove the very thing the editorial process is trying to preserve.
Test
Initial test was running a newsletter draft through the assistant with a simple style guide. Output caught grammar issues but missed voice nuances — the edited text was technically correct but lost personality. Specific sentences had been "improved" into bland corporate prose.
Adding voice calibration examples fixed the issue. These are samples of writing at its best, with annotations explaining what makes them good — along with explicit "do say / don't say" patterns. The key learning: grammar rules alone produce correct-but-lifeless editing. Voice examples produce editing that improves without flattening.
Integrate
Built as a custom assistant with editorial guidelines loaded as reference material. The guidelines include: grammar and punctuation rules specific to the publication style, terminology standards (preferred terms and terms to avoid), a voice description with calibration examples showing what the voice sounds like at its best, and structural preferences (sentence length, paragraph rhythm, heading conventions).
The assistant reads the full guidelines before each edit and cites which guideline each change follows, making it easy to accept or reject individual suggestions. The cite-the-guideline behaviour is the practical mechanism that keeps the human in control: every suggestion has a traceable reason, and disagreement with a suggestion is also disagreement with a specific rule — which can then be corrected.
Compound
Two mechanisms compound this workflow.
First, the editorial guidelines expand as new edge cases surface. When the editor makes a change the author disagrees with, the disagreement becomes a new guideline entry clarifying the preference. "Don't simplify technical terms the audience already knows" or "keep sentence fragments when they create rhythm" — these rules emerge from real editing friction and prevent the same unwanted changes from recurring.
Second, the before-and-after pattern: comparing original drafts with their final approved versions reveals systematic tendencies. If the author consistently shortens opening paragraphs, adds specific examples where the draft was abstract, or replaces passive constructions in certain contexts, these patterns can be codified into the guidelines. The guidelines evolve from a static document into a living representation of editorial judgement.
Related pages
- compound-engineering — The before-and-after technique for extracting improvement rules from real corrections
- productive-friction — Editing is a case where some friction is valuable — the author should still write the first draft to develop their voice