A shortlist of deterministic editing tools is short by nature, because the category is defined by what it excludes: anything that guesses. What remains is exact within its scope, which is a real advantage for the checks where completeness matters and no help at all for the ones that require understanding.
The shortlist by job
Named products rather than categories, because a category cannot be bought. Where a product carries optional generative features that can be left switched off, that is said rather than glossed over.
| Job | The named options | What it is exact about |
|---|---|---|
| Spelling and grammar | LanguageTool, whose core checking is rule-based and whose paraphrasing you can leave alone, or your word processor's own checker with its assistant features switched off | Mechanics, sentence by sentence. Every invented name in the book will be flagged as a mistake until you teach it one at a time |
| Editorial consistency and house style | PerfectIt, run against a style sheet you define | Capitalisation, hyphenation, abbreviations, lists and tables. Built for professional editing, so it is noisy on fiction that breaks rules deliberately |
| Readability and sentence habits | The Hemingway Editor desktop application, which uses fixed criteria and does not carry the Plus features | Sentence length, adverbs and passive constructions. Exact, and of debatable use in fiction |
| Word frequency and repetition | The frequency counts inside most editing applications, or a search in your word processor | Counting. There is no judgement in it to get wrong |
| Names and invented terms | The Novel Kit | Every proper noun and unusual term the manuscript contains, with counts and first appearances |
| Formatting and file checks | Desktop formatters, plus an EPUB validator on the finished file | Validation is exact. Whether a given formatter has since added generative features is a claim to check product by product |
| Continuity of meaning | None | Not available |
| Prose quality | None | Not available |
The last two rows are the honest boundary of the category, and they are not a gap that better engineering will close. They require understanding what the text says, and rules cannot do that.
Non-generative is not the same as deterministic
This is the distinction the category name hides, and it changes what you get. Marlowe generates no prose at all and is a reasonable choice for an author who objects to machine-written text, but its reports on plot, pacing and character are produced by a model making an assessment. Run it twice on a changed draft and you're reading a new judgement, not a recomputed count.
The tools in the table above give the same answer for the same file every time, which is what makes a finding closeable: you fix the thing, run it again, and the row is gone or it isn't. If your objection is to generated prose, the non-generative set is larger than this shortlist. If your objection is to guessing, this is the shortlist. Writing tools that use no AI at all sets out the two questions to ask a vendor.
Assembling a workflow
- A grammar checker, ideally one that runs locally, on a finished file rather than while you draft.
- A consistency checker, if you work to a style sheet and can tolerate the flag volume.
- Term extraction, for the invented vocabulary, which every checker above treats as errors rather than as the subject.
- Frequency analysis, for your own tics.
- Your word processor's own tools, which do more than most people use.
- A person, for everything above.
Run them in that order and each pass is cheaper than the one before it would have been. Fixing a name spelled two ways after the copyedit means re-reading the copyedit. The drafting half of a no-model toolkit is a separate question, covered in best non-AI writing tools.
Item six isn't a joke entry. The gap in the table is filled by a proofreader, a continuity reader or a beta reader, and a deterministic workflow that pretends otherwise is misleading.
Where this product sits
The Novel Kit is the names and invented terms row: rule-based extraction of proper nouns and unusual terms with counts and first appearances, no model anywhere, the same report every time, and no manuscript stored, since the file is transiently processed for the scan and never kept.
It is the row the others structurally cannot cover, because a spelling checker's only move against a word absent from its dictionary is to call it wrong, and a custom dictionary makes every spelling of an invented name correct, including the wrong ones. It also does not check grammar or ordinary spelling, so it sits alongside row one rather than replacing it.
Its limits are stated rather than implied, in what a rule-based checker cannot do.
A practical next step
Tick the rows in the table you have covered. The uncovered ones are your shortlist, and the bottom two tell you where to spend money on a person.
Sources checked
- Writing tools that use no AI at all
- Rule-based checking versus AI checking
- Best tools for checking a finished manuscript
Frequently asked questions
Is there a complete non-AI toolkit?
For mechanical checking, close to it. For anything requiring judgement about meaning, no, and that is a property of the problem rather than of the tools.
What is the strongest category?
Consistency and terminology, where deterministic methods are exact and generative ones are unreliable.
What is missing?
Anything that reads for sense: continuity of description, timeline coherence, character knowledge. Those need a person.