Our approach
Why we built an engine instead.
A lot of writing tools use AI now. Here is why we don't, and why we think it matters.
AI feedback teaches you to write like AI.
When you ask an AI to critique your prose, it compares your writing to the statistical patterns in its training data. It tells you what sounds like good writing because it has seen a lot of writing and learned what the average of good looks like.
That is a useful thing. It is not the same as developing a voice.
Voice comes from the choices you make that deviate from the average. The sentences that run longer than they should. The dialogue that doesn't resolve cleanly. The image that doesn't quite fit but is exactly right. An AI will suggest you fix these. A deterministic engine just measures them and tells you what it found.
Inkbreaker does not tell you your writing is good or bad. It tells you your average sentence length is 9.2 words, your genre target is 15, and your sentence length variation is higher than 80% of your recent submissions in this type. What you do with that is yours.
That distinction matters. The feedback that tells you what to think about your work teaches you to think what the feedback thinks. The measurement that shows you what is actually on the page leaves the thinking to you.
Measurement you can verify is different from judgment you have to accept.
When Inkbreaker tells you your passive voice is at 11%, you can check it. Count the sentences with passive constructions. Divide by the total. The number will match.
When Inkbreaker tells you your reading ease is 68.0, that number came from a formula published by Rudolf Flesch in 1948:
206.835 − (1.015 × words per sentence) − (84.6 × syllables per word)
Plug in your own counts. You will get 68.0.
Paste the same passage into Inkbreaker twice and you will get the exact same numbers, down to the third decimal. This is the definition of a deterministic system. There is no randomness, no interpretation, no probability. The engine runs the same calculation every time.
This matters because trust in a tool is proportional to how much you can verify it. You can verify everything Inkbreaker shows you. The formulas are public. The benchmarks are documented. The metrics are counts and ratios, not impressions.
We use the Coleman-Liau index alongside Flesch-Kincaid specifically because the two formulas use different inputs. One counts syllables, the other counts letters. They cross-check each other. When they agree, the score is robust. When they diverge, the engine tells you why.
That is not how AI works. An AI that tells you your prose "lacks clarity" is making a judgment. Inkbreaker tells you your Flesch-Kincaid grade is 12.4 against a fiction benchmark of 8. You can disagree with the benchmark. You cannot disagree with the math.
The feedback that matters most comes from humans who read carefully.
There are things a deterministic engine cannot measure. Whether your voice lands. Whether the tension in a scene is earned or manufactured. Whether the dialogue sounds like a real person talking or a writer's idea of one.
These things require a reader. Not a language model trained to simulate a reader. An actual person who brought their own reading history to your page and had a response.
Inkbreaker's model is straightforward: the engine handles what is measurable, humans handle what isn't. Your metrics tell you what is on the page. A writer in your genre tells you how it reads.
AI tries to do both. It produces measurements that are actually opinions (your prose "feels dense") and opinions that are actually measurements ("your sentences average 22 words"). The categories blur. The feedback becomes harder to act on because you cannot tell what is verifiable and what is a guess.
We think the cleaner model is better. Know what the numbers mean. Know what only a reader can tell you. Ask for both from sources equipped to give them.
That is what Inkbreaker is built for.
Where we approximate, and what that costs
Four of our metric categories address craft questions that competitors typically solve with language models. We solve them with public lexicons and deterministic rules. Each approach has limits worth knowing before you read the score.
Sentiment
Inkbreaker reads the emotional register of your prose by looking up each content word in the Warriner et al. (2013) lexicon. The lexicon assigns 13,915 English words a 1-to-9 score for valence, arousal, and dominance, scored by thousands of human raters. We average across the words your passage matches. A competitor's sentiment model will tell you the same passage scores 0.72 positive without telling you why. We will tell you which words drove the average, and what the per-paragraph arc looks like. The metric cannot read negation or sarcasm. "She was not happy" looks up "happy" and reports a positive register. Read the score as a directional signal, not a verdict.
Coherence
Inkbreaker measures coherence by tracking word-level threading across sentences. Adjacent sentences share content words. Stems recur across paragraphs. The metric is built on Halliday and Hasan's foundational 1976 work on lexical cohesion, the same approach published academic NLP used for decades before embedding models existed. A model-based coherence score will tell you two sentences are semantically similar. We tell you which words bridge them, and where the bridge fails. The metric cannot judge whether your ideas logically follow. A passage can have high lexical cohesion and still be logically incoherent. Read the signal for the threading question, not the meaning question.
Grammar Patterns
Inkbreaker is not a grammar checker. We do not flag errors. We do not suggest corrections. Dedicated grammar checkers handle that well. We are not competing on it. What we measure is the pattern level: how often your sentences are simple, compound, complex, compound-complex. How often you reach for a subordinate clause. How often you turn a verb into a noun. These are stylistic fingerprints, not grammar mistakes. A varied complexity distribution is not "correct" the way subject-verb agreement is. It is a craft signal you can choose to act on. The framing matters. We tell you the pattern. You decide what it means for your draft.
Originality
Inkbreaker does not detect AI-generated text. We are not in that business. Originality detectors built on language models false-flag human writing constantly and miss tuned AI output entirely. What we measure is structural variety: how often your sentences open with the same grammatical template, how varied your sentence complexity is across the passage, how the rhythm of paragraph-level sentence length swings. We call the composite Syntactic Variety because that is what it measures. Some of the best writing in history is highly structurally predictable. The number is a mirror for you to look in, not a verdict for the engine to deliver. Use it that way.
What the engine cannot measure
The engine is a measurement instrument. There are several things it does not measure, and never will, because they require something the engine does not have: a reader who brought their own life to your page.
Semantic coherence
Whether your ideas logically follow from each other is not a property of the surface text. A passage can be grammatically tidy and rhythmically varied while saying nothing. The engine can sometimes flag the shape of incoherent prose (vivid disconnected images, dressed-up abstraction). It cannot tell you whether the sentences are actually arguing for something, or whether the second paragraph follows from the first.
Emotional resonance
Whether your writing moves a reader is the question every writer is trying to answer. No deterministic count can answer it. A reader who has lost someone reads a grief scene differently than a reader who has not. The engine has no access to that.
Originality of thought
Whether the ideas themselves are interesting is a judgment the engine cannot make. It can tell you whether your sentence openings are templated, whether your vocabulary varies. It cannot tell you whether the idea itself is interesting.
Truthfulness
For nonfiction, whether your claims are accurate is the responsibility of you and your editor. The engine measures how the prose works as prose. It does not fact-check.
Intent
Whether a stylistic choice was deliberate or accidental looks identical from the outside. A run of short fragments may be a craft decision or a draft that ran out of energy. The engine will report the pattern. It cannot read your mind about the reason.
This is why peer feedback exists in the product. The engine measures the craft of prose. Human readers measure everything else. Both are necessary; neither is sufficient on its own.
Inkbreaker's prose engine is built on Flesch-Kincaid (Kincaid et al., 1975), Gunning Fog (Gunning, 1952), Coleman-Liau (Coleman and Liau, 1975), and silent reading research from Brysbaert (2019). Every metric is a count, a ratio, or a formula applied to those counts. Nothing is inferred. Nothing is generated.
The only place AI touches Inkbreaker is content moderation, where we use it to flag submissions that violate our community guidelines. Every exercise, every piece of feedback, every published story is written by a human.
If you want to see the engine in action, the Prose Grade tool is free. No account needed, nothing saved.