Writing analysis
See what needs attention before you change it.
Syft checks a draft for concrete writing problems — grammar, spelling, clarity, conciseness, style, consistency and academic convention — and shows each one on the sentence it came from, with the reason it was raised.
Bring a draft — paste it in or open a file — and Writing opens on it.
A report, not a rewrite · No score, no grade
The attendance pilot ran at three schools for a single term, and the steering group has asked for a reading of it before the summer timetable is set. Due to the fact that the scheme ran for only one term, the data is thin, but the direction is consistent across all three sites.
Northfield recorded a increase of 14 per cent in morning attendance. Beck Lane recorded 9%, and the pattern there was noticable within a fortnight; staff at the third site reported the same thing a little later, which suggests the effect is real rather than a artefact of one school's record-keeping, although it is important to note that we cannot rule out a seasonal pattern given that the pilot ran over a single term and the comparison period was drawn from a year with a different timetable.
What it checks
Seven areas, inside one lens.
They are categories of a single inspection, not seven separate tools. One run covers all of them and the panel reports every category, including the ones that found nothing — a category that disappears when it is clean is indistinguishable from one that did not run.
- Grammar
- Agreement, tense, sentence structure, articles and punctuation.
- Spelling
- Misspellings, typos, doubled words and confused word forms.
- Clarity
- Sentences that are hard to hold in one pass.
- Conciseness
- Words that can go without losing meaning.
- Style
- Voice, variety and emphasis.
- Consistency
- The same choice, made the same way throughout.
- Academic writing
- Objective academic conventions — not your personal voice.
Two honest edges. Syft does not ship a full dictionary, so a clean spelling result reads “no known spelling issues” — what the lens checked for and did not find, rather than a verdict on every word in the draft. And passive voice is not checked at all: a passive is not an error, and the rule that looked for it was switched off during calibration for flagging correct prose.
What a finding is
The span, the reason, and how sure the rule is.
Every finding names the characters it addresses, says what it thinks is wrong in one line, and carries the rule's own explanation of why it fired here. Nothing is summarised into a verdict about the draft as a whole.
- Stock fillerconcisenessSuggestionApply
Due to the fact that
“Due to the fact that” is 5 words for what “because” says in 1.
- Article choicegrammarWarningApply
a
“increase” begins with a vowel sound, so it takes “an”.
- Percentage style variesconsistencySuggestion
9%
This document writes percentages as “per cent” 1 times; this one uses the other form.
- MisspellingspellingWarningApply
noticable
“noticable” is not a word. The spelling is “noticeable”.
- Over-long sentenceclaritySuggestionApply
; s
75 words. It can be split where the meaning already breaks, at “; s”.
- Article choicegrammarWarningApply
a
“artefact” begins with a vowel sound, so it takes “an”.
- Stock fillerconcisenessSuggestionApply
it is important to note that
“it is important to note that” is 6 words for what “note that” says in 2.
- Signposted openerstyleSuggestionApply
However, t
3 sentences in this document open with a connective. “However” is one of them, and this sentence stands without it.
- Signposted openerstyleSuggestionApply
Moreover, t
3 sentences in this document open with a connective. “Moreover” is one of them, and this sentence stands without it.
- Signposted openerstyleSuggestionApply
Furthermore, t
3 sentences in this document open with a connective. “Furthermore” is one of them, and this sentence stands without it.
- Strong assertionacademicInfo
It is obvious that
“It is obvious that” asserts certainty, and this sentence names no source. Check whether the claim needs evidence here.
The Writing lens’s own output for the draft above, computed on this page.
Three severities
- Warning
- Near-certainly wrong, and the lens can say what it should be instead.
- Suggestion
- A real improvement, but a judgement that belongs to the writer.
- Info
- An observation worth seeing. Often entirely correct as written.
Severity is about how certain the lens is and what the consequence would be, never about urgency. There is no “critical”: an ordinary grammar slip is not an incident.
The marks on the draft carry only that — one weight for a warning, one for everything quieter. Not seven colours for seven categories; the category has a word in the panel instead.
The draft above produced 11 findings in 192 words, across all seven areas. That number is a count of what was found, not a mark out of anything: Syft does not score writing, and a draft with more findings is not a worse piece of work than a draft with fewer.
What you can do with one
Three actions, and none of them is a conversation.
A finding is a row you can act on, dismiss, or ask about. There is no chat, no assistant and nothing to prompt: the lens already said everything it knows when it raised the finding.
9 of the 11 findings above carry a correction. The rest are diagnostic: the lens can tell you what it sees without being able to fix it safely, and it says so rather than offering a guess.
- Apply
- Replaces exactly the characters the finding addresses, where the rule carries a correction that was validated separately. Nothing around the span moves, and the change lands in the document's history like any other edit.
- Dismiss
- Removes the finding from the list and leaves the text alone. The decision is remembered for that document, so the same rule does not raise it again on the next run.
- Explain
- Shows why this rule fired on these characters: the pattern it matched and the reason it is a problem. It is the rule's own account, not a generated one.
Diagnosis and action
Writing finds it. Improve changes it.
The two are separate on purpose. One is a report and the other is an edit, and keeping them apart is why Syft will tell you about a problem it cannot safely fix instead of quietly rewriting a sentence to make the report look clean.
Writing diagnoses
It inspects the draft and lists what it found, with the span, the severity and the reason. It changes nothing on its own, and a finding you never act on costs the document nothing.
Improve acts
For a finding with a safe correction behind it, Improve proposes the replacement for that span and nothing else — a word, a phrase, a sentence joint, never the paragraph around it.
Calibration
Built to be worth reading, not to find the most.
A writing checker earns its place by being right often enough that you stop second-guessing it. Syft's was calibrated against that standard rather than against how much it could flag.
Measured against prose annotated by hand. Every rule was evaluated on a frozen corpus of 120 passages whose ground truth was written by a person, never by running the analyser and keeping its own answers.
Clean prose produced nothing. 52 of those passages are controls: correct, publishable writing included to see what the lens says when there is nothing to say. It returned no warnings, no suggestions and no observations on any of them.
Noisy rules were switched off, not hidden. Of 83 rules in the inventory, 62 are active and 21 are disabled. A rule that flagged correct writing too often was turned off rather than demoted to a quieter severity where it would still interrupt you.
Corrections were validated separately. Offering to change your text is a stronger claim than pointing at it, so every automatic correction was held to its own evidence bar. 35 rules kept theirs. The rules that could not show enough evidence for a correction kept the diagnosis and lost the button.
Syft does not catch every writing issue, and it is not trying to. It is built so that what it does raise is usually worth your attention.
Scope
What a deterministic check does not attempt.
Writing matches patterns it can define exactly. Some real errors need a parser or an understanding of meaning, and those are recorded as out of scope rather than chased with a brittle rule that would cost more in false alarms than the catch is worth.
- Agreement across a phrase
- “The results of the survey was surprising.” The subject sits behind a prepositional phrase, and the rule refuses to reach across one. That refusal is what removed its false positives on correct prose.
- there / their by context
- Only a short closed list of following words is safe without a parser. Before an arbitrary noun, the lens says nothing.
- effect as a verb
- “The error effected three invoices” is missed outside the one frame the rule can recognise.
- Meaning
- Whether a claim is true, whether a source supports it, whether a pronoun refers to what you meant, whether a tense holds across a passage. None of that is a pattern, and the lens does not guess at it.
The same restraint shapes the academic checks. When Syft sees a claim asserted at full strength with no source beside it, the finding is called a strong assertion — not an unsupported one. A pattern engine can see how firmly something is written; it cannot know whether the evidence exists.
Part of Syft
Four capabilities, one workspace.
They read the same document and the same analysis, so a signal you find in one is the passage you revise in another.
See the whole workspace — documents, metadata, provenance and code as well as writing.
Check a draft you are working on.
Paste it in or open a file, and see what the lens finds before you change a word.