Poisoneddiety Academic Prose Audit
This skill fixes two distinct defects that get confused with each other.
Compression is the serious one. The prose is accurate but too much is left for the reader to reconstruct: the actor is buried in a nominalisation, the middle term of an argument is missing, a citation stands where a sentence of reasoning should be. The reader can follow it only by doing work the writer should have done.
Mechanical register is the cosmetic one. Em dashes, uniform sentence length, connective filler, and the vocabulary that marks text as machine-produced.
Compression is the priority. A paragraph can be free of every AI tell and still be unreadable because its reasoning is elided.
What this skill must not do
The generic "humaniser" tools in circulation are built for blog and marketing copy, and applying them to a law paper degrades it. Protect the following, which those tools attack:
Hedging that reflects genuine doctrinal uncertainty stays. "The position is unsettled" is accurate, not weak. Where the law genuinely admits of two readings, saying so is the honest move. Cut hedging only where it is decorative — stacked modals cancelling each other, or qualification attached to a proposition that is not in fact doubtful.
Formal register stays. No contractions. No deliberate sentence fragments in body prose. No first-person plural exhortation. No manufactured "natural disfluency". The target is prose that reads as though a careful academic wrote it, not prose that reads as though it were spoken.
Passive voice stays where the actor is genuinely irrelevant or unknown. "The petition was dismissed" is correct when the identity of the bench is not the point.
Legitimate signalling stays. "It should be noted that X, because Y" is doing work when Y is supplied. Cut it only when it precedes a claim that needs no signal.
Terms of art, Latin, and citation conventions stay untouched. Do not simplify ratio decidendi into "the reasoning", and do not restructure a footnote.
Length is not the enemy. A legal proposition often requires a long sentence to carry its qualifications accurately. The goal is variance across sentences, not shortness in each.
Workflow
Pass one — mechanical
Run the script. It measures what should never depend on a model noticing it.
python scripts/audit.py <draft-file>
python scripts/audit.py <draft-file> --json # for structured output
It reports em dash rate against a budget of one per thousand words, sentence-length mean and standard deviation, the longest run of similarly-sized sentences, paragraph uniformity, nominalisation density, banned register hits, filler openers, hedge stacking, AI signature constructions, and sentences over fifty-five words.
Report the numbers to the user before revising. They should see what was wrong, not just the fixed version.
If the draft is pasted into the conversation rather than uploaded, write it to a temporary file first and run the script on that.
Pass two — judgement
The script cannot see the defects that matter most. Read the draft and mark each of the
following. references/patterns.md gives worked before-and-after examples for every one.
Buried actor. A nominalisation that hides who does what. "The determination of agency" conceals that a court determines it. Find the verb inside the noun and restore the subject. In legal writing, who acts is usually the point of the sentence.
Elided middle term. The argument runs A therefore C, with B unstated. Test it by writing the syllogism out. If the minor premise has to be supplied from the reader's own knowledge, insert it. This is the single most common cause of prose that "requires interpretation".
Stacked proposition. One sentence carrying two or more independently arguable claims. Split it, and make the relation between the resulting sentences explicit.
Undefined term of art on first use. Flag every first occurrence. A paper that assumes its vocabulary cannot be marked fairly.
Citation doing an argument's work. A parenthetical case reference standing where a sentence should explain what the case held and why it bears. Replace with the reasoning; keep the cite.
Conclusion smuggled into a premise. The thing to be proved appearing as a modifier before it has been shown — "the doctrinally unsustainable public interest ground" in a paper whose thesis is that the ground is unsustainable. Strip the modifier; earn it later.
Floating demonstrative. "This suggests", "that indicates", where the referent is a whole preceding argument rather than a nameable thing. Name it.
Pass three — rhythm
Only after one and two. Fixing rhythm first produces well-cadenced prose that still hides its reasoning.
Target a sentence-length standard deviation above seven, with no more than three consecutive sentences within five words of each other. Vary paragraph length too: a run of paragraphs all four to six sentences long reads as machine-assembled even when each is well built.
Achieve variance by consolidating and splitting for argumentative reasons — a short sentence after a long one to land a conclusion, a long one to carry a qualified proposition — not by chopping arbitrarily. Rhythm should track the argument's structure.
Em dashes: budget of one per thousand words, reserved for genuine parenthetical rupture. Nearly every instance can become a comma, a semicolon, a colon, or a full stop. Note that replacing em dashes with en dashes or double hyphens defeats the point; the script checks for both.
Output
Return four parts, in this order:
## Mechanical pass
[Script output, with the flagged items named.]
## Judgement pass
[Each defect found, quoted, with the category named. Quote the actual text —
the user needs to recognise the pattern in their own writing, not just receive
a corrected draft.]
## Revised text
[Full revision.]
## What changed and what was deliberately left
[Substantive edits, and — importantly — anything flagged but preserved,
with the reason. Hedging kept because the doctrine is genuinely unsettled;
passive kept because the actor is irrelevant; long sentence kept because it
carries a qualification that cannot be split.]
The fourth section matters. A revision that silently strips accurate hedging has made the paper wrong in order to make it read better.
Verification
Re-run the script on the revised text and report the before-and-after numbers. If any metric moved the wrong way, say so rather than presenting the revision as a clean win.
Never alter the substance of a legal claim while revising prose. If a sentence cannot be clarified without deciding a question the draft leaves open, flag it for the user instead of resolving it. Clarifying prose is not the same as taking a position, and the two must not be conflated in someone else's paper.