Humanist

Humanist: edit the draft, then read it

The goal is a better piece of writing, not a disguised one. In a blind study where editors judged quality alone, the processed draft beat the raw draft 22 times out of 22. In a separate study, the same processing moved detection of machine authorship by no measurable amount. Both of those are true at once, and the first is the one worth working for.

So the work splits into a pass that rewrites, a checker that reports, and a read that catches what no regular expression will. Run them in that order, and read the note on what each stage is actually known to do.

Before anything else, name the surface. Public long-form, internal document, short-form social, quotable snippet. The --lenient flag, the readability targets and the calibration bands all differ by register, and a draft judged against the wrong one produces confident nonsense in both directions.

What this does not do

Three non-goals. The first two are impossible here; the third was attempted, measured, and abandoned.

It does not detect AI text. Given a passage, this skill cannot tell you whether a machine wrote it. Nothing in the repository does that, the accuracy of tools that claim to is poor, and a false accusation is worse than an unanswered question.

It does not defeat AI detection or watermarking. Statistical watermarks live in word choice across a whole passage, not in characters or phrases, and no rewrite can promise anything about them.

It does not make prose pass as human, and that was the original goal. Across roughly 700 blind judgments: raw output was identified as machine-written 100% of the time and fully processed output 97.8%, a difference smaller than the design could resolve. Explicit instructions for "writing more human" fooled 0 of 17 judges. Length changed nothing. Genuine human prose was never once misidentified.

The reason is worth carrying into the work: the pipeline removes the tells it encodes and worsens the ones it does not. AI vocabulary fell from 70 judge citations to zero while uniform beat rate rose from 60 to 101. A subtractive rewrite makes prose more uniformly well-shaped, and uniform good shape is itself a tell. If you want prose that reads as a person's, that comes from the person: the specifics only they hold, unevenly distributed, with the rough parts left in. A pipeline cannot supply it, and step 3 below is where it has to enter.

The pipeline

Step 1: the rewrite pass

Read third_party/humanizer/SKILL.md in full and apply it. It is Siqi Chen's MIT-licensed humanizer, vendored here, and it is the subtractive half: it strips inflated significance, AI vocabulary, negative parallelisms, filler, fake-candid openers and the rest of its 33 patterns, rewriting rather than deleting.

Skipping this wastes everything after it, because the checker would drown in findings that one rewrite clears wholesale.

Two cautions the pipeline's own audit surfaced:

• Do not swap an em dash for a spaced hyphen and consider it handled. That produces a construction the composition read then has to catch. An em dash usually wants the sentence rebuilt around it.
• The rewrite pass is itself machine output. Everything it produces goes through step 2. That is the entire reason step 2 exists.

Step 2: the checker

python humanist.py draft.md

It sweeps the draft against the rules derived from ai-tropes.md, unwraps soft line breaks first (a hard-wrapped draft otherwise defeats every multi-word pattern), and reports counts. It never edits.

Read the frame it prints before the findings. It states the word count, which config it loaded, how many of its rules are active, how many were overridden, and the readability grade. A run with rules silenced cannot masquerade as a clean sweep, which is the point.

Severity. FAIL does not ship in public copy and sets exit 1. WARN is judged in context and counted, so repetition stays visible. --lenient downgrades the format and phrasing registers to notes for internal documents, which legitimately use bold bullets and plainer hedging on purpose.

Read WARN counts as density, not as a list of errors. For the largest family of tells, the not-X-but-Y antithesis above all, any single instance is usually fine and the pattern is what gives a draft away. One is nothing. Six in eight hundred words is the tell.

Exit codes. 0 clean, 1 FAILs remain, 2 the tool could not do its job. Never treat 2 as a verdict; it means the run did not happen.

What this stage is known to do, honestly. In a blind quality study the rewrite pass beat the raw draft 16 of 16 times. Adding this checker on top of it produced no measurable further gain: it changed nothing at all in 11 of 18 cells, and where it did act the preference could not be separated from chance. Treat its output as material for the composition read, which is what it is good for, rather than as a stage that improves the draft on its own. Its clean report is not evidence the prose is finished.

Step 3: the composition read

Five tells are manual forever, because no regular expression catches them without lying about its precision. After the checker is clean, read the piece aloud:

1. Dead metaphor. One metaphor beaten five to ten times. Introduce it, use it, move on.
2. Fractal summaries. Summaries of summaries at every level. One close per piece.
3. Analogy stacking. Rapid-fire borrowed authority. Your own receipts outrank anyone else's.
4. One-point dilution. One argument restated ten ways to feel thorough. State the conclusion once and end there. A case that takes 800 words gets 800 words.
5. Near-verbatim repetition. Exact repeats are caught mechanically; the near-verbatim kind needs the read.

Then check the five laws below at paragraph scale. Drafts scoring zero FAILs have still contained tells found only on a read-aloud. The read makes the final call. The checker exists to make the read cheaper, never to replace it.

This step carries more weight than the two before it. The measured tells that survive a clean checker run are all rhythmic and structural rather than lexical, and ai-tropes.md has them under "The rhythm tells" with what to do about each. The short version: vary paragraph length so no two land the same number of beats, end half your sections flat rather than on an epigram, hedge only where the writer would genuinely be unsure, leave one objection standing, and put in the specifics only the author has. None of that is checkable by machine, all of it is what a reader notices, and the author has to supply it.

The five laws

The deepest family of tells is a sentence about the material instead of the material. Each instance is well-formed English, which is why surface rules miss it.

1. No sentence may justify a choice the page has already made.
2. No sentence may announce the writer's own honesty. Admitting a concrete fault is fine; it costs the writer something. Cut any line that compliments its own candor.
3. No sentence may tell the reader how to weigh material it has not shown yet.
4. No phrase may assert that the material deserves the reader's attention.
5. The page does not manage the reader's verdict. It supplies material, not conclusions about the material. This one subsumes the other four.

Quoted sources

python humanist.py draft.md --strip-quotes

The checker cannot tell your prose from a source you quote, and a quoted source will happily trip style rules its author never agreed to. --strip-quotes removes blockquote lines before scoring and prints how many. Run it on anything carrying quoted material, and remember why it exists: a quoted source is evidence, and editing evidence to satisfy a style rule is falsifying it.

Web copy

Never run the checker on raw HTML or on text scraped from a rendered page. The rendered DOM carries theme chrome (nav, footer, promo widgets), and every one of those strings is a false positive. Export the body HTML from the CMS instead:

python tools/html2prose.py body.html > body.txt
python humanist.py body.txt --strip-quotes

html2prose.py maps headings to #, blockquotes to >, list items to -, excludes code blocks by design, and prints its element counts to stderr so a partial extraction cannot be reported as a full one.

Calibration

Out of the box the checker runs its universal rules against wide generic bands labeled UNCALIBRATED. Voice bands are meaningless until measured from the actual author.

python humanist.py --calibrate my-writing/ --register general
python humanist.py draft.md --mode post --register general

Point --calibrate at a directory of your own .md/.txt writing. Each run pools that directory into one corpus and produces one band set stored under --register, written to humanist.config.json in the current directory. Multiple registers are multiple named band sets: run it once per register, each on its own directory. A technical how-to and a first-person essay are different measurements and differ more than the markers do.

The sampling law, binding. A rare-event rate cannot be tested below the word count that yields an expected count of about three. Below that the checker prints UNTESTABLE instead of a number. A marker rate from a short piece is noise in both directions: do not report it, and do not let a clean short sample count as evidence the voice landed.

Calibration pitfall. Measure published writing, not chat logs. Collaborator and chat corpora over-predict first person and question rate, because collaborating inflates both. Calibrate on what you shipped to an audience, unassisted.

Configuration

House taste is config, never core. config.example.json documents each key with worked examples; copy it to humanist.config.json or pass --config.

• severity_overrides. Keys are rule IDs, matched exactly. Run python humanist.py --rules for the list. An unknown ID or an invalid severity is a hard error rather than a silent no-op, because the previous behavior turned a FAIL into an uncounted note and let the run report CLEAN.
• extra_banned_words / extra_warn_words. Per-house bans with your own rationale. A blanket ban on a common English word is house taste, not universal law, which is why the shipped set contains none.
• never_flag_words / exempt_patterns. How a real author's measured tics survive the sweep. Exempt patterns are anchored at the hit, so an empty drum-roll still fails while an anchored report survives.
• fk_quotable / fk_body. Readability targets, checked and reported as a WARN. An audience-friction measure, not a voice measure: a real voice at grade 9 beats slop at grade 6.

Two reconciliation tests generalize whenever a rule and a real sentence collide. Genuine questions survive; only the self-answered drama form is a tell, and the test is to flatten the question into a statement, and if nothing is lost it was staging. Literal industry senses survive; a platform ecosystem and a software framework are legitimate, and abstract use is the offense.

Why the defaults are calibrated the way they are

The rule set this skill inherited was swept over 24 known-human documents (107,191 words, published 1854 to 2018, across four registers) and blocked 22 of them. 94.6% of its 919 failures came from two rules, curly quotes and em dashes, which describe how a file was encoded rather than how it was written. Jack London's chapter scored 199 failures, all 199 of them quotes and dashes.

After recalibration the same corpus produces 8 failures and blocks 4 documents. Every rule that was demoted was demoted against evidence, not taste:

• AI vocabulary split into two tiers. Every one of its 15 hits on the human corpus was ordinary English: Obergefell's "fostered and then adopted a baby boy," the Rogers Commission's "crucial information about the O-ring damage," Thoreau's "elevate his life," RFC 793's "utilize the SYN control flag." A small anchored FAIL tier remains for constructions with no innocent use.
• Negative parallelism became a counted WARN. Its detection was widened, since it used to miss three of the five forms its own name claimed, but 10 of its 17 corpus hits were the correlative "not only X but Y" as written by Thoreau, Du Bois, Russell, and by Strunk in Elements of Style.

This matters for how you read the output. A checker that fails nine of ten published essays does not measure what it claims to, and its users learn to ignore the number. tests/fp_guard.py is a CI gate that holds this line: any rule change pushing the false-positive rate back up fails the build.

The honesty rule

The checker never edits. It only reports. The person decides. Every fix is made by the author or with the author's sign-off, because a checker that rewrites is a checker that can falsify without leaving a trace.

Standing corollaries:

• Never soften a claim to make it flow. Every number keeps its source and its base.
• Never add a story, an example, or a number the author did not supply. Leave [AUTHOR: ...] placeholders until they fill them.
• The medium decides length. When a register is uncalibrated, shorter is the safer error for anything public.
• A rule change is not an implementation fix. Widening a regex so it matches its own stated law is maintenance; changing the law is a decision. Both get swept over the false-positive corpus before they ship.
• What makes prose attributable is its specifics. Strip them and the line could have come from anyone. What cannot be faked is the evidence only you hold: the events you witnessed, the numbers whose denominators you can produce. This is the one input no stage of this pipeline can generate, and the measurements say it is the input that matters most.

Companion tool

app/inkwash.html at the repository root is a single offline page that strips invisible characters and machine typography out of finished prose, and reports the metadata hiding inside imported .docx and .html files: zero-width characters, smuggled Unicode tag payloads, homoglyphs, exotic spaces, curly quotes. It runs entirely in the browser and reports every change it makes. It works on characters, so it cannot touch a statistical watermark, and it says so on the page.