2.1.0

Adds one Scope rule and one example file. The other rules are unchanged.

Voice. When drafting in a person's own voice from their writing
sample or a voice profile, that person's habits now outrank the style
rules, except the rule against fabricating. Before this, the skill
removed dashes, negated contrasts, one-line endings, and stock idioms
even when they were the author's own. voices/ashwin-sathian.md is an
example profile for the skill's author: sentence-length and pronoun
counts from about 55,000 words of his writing, and descriptions of his
habits. It contains no quoted text and is meant only for drafts under
his name.

Test. Fresh model instances wrote a tutorial section and a film
note under 2.0.0, under 2.0.0 with the profile, and under 2.0.0 with an
added instruction to write "80% of the way to ASD-STE100". One model
judge, shown the pieces unlabelled beside real passages by the author,
ranked the profile drafts closest to him on both tasks and the STE
drafts farthest. The profile drafts moved toward his use of "we" and
of questions, and after one revision of the profile matched his rate of
contractions. They did not match his sentence length (15 words on
average against his 17 to 18, with far fewer long sentences) or his
dashes (none against five to six per 1,000 words). An early profile
draft also invented a first-person experience. The profile now warns
against that by name.

That is two tasks, one judge, and one writer model, and the second
profile round was checked by counts and not re-judged. The pieces and
the judge's packet are not published because the packet contains the
author's unpublished writing.

Not adopted. A Simplified Technical English filter was considered
after a suggestion by Andrej Karpathy in October 2026 and left out. In
the test above it cut average sentence length to 11 or 12 words with
little variation, and Wikipedia's "Signs of AI writing" page now lists
shorter sentences of 8 to 20 words as a habit of mid-2026 models.

2.0.0

The rules were rewritten against 2026 sources, and the rewrite was
tested blind before release. Existing installs will see different output:
that is the reason for the major version.

Why. The tells this skill targeted have moved. A July 2026 study by
The Economist of four major models found the signal now sits in long
words, long "and"-chained sentences, and thin punctuation, with the em
dash narrowed to Claude alone. Wikipedia's AI-vocabulary list for mid-2025
onward is four words, one of which was on its 2023 list. Anthropic's own
prompting guide names "mannered prose" (metaphor in place of direct
statement) as a habit of its current model. Version 1.x covered none of
this, and an audit of its own examples found that its rewrites had
swapped 2023 tells for 2026 ones: negated contrasts, one-line closers,
fragments for effect. They had also added facts the originals did not
contain. Details and sources: reference/research/2026-update.md.

Rules that changed.

• "Vary sentence rhythm on purpose" is replaced by "let sentence length
  follow the content". The old rule produced short sentences written for
  effect.
• "Cut hedge-intensifiers (rather, very, little, pretty)" is replaced by
  "hedge each uncertain claim once". The old rule came from a 1959 style
  guide and no source reported those words as a model habit.
• The dash rule no longer carries a number ("one or two per paragraph"),
  which had no source.
• New rules: say each thing once; replace a figure of speech with the
  fact it stands for; prefer the verb to its noun form and the common
  word to the long one; stop when the content stops (no closing line
  that repeats the paragraph, no announcing a point, no denying a view
  the reader does not hold).
• "Never invent" is now its own rule and covers plausible-sounding
  incidents, mechanisms the writer does not know, and anything a rewrite
  adds to its source. Blind testing caught both 1.1.1 and an early 2.0.0
  draft inventing a past incident for an internal blog post.
• The tells table is cut from nine rows of mostly 2023 vocabulary to six
  patterns that no rule already names.
• Rules and Scope are written as plain sentences. They were bold-lead
  bullets, a format the repo's own catalog lists as a tell.

Scope. All 1.1.0 carve-outs remain. Added: summaries (an abstract or
TL;DR exists to restate), speeches, poetry, and brand voice. The
non-English entry no longer says the tells do not transfer; a 2026 study
found model-preferred vocabulary recurring across 34 languages.

Trigger. The description now fires on prose of a paragraph or more
and on requests to humanize text, and says what the skill is not for
(proofreading a person's own writing, evading AI detectors). In 1.x it
fired on any written text, including one-line commit subjects where
Scope then told it to do almost nothing.

Detectors. SKILL.md and the README now say plainly that the skill
does not change AI-detector scores, with the research behind that in
reference/research/2026-update.md §3. The ai-detection keyword is
removed from the plugin manifest.

Validation. Replaced the 1.x method (a judge who knew which passage
was which, scoring against the skill's own list) with shuffled blind
pairs judged by two models, plus scripts/measure.py for descriptive
metrics. In the released round both judges preferred 2.0.0 to no skill
in 5 of 6 pairs, and to 1.1.1 in 5 of 6 and 4 of 6. An earlier 2.0.0
draft did no better than 1.1.1 and was revised. Losses, limits, and all
the raw pairs are in reference/validation-note.md and
reference/validation-2.0.0/.

Also. Four worked examples (two rewritten, two new in 2026 styles),
each noting what the 1.x rewrite got wrong. reference/claude-tics.md
records habits reported for current Claude versions with an evidence
grade for each. reference/oss-skills-review.md has an October 2026
re-survey of the most-used alternatives and what was taken from them.
SKILL.md is about a quarter longer than in 1.1.1.

v1.1.1

Audited against Anthropic's actual plugin submission requirements —
nothing needed fixing (manifest schema, validation, directory structure,
and security surface all already clean).

The community-marketplace submission form requires an Anthropic Console
account gated behind purchasing usage credits. Rather than pay for that,
this release adds .claude-plugin/marketplace.json, making the repo its
own self-hosted marketplace with no Anthropic review, account, or cost:

/plugin marketplace add AshwinSathian/humanize-writing-skill
/plugin install humanizing-writing@humanize-writing-skill

Verified end-to-end locally (add, install, uninstall) before shipping.

Full detail in CHANGELOG.md.

v1.1.0

First tagged release. No prior version was ever released to external
users, so this covers everything: initial 1.0.0 build plus the two
adversarial review rounds that followed it.

1.0.0 (initial): built from three research passes (academic
detection literature, editorial/practitioner style guides, a
cross-referenced catalog of 27 AI-writing tells) and a teardown of 13
existing public "humanizer" skills. Weights structural guidance above
a banned-word list.

1.1.0 additions:

• Genre/voice scope-gating in SKILL.md — the structural rules were
  correct for discursive essay prose but stated as universal; red-team
  testing found they regressed API docs, legal text, marketing copy,
  fiction, non-English text, and edits to someone else's prose. Fixed
  with an explicit ## Scope section, verified case by case with fresh
  subagents rather than assumed to work.
• Explicit user-instruction-priority rule: an explicit user request now
  overrides the skill's defaults, with the skill's own text saying so.
• A July 2026 cross-model study (verified, not assumed) narrows the
  em-dash tell to Claude specifically — the model this skill runs on.
• Two citation figures softened to "unconfirmed" rather than stated as
  fact, after a full citation audit found everything else checked out.
• Tables of contents on every reference file over 100 lines.
• README: install verification, uninstall, and feedback/contributing
  sections; CHANGELOG.md added.

Full detail in CHANGELOG.md.