/declauding

declauding - Changelog

All notable changes to the declauding skill are documented in this file. The format is based on Keep a Changelog.

[0.9.3] - 2026-09-22

Fixed

• trigger at task start; fixes from a live subagent test (0.9.3)

[0.9.3] - 2026-09-22

Changed

• Description triggers at the start of the task ("Load at the start of any task whose deliverable is prose another person will read … before drafting it"), covers results saved to a file, and names postmortems. The 0.9.2 wording ("run as the last step … before the push, post or publish") loaded in 0 of 3 subagent tasks that ended in a README, a PR description and a postmortem; the new wording loaded in 2 of 3 on the same tasks, and one of the two ran the lint on its draft. Skills get chosen when the task is read, not after the draft exists.

  Confirmed outside any repository (headless claude -p from an empty directory, no CLAUDE.md, same three tasks with a neutral diff for the PR): the new wording loaded in 15 of 15 runs, always before drafting, and 14 of 15 linted their own draft; the 0.9.2 wording loaded in 1 of 9. A code-only control loaded it in 0 of 2. Inside claude-workspace the new wording loaded in 2 of 3, where the repo's own prose routing competes with it.

• Step 2b gives a fallback for contexts that cannot spawn a subagent: answer the review prompt yourself if you did not write the draft; if you did, answer it and say the review was not independent.

• models/opus-5.md editor checks: leave a plain factual title alone, and re-read rewritten sentences for entry-2 reversals. Both failures appeared in an Opus 5.5 test pass that had read the profile.

[0.9.2] - 2026-09-22

Fixed

• description as an invocation trigger (0.9.2)

[0.9.2] - 2026-09-22

Changed

• Frontmatter description rewritten as an invocation trigger. It opens with the agentic case (the last step before any prose Claude drafted reaches another reader, without being asked) and names the artifacts: PR description, commit message, README, doc, blog post, issue or review comment. The old text spent its first 600 characters cataloguing tic families, which describes the skill and does not tell a model mid-task when to load it. Adds a boundary against challenging. 1017 -> 786 characters.

[0.9.1] - 2026-09-22

Changed

• model profiles as instructions, not results (0.9.1)

[0.9.1] - 2026-09-22

Changed

• references/models/*.md rewritten as instructions. Each file now says where to look in that model's drafts and what to cut, with one specimen per rule, plus what to check in your own rewrite when you are that model. Scores, confidence intervals and entry counts are gone from the files; each keeps one evidence line pointing at oaustegard/experiments model-register-drift/. The numbers told the reader how the profile was earned and did not change what the pass does.
• SKILL.md Model profiles table and step 2 note state where each model's tics sit rather than how it ranked. Extending tells profile authors to write instructions and leave the scores in the experiment.

[0.9.0] - 2026-09-22

Added

• per-model profiles via progressive disclosure (0.9.0)

[0.9.0] - 2026-09-22

Added

• references/models/ — per-model profiles loaded on demand: opus-5.md (Opus 5, and Opus 5.5 as an unmeasured prior), opus-4.md, sonnet.md, haiku.md, fable.md. Each has an as author half (entry composition of that model's drafts, where the linter under- or over-reads it, which workflow steps carry the pass) and an as editor half (how that model fails running this skill). SKILL.md gains a Model profiles section with the index and the loading rule: author profile when the draft's model is known, own profile when the running model is known, one file for both when Claude cleans its own draft.
• Entry 3 lint nouns: fix, heuristic, lesson, insight, takeaway, culprit, cause, trap, trick, win. Opus 5 wrote "the real fix" and "the actual heuristic" and both passed 0.8.1. work and cost were tried and dropped: "the actual work" and "the real work" appear three times in CPython's stdlib, all literal. Specimens and negatives in tests/sample-tics.md.

Changed

• The five earned-exception tables moved from SKILL.md to references/exceptions.md; SKILL.md keeps a pointer, two worked cases and the modifiers-carry-content rule. Step 3 loads the file before any deletion.
• The four-family narrative in "The one pattern" is now a table.
• Header "The author's own writing outranks this skill" is now "Author's writing sample" (entry 7, found by the structural review of this release).
• SKILL.md 23,970 → about 21,400 bytes.

Provenance

Evidence is oaustegard/experiments model-register-drift/ (2026-08-23, and the Fable 5.1 addendum of 2026-09-01): ten samples from six models, blind pairwise staging judgments, and the hand-adjudicated specimens in adjudication.json. The hand count was retracted as a score there and is used here only for entry composition. Editor-half evidence is claude-skills PRs #769, #771, #779 (Opus-line passes) and #791 (Sonnet audit agents deleting evidence with framing). No Opus 5.5, Opus 4.x-as-editor, Haiku-as-editor or Fable-as-editor measurement exists; those sections say so.

[0.8.1] - 2026-09-09

Other

• prompt-audit: dated prompting patterns across the skill catalogue (#791)

[0.8.0] - 2026-08-30

Other

• declauding 0.8.0: content preservation and a fitted staging rank (#781)

[0.8.0] - 2026-08-30

Added

• scripts/declaude_diff.py — step 5 of the workflow as an exit code. Compares a draft against its rewrite and reports what the edit LOST and what it ADDED. Membership for numbers, names, quotations, code and link targets; counts for superlative, scope, negation and hedge constructions. Standard library only. --git PATH --ref REF compares a working tree against a ref.
• scripts/declaude_rank.py and assets/staging-axis.json — optional stage 1.5. A fitted direction in sentence-embedding space, the mean of embed(was) - embed(now) over the 41 matched pairs in register.md, that sorts a draft's sentences by how staged they look. Shortlists; decides nothing. --fit refits and reports leave-one-out. Needs torch and transformers, which no other stage does.
• references/preservation.md — the measurements behind both, negatives included.

Provenance

Written after a pass over a published post dropped a claim three times. The linter found 1 of the 8 tics in that draft; the editor caught the three regressions on self-review and nothing else would have.

Measured before shipping:

• Cosine similarity cannot do the preservation job. On the three real cases the lossy rewrite scores 0.9396 against the source and the faithful rewrite 0.9128 — inverted on the pair that matters, and a threshold passing the first would flag a register-only edit at 0.8914. Paraphrase invariance is what an encoder is trained for. Hence the lexical design.
• Counting superlative tokens misses the case the check exists for. "The format that most invites staged reveals" rewritten as "more than most formats do" keeps most. Counting constructions catches it; the first cut of the regex did not, and missed its own motivating case in testing.
• The rank axis is 76% leave-one-out on its own pairs, and on the pre-edit post put all nine subsequently-edited sentences at a median rank of 13 of 53 against a chance median of 26 (permutation, 200k draws, p = 0.031).
• It cannot rank documents: -0.26, p = 0.46 against ten human-graded samples. A cruder centroid axis scored -0.10, p = 0.79. The regex scan scores -0.37 on the same ten and the sign is negative, because its rate measures the flat lexical tells and the samples carrying those stage least. No document score is exposed.

False-positive budget

declaude_diff.py is silent on a file against itself, on a register-only edit that changes no claim, and on the shipped version of the post that motivated it. It reports both real regressions from that pass. Presence rather than frequency: a name used twelve times and now thirteen was not invented, and reporting it is how a guard becomes noise.

[0.7.0] - 2026-08-29

Added

• Entries 48 to 52, a fourth group: the confiding-essayist register. 48 performative candour ("let's be honest", sentence-initial "Honestly,", "you don't have to take my word for it"), 49 stranded auxiliary reversal ("The tool died; the data didn't."), 50 retroactive significance ("that's why X mattered"), 51 totalizing designation ("that's the whole point", "the only release notes I trust"), 52 obituary headline ("X is dead", "long live X"). Each has an earned column.
• Lint categories candour, reversal, retroactive, totalizing, obituary, and an obituary check on header text. "X is dead" is left alone inline — a dead process is a dead process — and flagged only in a header.
• _echo_runs, entry 26's third shape: two consecutive sentences sharing a five-word run that covers at least half the shorter one. Anaphora keys on the opening words and the triad regex on the commas; neither reaches A shopping cart is an object in the system. A chat room is an object in the system.
• _question_runs, entry 10 without the fragment answer: two or more questions in a row where at least one after the first is clipped.
• Entries 50 to 52 added to STRUCTURAL_ENTRIES in declaude_review.py. A retroactive grade is only visible against the passage it grades, a totalizing claim is earned when the set is named nearby, and an obituary headline is a header. 48 and 49 are lexical and stage 1 reaches them.
• Specimens for all five entries plus the two detectors in tests/sample-tics.md, with four negatives that must stay silent.

Changed

• Entry 3 gained the staged variants of its gesture: "here's the twist / catch / kicker / rub", "that's the part", "my favourite part of".
• Entry 10 gained the stacked-question run.
• Entry 18 gained bare sentence-initial "Turns out"; the rule only reached "it turns out".
• Entry 19 gained "batteries included", "zero config" and "small enough to fit in your head".
• Entry 26 is now three shapes rather than two.
• is the whole point moved off the aphorism rule; entry 51 owns it.

Provenance

Simon Willison's llm-cliche-highlighter added fifteen patterns and three structural detectors on 2026-08-27, in simonw/tools#322 and #323. Eight of the fifteen were already covered here in some form and extended existing entries. Five had no entry and became 48 to 52. Two of the three detectors are ported; the third, his colon-into-a-triple, is already the comma-list half of entry 26.

Several ported regexes are narrower than the source, which is tuned for essays rather than for technical prose. His x-is-dead fires on any "X is dead"; here it fires in headers only. His echo detector needs one shared four-gram; here it needs five words covering half the shorter sentence. His question run needs two questions; here one of them after the first must be clipped, because two full questions in sequence is how a person changes subject.

False-positive budget

tests/sample-clean.md still reports zero. It did not on the first cut: the question-run rule fired on So, what's next? Is this a project that starts and ends with DeepSeek v4 Flash?, which is antirez changing subject, and the clipped-question condition came out of that.

On 180,929 words of Python stdlib docstrings the five new categories fire 4 times (0.022 per 1,000 words), all four on entry 49 and all four the ordinary ellipsis the entry names as earned. _echo_runs fires 36 times (0.199 per 1,000), between the existing anaphora detector at 0.077 and fragment cadence at 0.298. The first cut of that detector, keyed on a shared five-gram alone, fired 91 times on phrases like "is the same as using" inside sentences that were otherwise unalike; requiring the run to cover half the shorter sentence is what took it to 36. _question_runs fires 0 times.

tests/sample-tics.md now reports 167 candidates across 42 categories, SKILL.md --skip-quoted 13 and README.md --skip-quoted 18. The extra hit on SKILL.md is the reuse detector counting a fifth parallel table label; one of the eighteen on the README is entry 49 on "Symmetry and antithesis are not", which is the earned ellipsis.

[0.6.0] - 2026-08-28

Other

• declauding 0.6.0: entries 43-47, the register the pass produces (#779)

[0.6.0] - 2026-08-28

Added

• Entries 43 to 47, a fourth group: the flat-certainty register this skill's own output lands in. 43 flat-certainty adverb ("provably safe", "quietly dropped"), 44 juridical register (a linter that rules, a default that is a carve-out), 45 verification-provenance compound ("byte-identical", "mutation-checked", the re- prefix claiming a second pass), 46 privative coinage ("unguarded", "unwired", "vacuous"), 47 exhaustive negation ("nothing in it does X", "nowhere else for it to be"). Each has an earned column; none is a blocklist.
• references/corpus.md — the measurement behind them, the method, and what the numbers do not support.
• Lint categories flat-certainty, juridical, verification, privative, exhaustive, and a corpus-register density line reporting the draft's rate of the cluster's top-150 vocabulary.
• Earned-exception rows in SKILL.md for entries 43 to 47, and specimens plus four must-stay-silent negatives in tests/sample-tics.md.
• Entry 47 added to STRUCTURAL_ENTRIES in declaude_review.py: an exhaustive negative is earned when its scope is named, and only the surrounding sentences say whether it is. 43 to 46 are lexical and stage 1 reaches them.

Fixed

• load_register in declaude_review.py ended an entry only at the next numbered heading, so selecting the file's last entry pulled the trailing "Quick self-check" section into the prompt with it. Any #/## heading now closes the block. Latent until 47 became the last entry.

Provenance

Louis Abraham's load-bearing clusters GitHub pull request descriptions by vocabulary — 461,121 descriptions, 85 whole weeks, 2025-01-06 to 2026-08-17, ten clusters, no time parameter in the fit. One cluster is 0.70% of the first eight weeks and 39.5% of the last four, rising 1.24 points a week. Its highest-lift words are load-bearing 39x, plainly 34x, quietly 30x, refusal 28x, re-derived 27x, asserted 25x, nobody 25x, genuinely 24x, outright 23x, byte-identical 23x.

That is not the slop vocabulary of entry 20. It is the register this skill aims at, and it is now the fastest-growing way of writing a pull request description there is. A pass that removes 42 staging tics and lands here has traded one detectable shape for another.

False-positive budget

tests/sample-clean.md still reports zero. On 115,920 words of Python stdlib docstrings the five new categories fire 18 times (0.16 per 1,000 words), ten of them on silently. is simply was cut from the adverb rule for scoring two of those and not appearing in the corpus list at all. unsatisfiable was cut from the privative rule as a SAT term.

The corpus-register line is a register locator and not an authorship detector. load-bearing's own human-written README scores 1.51 per 100 words of body prose, above this skill's SKILL.md at 1.47, on nothing x9, carries x3, alone x2, never x2 and half x2. Human stdlib docstrings run 0.08 median and 0.32 at worst over 46 chunks of 2,500 words. The note on the line says this, because a reader who takes it for a detector will cut nothing and measured out of correct sentences.

tests/sample-tics.md now reports 136 candidates across 35 categories, SKILL.md --skip-quoted 12 and README.md --skip-quoted 15 — the README's three new locator hits are the repository's name, which entry 19 lists as a metaphor and --skip-quoted cannot tell from a proper noun.

[0.5.1] - 2026-08-23

Other

• declauding 0.5.1: bring the README current, scope the header check (#772)

[0.5.1] - 2026-08-23

Changed

• README brought current. It described 37 entries and two register halves, which was already stale by one before this release — 0.4.0 added entry 38 without updating it. Now three groups, 42 entries, a table for 39 to 42, refreshed corpus counts, and the provenance of the new block.
• Nominalized-header check scoped to a prepositional tail ("Authorship in a sourced child") or a two-word modifier plus abstraction ("Spawn availability"). A bare one-word nominalization is a legitimate label for a term the document defines, and the first cut fired on this skill's own ## Overcorrection and ## Calibration.

Added

• Earned-exception rows in SKILL.md for entries 39 to 41, including the one-word-label carve-out the header check now honors.

[0.5.0] - 2026-08-23

Other

• declauding 0.5.0: entries 39-42 (#771)

[0.5.0] - 2026-08-23

Added

• Entry 39, welded epigram — a maxim joined to a fact with "and" or "so", inside one sentence. Entry 12 only catches the paragraph-final form.
• Entry 40, spec-ese — demonstrative subjects, reflexive emphatics, and latinate state verbs ("idles", "holds no", "awaiting input") for a program doing nothing.
• Entry 41, nominalized header — the standard overcorrection from entry 7. Fleeing a verdict header into an abstraction passes entry 7's regex and fails its table-of-contents test.
• Entry 42, contents-list standfirst — "X, plus the Y the docs leave out".
• spec-ese lint category, three new aphorism and two new announce patterns, and a nominalized/gerund header check.
• Two self-check questions covering the welded epigram and the read-aloud test.

Diagnosed on a create_session reference artifact whose four section headers were all rewritten from verdict-shaped into nominalized by a single declauding pass, after which the linter reported the document clean.

[0.4.0] - 2026-08-21

Other

• declauding 0.4.0: entry 38, announce-then-deliver (#769)

[0.4.0] - 2026-08-21

Added

• Entry 38, announce-then-deliver. A counted noun-phrase label standing in front of the content it names: "One factual note, not fixed:", "Two things I did not do.", "One judgment call left alone." The label carries nothing the content does not, and the fix is usually to delete it and start with the thing. Four announce rules in declaude_lint.py reach the regular forms; the bare "One note:" shape is left alone because a pattern for it over-fires on legitimate list intros.

  Found by an author flagging it in this skill's own output — an annotated register pass whose frame prose used the shape nine times, twice as a section header, in a document whose central finding was bad headers. None of the existing entries covered it: entry 5 is a placeholder in the subject slot, entry 17 narrates the writer's procedure rather than labelling the content, entry 29 is the bulleted cousin, and the staging rule at "announces a list before giving it" targets demonstratives (Here is what X:) rather than counted labels.

  Fired before trusting: 6 of 6 specimens caught, 6 of 6 negatives silent (including luria's "One decision, one thing." and the corrected form "One note I did not fix:"). Dropping the qualifier requirement from the first rule over-fires on the aphorism, which is what holds the precision. Zero new hits across all four existing test corpora.

Fixed

• Register entry 37, Dressed metaphor, was written # 37. where every other entry is ## N., so it rendered as an H1 and fell out of any ^## [0-9] count. Unrelated to the above; picked up while editing the file.

[0.3.0] - 2026-08-17

Other

• declauding v0.3.0 — two stages, folded from #762, #763 and the recall pass (#764)

[0.3.0] - 2026-08-17

Three branches folded into one version: a recall pass on the linter, the structural-pass work from #762, and the guard work from an unmerged 0.2.2. They were three answers to the same complaint and shipping them separately would have meant three 0.3.0s.

The shape that results is two stages. Stage 1 is declaude_lint.py, which now sees HTML and detects triads, reuse and repetition. Stage 2 is declaude_review.py, which sends only the slots regex cannot judge — headers, opening sentence, closers, isolated lines — to a model. Neither subsumes the other: stage 1 over-flags a comma that belongs to a citation, stage 2 reads it in context; stage 1 finds a verdict header every time, stage 2 finds the aphoristic closer no regex reaches.

Measured by scoring the linter against a hand annotation of the same 900-word draft: 31 candidates, 24 of them real, against 68 instances a full sentence pass found. Recall 35 percent, precision 77. After this release the same draft reports 60 candidates, 45 real, for 66 percent recall at 90 percent precision, with tests/sample-clean.md still at zero.

Every miss had the same cause. The rules were closed vocabulary lists, and real prose used a word that was not on the list: the participle rule listed highlighting|underscoring|emphasizing and walked past hitting, genuinely had a six-word whitelist and walked past genuinely considered, and the (part|thing) that required matters|counts|transfers and walked past the thing that made it slow. Widening those to shapes rather than words recovered 21 instances at zero cost to the clean corpus.

Added

• Forced triad detection, category triad. Entry 26 had no rule at all, and it was the largest single category in the evaluation draft at seven instances. Two detectors: a comma list with negative lookaheads for relative pronouns, and an anaphora form covering three consecutive sentences or three consecutive clauses opening on the same two words. The anaphora form catches It was a message. It was a permission slip. It was an out. and something to steer toward, something to be measured against, something to fear, neither of which the comma rule can see. The new earned-exception row for entry 26 is the guard on it, and entry 27 read in the right direction is the guard on the repetition statistics below.
• reuse block. Groups the hits already found and reports any construction used more than once, with its line numbers. No new rules, no new scanning. The evaluation draft used the part that three times and the linter previously reported three unrelated hits.
• Per-instance em-dash gotcha, category em-dash: a dash whose clause runs to the end of the sentence, which is the drum-roll shape rather than the aside. Density was already reported; the individual dashes were not. Entry 16 gains a demotion test: replace the dashes with commas, and see whether the aside survives. A beat does not, because the beat was the punctuation.
• Document-level repetition statistics: repeated sentence openings and repeated content trigrams. Uniformity is the tell, and it is measurable without a reference corpus, which is why this is a statistic rather than a model.
• Widened rules for participle tails at sentence end, genuinely as a self-grade, the part/thing that, Nobody claims, Not X. fragments, represents a kind of X, sentence-initial from-X-to-Y, Here is what X:, which was this:, that's not quite right, a doubled adverb, and the entire architecture of the thing.
• Specimens for all of the above in tests/sample-tics.md, which now reports 91 candidates across 28 categories, up from 62 across 24.

Stage 2, from #762:

• scripts/declaude_review.py. Extracts headers, opening and closing sentences, and isolated one-sentence paragraphs, then judges only those against the structural entries in references/register.md. Gemini or Anthropic key, --emit-prompt fallback, --slots for extraction only. Against the structure fixture across four runs it caught the subtitle, both coy headers and the aphoristic closer every time, with zero false positives on the eight clean slots — including EU AI Act, Article 50, where the comma belongs to a citation and stage 1 flags it.
• Entry 26 added to its register slice. Stage 1 detects a triad deterministically but cannot judge whether the content had three things, and a closer is where the padded ones land.
• tests/sample-structure.md and tests/sample-structure.html, built from headers that shipped past a clean report.
• Workflow step 2b, with the warning not to run stage 2 in the context that wrote the draft. A model reviewing its own prose is the actor that chose the words.

From the folded 0.2.2 branch:

• Earned exceptions now covers 17 of the 37 entries, in two tables, with the entry number on every row. It covered 5, all from the first block, while the second block carried its exceptions as prose buried inside individual entries.
• The paragraph naming what hides inside a watched phrase: superlatives, rankings, simultaneity, scope words, and the condition attached to a hedge.
• A dropped-claim check in workflow step 5 and as self-check 11, asked as a question and answered claim by claim. A dropped superlative leaves the paragraph looking intact, which is why a read-through misses it.
• Earned when notes on entries 25, 27, 28, 29 and 34, which had none.

Fixed

• A one-line paragraph that is a line of dialogue reported as a drama line break. tests/sample-clean.md now carries a four-line exchange as the regression control and still reports zero.
• aphorism was missing from CATEGORY_ORDER, so its hits sorted after the document-level density block instead of with the sentence-level tells.
• The reuse counter treated two rules matching the same span as two uses. Deduplicated on offset, so a genuine repeat on one line still counts twice.
• Entry 11 explained itself with "the negation-first shape wearing a costume", which is entry 37. The skill cannot ban a shape in one entry and use it in another. (From the folded 0.2.2 branch.)
• declaude_lint.py scanned markdown headings only, so every header rule was silent on an HTML draft: three thesis-shaped headers, a comma-clause header and a coy header shipped past a clean report. HTML is now flattened before scanning, and masthead .subtitle / .eyebrow / .post-meta elements are scanned as headings, because a subtitle is a header by every test that matters. Step 2 also now says to lint every string that reaches the reader, since a builder taking the subtitle as a CLI argument hides it from any file scan. (From #762.)

Changed

• Workflow step 2 no longer claims the linter misses every structural tic, which stopped being true with this release. It now says which third it does miss: staged paragraph shape, staged closers, dressed metaphor, and every earned exception in the tables. That third is the expensive one and it stays with the sentence pass.
• Step 2 shows --skip-quoted alongside the bare invocation. The evaluation run did not use it and spent two of its seven false positives on the draft's own quoted specimens.

Not shipped

A TF-IDF classifier over model-written and human-written corpora was considered and rejected. Bag-of-n-grams is a strictly weaker feature space than regex plus the positional checks already here, so it can surface no tic class this cannot; its output is a score with no span and no register entry, which does not fit a tool whose contract is candidates a reader can check; and it would learn "not this author" rather than "machine-shaped", which conflicts with the rule that the author's own writing outranks the register. It remains useful offline as a rule miner: train it, read the coefficients, hand-pick the real ones, ship regexes. The document-level repetition statistics above are the part of that idea that survives, and they need no corpus and no training. The model-shaped part of the problem it was aimed at is answered by stage 2 instead, which sends slots to a model that can name the tic and cite the entry rather than returning a score.

A widened bolded-bullet rule was also cut. It caught one real instance in the evaluation draft and fired seven times on this skill's own "Leave these alone" list, where the labels are exactly the real index the folded branch wrote into entry 29's earned form. The rule now tests for the restatement with a backreference, which is entry 29's actual tell, and the looser shape stays with the sentence pass. Two branches reached that boundary independently, one by reading the register and one by watching a rule misfire.

[0.2.1] - 2026-08-16

Other

• declauding 0.2.1 — register entry 37, dressed metaphor (#761)

[0.2.1] - 2026-08-16

Added

• Register entry 37, dressed metaphor: a figure of speech standing in for a mechanism you could have named ("wearing the costume of", "dressed up as", visceral imagery on a mundane observation). Entry 6 covers the locational special case; this is the general one. The entry states that no regex reaches it, and both its specimens shipped past a clean lint report, one of them into a patch headed for another project.
• Lint rule for mid-paragraph X is A, not B. The existing rule anchors on end-of-line, so it caught the construction only as a closer.
• Three specimens for entry 37 in tests/sample-tics.md, which now reports 62 candidates across 24 categories. Two of the three are invisible to the linter by design, which is the entry's point.

[0.2.0] - 2026-08-16

Other

• declauding v0.2.0 — absorb the encyclopedic and chatbot register from humanizer (#760)

[0.2.0] - 2026-08-16

Absorbs what a comparison against blader/humanizer v2.9.1 (MIT) showed this skill was missing. Humanizer packages the Wikipedia AI Cleanup project's Signs of AI writing; its coverage of encyclopedic and chatbot slop is broader than the two vocabulary entries this register had.

Measured before porting: a probe of 19 humanizer specimens produced 2 candidates from declaude_lint.py, neither for the right reason. It now produces 33 across 13 categories.

Added

• Register entries 24 to 36, in a second block that names itself as a different family from the staging mechanisms: copula avoidance, participle tail, forced triad, elegant variation, false range, inline-header list, typographic tells, chatbot residue, filler and hedge stacking, speculative gap-filling, diff-anchored documentation, subjectless fragment, predicate-position hyphenation. The block states that several of them are phrase lists rather than mechanisms and will leak.
• "Do not invent specifics" in SKILL.md, and a fourth failure mode in workflow step 5. De-vaguing a sentence is how a register pass fabricates, and the skill previously warned only against changing claims, not against supplying a name or number the source does not have.
• Voice-sample precedence. A sample of the author's writing outranks every rule here, including the em-dash density guard.
• Three invocation modes: pasted text, file, and embedded (another agent calling this as one step, which returns text and nothing else).
• "Leave these alone" section: the positive signals of human writing, and the things that are not tells on their own.
• Linter rules for the new lexical entries, plus Title Case heading detection and document-level curly-quote and emoji counts. Eleven new categories.
• 13 new specimens in tests/sample-tics.md, which now reports 58 candidates across 23 categories. tests/sample-clean.md stays at 0.

Changed

• Content preservation now licenses structural rearrangement: every claim survives, but paragraphs may merge or split and depth need not be uniform.

Fixed

• --skip-quoted blanked spans before lines, so a bold marker inside a table cell could mis-pair the italic regex across lines and leave that row's specimens visible. It also never blanked code, so a tell quoted in backticks reported as a hit. Fenced blocks and inline spans are blanked now, and the line pass runs first.
• The emoji count included U+2190 to U+21FF and U+2300 to U+23FF, so a plain arrow or a technical symbol in ordinary prose reported as decoration. Narrowed to the emoji-presentation blocks.

[0.1.1] - 2026-08-16

Other

• declauding v0.1.1 — add README, --skip-quoted, and the fixes that missed #755 (#756)

[0.1.1] - 2026-08-16

Added

• README.md. Missing from 0.1.0: the commits carrying it landed on the branch after the merge and were lost when the branch was deleted.
• --skip-quoted on the linter. Blanks blockquotes, table rows, *italic* spans and <q> elements while preserving line numbers, so a document that quotes bad prose as specimens does not report its own examples. Handles italics that wrap across lines.

Fixed

• The earns/wants/demands agency rule fired on ordinary second-person prose ("if you want the noise"). Added a lookbehind for personal pronouns.
• Density checks counted headings, table rows and list bullets as sentences, inflating fragment density on any structured document.
• The skill's own prose failed its own linter: two coy headers, four X, not Y closers, one abstraction-agency subject, and a negation-first closer sitting directly beneath the sentence explaining why negation-first closers are a tic.

[0.1.0] - 2026-08-16

Added

• Initial release. Two modes: clean rewrite (default) and annotated HTML diff.
• references/register.md: 23 entries grouped by mechanism, each with surface tell, why, fix, and a before-and-after from a real published draft. Grouping is by mechanism rather than phrase because phrase blocklists miss the next paraphrase.
• references/annotating.md: spec for the annotated diff.
• assets/annotated.template.html: self-contained output template. No build step, no CDN, opens from file://.
• scripts/declaude_lint.py: stdlib-only scan for lexical tells plus header shape, one-line-paragraph beats, fragment runs, em-dash density and sentence-length monotony. Exits 1 on candidates.
• tests/sample-tics.md (29 candidates, 12 categories) and tests/sample-clean.md (0). The clean corpus is human-written prose and holding it at zero is the linter's design constraint.
• Overcorrection guard and earned-exception table in SKILL.md.
• Workflow steps 1 and 6: read the whole piece first, and report contradictions separately rather than fixing them silently.

Other

• declauding v0.1.0 — LLM prose tics in, human technical prose out (#755)