Add humanizer-classics: book-grounded writing skill (v3 fork)

humanizer-classics/.github/ISSUE_TEMPLATE/new-book.md

@@ -0,0 +1,67 @@
+---
+name: New book proposal
+about: Propose adding a new foundational writing book to the canon
+title: "[New book] "
+labels: book-proposal
+---
+
+## Book
+
+**Title:** *...*
+**Author:** ...
+**Edition + year:** ...
+**Publisher:** ...
+**ISBN (optional):** ...
+
+## Why this book belongs in humanizer-classics
+
+3-5 sentences on what this book teaches that the existing rules don't already cover. What's the unique angle? Who is this book for? When does its advice matter most?
+
+## Proposed rule prefix
+
+E.g., `W` is taken (Wikipedia), `Z` is taken (Zinsser), `S` is taken (Strunk & White), `H` is taken (HBR Guide). Pick a single letter not already used — usually the first letter of the author's last name. Confirm by checking `references/_rule-index.md`.
+
+## Proposed rule list
+
+Sketch 3-7 rules you'd add from this book. Each should have:
+
+- **Rule name** (one line, imperative)
+- **Source** (chapter, page)
+- **What it adds** (why this rule isn't already covered by Z/S/H/W)
+
+Format:
+
+| ID | Rule name | Source | What it adds |
+|----|-----------|--------|--------------|
+| X-1 | ... | ch. N | ... |
+| X-2 | ... | ch. N | ... |
+| X-3 | ... | ch. N | ... |
+
+## Voice and pull-quote sample
+
+Paste 1-2 short pull-quotes (10-25 words each) you'd plan to cite. This helps reviewers gauge whether the book's voice fits the project.
+
+## License / fair use
+
+Confirm:
+
+- [ ] You can cite the book under fair use for short pull-quotes (10-25 words for educational commentary)
+- [ ] You're not proposing to reproduce extended excerpts, chapter summaries, or proprietary material
+
+## Plan
+
+- [ ] I plan to submit the PR myself
+- [ ] I'm flagging this for a maintainer to pick up
+
+If submitting yourself, see `CONTRIBUTING.md` for the file format and the acceptance bar. The PR should include:
+
+- A new file at `references/<author>-<short-title>.md`
+- Updates to `references/_rule-index.md` (add the new rule rows + cross-references)
+- Updates to `SKILL.md` (add the new book's rules to the catalog and the references list)
+- Updates to `README.md` ("Books currently included")
+- At least one new corpus sample at `tests/corpus/` exercising rules from this book
+- A version bump (minor — e.g., 3.0.x → 3.1.0)
+
+## Acknowledgments
+
+Anyone you'd like credited (yourself, the book's author, anyone whose review helped shape the proposal).

humanizer-classics/.github/ISSUE_TEMPLATE/new-rule.md

@@ -0,0 +1,73 @@
+---
+name: New rule proposal
+about: Propose a new rule for an existing book file (or for the Wikipedia detection layer)
+title: "[New rule] "
+labels: rule-proposal
+---
+
+## Source
+
+Which book is this rule from?
+
+- [ ] Zinsser, *On Writing Well*
+- [ ] Strunk & White, *The Elements of Style*
+- [ ] Garner, *HBR Guide to Better Business Writing*
+- [ ] Wikipedia, *Signs of AI writing* (detection layer)
+- [ ] Other (please specify and consider opening a `new-book` issue instead)
+
+**Specific reference:** chapter, section, or page number — be precise so the rule can be traced back to the source.
+
+## Proposed rule ID
+
+E.g., `Z-7`, `S-6`, `H-6`. Check `references/_rule-index.md` to confirm the next available number.
+
+## One-line rule name
+
+Imperative or prescriptive form. E.g., "Cut throat-clearing openers" or "Use the active voice."
+
+## Pull-quote (10-25 words from the source)
+
+> "..."
+> — Author, ch. N
+
+## Detection cue
+
+What pattern in the input text signals this rule should fire? Be specific — keywords, sentence shapes, or structural tells.
+
+## Problem
+
+2-4 sentences. What does the LLM do wrong? Why does it sound off to a human reader?
+
+## Before / After
+
+**Before**
+> [Realistic AI-generated text — not a strawman]
+
+**After**
+> [The same passage rewritten following the rule]
+
+## Cross-references
+
+Which existing W-rules does this rule help fix? Which other book rules does it interlock with?
+
+## Context tags
+
+Which contexts should this rule fire in? Pick from: `all`, `memo`, `email`, `blog`, `book-draft`, `technical-doc`, `dictation`, `meeting-notes`. Justify if you're picking a non-obvious set.
+
+## Why this rule isn't a duplicate
+
+How does this rule add something the existing rules don't already cover? If it's a sharper version of an existing rule, why does the project benefit from both?
+
+## Acceptance bar self-check
+
+- [ ] Traceable to a specific page or chapter
+- [ ] Maps to a real AI-failure mode
+- [ ] Adds something the existing rule set doesn't cover
+- [ ] Before/after example is realistic, not a strawman
+- [ ] Detection cue is specific enough to actually find the pattern
+- [ ] "How to apply" gives a mechanical move
+
+## I plan to submit a PR
+
+- [ ] Yes, I'll open the PR (one rule per PR per `CONTRIBUTING.md`)
+- [ ] No, I'm flagging this for a maintainer to pick up

humanizer-classics/CHANGELOG.md

@@ -0,0 +1,64 @@
+# Changelog
+
+All notable changes to humanizer-classics. Format roughly follows [Keep a Changelog](https://keepachangelog.com/).
+
+## [3.0.0] — 2026-04-29
+
+### Added
+
+- Initial v3 release as a fork of `humanizer` v2.5.1.
+- New architecture: slim `SKILL.md` dispatcher (~250 lines) + per-source `references/` files lazy-loaded as rules fire.
+- 24 new craft rules sourced from foundational writing books, each with a citation and pull-quote verified against the source PDFs:
+  - **Zinsser, *On Writing Well*** (25th Anniversary Edition, 6th ed., HarperResource, 2001) — Z-1 through Z-10:
+    - Z-1: Cut clutter — every word that does no work (Ch. 3)
+    - Z-2: Use short, plain, Anglo-Saxon words (Ch. 3, Ch. 6)
+    - Z-3: Active verbs do the work; kill nominalizations (Ch. 10)
+    - Z-4: Strip qualifiers (Ch. 10 — Little Qualifiers)
+    - Z-5: Be present on the page; have a self (Ch. 4)
+    - Z-6: Endings matter — quit when the work is done (Ch. 9)
+    - Z-7: A lead must capture the reader immediately (Ch. 9)
+    - Z-8: Maintain unity of pronoun, tense, and mood (Ch. 8)
+    - Z-9: Most adjectives and adverbs are unnecessary (Ch. 10)
+    - Z-10: Write for yourself, not for an imagined mass audience (Ch. 5)
+  - **Strunk & White, *The Elements of Style*** (4th ed., 1999) — S-1 through S-9:
+    - S-1: Omit needless words (II.17)
+    - S-2: Use the active voice (II.14)
+    - S-3: Put statements in positive form (II.15)
+    - S-4: Use definite, specific, concrete language (II.16)
+    - S-5: Do not overstate (V.7)
+    - S-6: Express coordinate ideas in similar form — parallel construction (II.19)
+    - S-7: Place the emphatic words of a sentence at the end (II.22)
+    - S-8: Avoid a succession of loose sentences — mechanical singsong (II.18)
+    - S-9: Do not affect a breezy manner — calibrating partner to Z-5 (V.9)
+  - **Garner / HBR, *HBR Guide to Better Business Writing*** (1st ed., 2012) — H-1 through H-5:
+    - H-1: Lead with the bottom line (pyramid principle)
+    - H-2: Write for the skim-reader
+    - H-3: One idea per paragraph
+    - H-4: Imperative for instructions
+    - H-5: Cut throat-clearing openers
+- `context_tags:` field on every rule for conflict resolution across forms (memo, email, blog, book-draft, technical-doc, dictation, meeting-notes).
+- Cross-reference graph in `references/_rule-index.md` mapping detection rules (W) to fix rules (Z/S/H).
+- Granola live integration via `mcp__granola__list_meetings`, `mcp__granola__get_meeting_transcript`, etc. Workflow documented in `references/granola-meeting-transcripts.md`.
+- Wispr Flow dictation guidance (no MCP integration; comes through as pasted text).
+- `tests/corpus/` with golden samples + `tests/REVIEWING.md` for manual reviewer checklist.
+- `CONTRIBUTING.md` with one-rule-per-PR norm, rule acceptance bar, pull-quote licensing.
+- GitHub issue templates: `new-rule.md`, `new-book.md`.
+
+### Changed
+
+- 29 Wikipedia "Signs of AI writing" rules renamed from numeric IDs (1-29) to prefixed IDs (W-1 through W-29) and lifted verbatim into `references/wikipedia-signs-of-ai-writing.md`. Content unchanged.
+- Two-pass humanization process now context-aware: the chosen context tag governs which book rules fire. (E.g., Z-5 "be present on the page" doesn't fire on memo or technical-doc contexts; H-1 "lead with bottom line" doesn't fire on book-draft contexts.)
+- Voice calibration section retained from v2.x verbatim.
+- Personality and soul section retained from v2.x verbatim.
+
+### Migrated from v2.x
+
+- All 29 rules (Wikipedia "Signs of AI writing")
+- Voice calibration feature
+- Two-pass audit process
+- Personality and soul guidance
+- Full example with before/draft/audit/after
+
+### Note on coexistence with v2.x
+
+`humanizer-classics` is a fork, not a replacement. `humanizer` v2.5.1 remains stable for current users at `~/.claude/skills/humanizer/`. v3 installs alongside as `~/.claude/skills/humanizer-classics/`. Future work may merge them once v3 has proven out.

humanizer-classics/CONTRIBUTING.md

@@ -0,0 +1,122 @@
+# Contributing to Humanizer Classics
+
+Thanks for helping make this a better tool for people who want their AI-assisted writing to sound human. This is meant to be an ongoing repository — new books, new rules, refinements to existing ones — and contributions from people who actually do this work for a living are the point.
+
+## What you can contribute
+
+| Contribution | Effort | PR shape |
+|--------------|--------|----------|
+| Add a new rule to an existing book | Small | One rule per PR, ~30 lines |
+| Refine an existing rule (better example, sharper detection cue) | Small | One rule per PR |
+| Add a new book reference file | Medium | One book per PR, 3-7 rules |
+| Add a new corpus sample to `tests/corpus/` | Small | One sample pair per PR |
+| Refine the Granola/Wispr workflow | Medium | Update `references/granola-meeting-transcripts.md` |
+| Propose architectural changes | Large | Open an issue first, discuss before PR |
+
+## One rule per PR
+
+The repo follows a *one rule per PR* norm. This matches the cadence the upstream `humanizer` repo has used since v2.0 (e.g., commits "Add passive voice rule (#80)", "feat: add hyphenated word pair overuse pattern (#42)"). Small PRs are easier to review, easier to revert, and force each rule to stand on its own.
+
+If your rule depends on another change (e.g., a new book file), bundle the dependency into the same PR — but keep the *first new rule* in the same PR as the book file, so reviewers can see the rule format in context.
+
+## How to add a new rule
+
+1. **Decide which book file it belongs to.** If the rule is sourced from one of the existing books, add it there. If it's from a new book, copy `references/_template-book-rules.md` to `references/<author>-<short-title>.md` and start with that template.
+
+2. **Follow the rule format.** Every rule has these sections, in order:
+   - `### <ID>. <One-line rule name in imperative or prescriptive form>`
+   - `**Source:**` — book, edition, chapter, page (as specifically as you can)
+   - **Pull-quote** — 10-25 words from the source, attributed
+   - `**Cross-references:**` — which W-rules this helps fix; which other book rules it interlocks with
+   - `**Context tags:**` — `all` or any combination of `memo`, `email`, `blog`, `book-draft`, `technical-doc`, `dictation`, `meeting-notes`
+   - `**Detection cue:**` — what pattern in the text signals this rule should fire
+   - `**Problem:**` — 2-4 sentences on the failure mode and why it sounds off
+   - `**Before** / **After**` — a realistic, non-strawman example pair
+   - `**How to apply:**` — the mechanical move; 1-3 sentences a writer can run on autopilot
+
+3. **Update `references/_rule-index.md`** — add a row to the rule-ID table and (if applicable) a row to the cross-reference graph.
+
+4. **Add a corpus sample if the rule introduces a new pattern.** Drop a `*.md` + `*.expected-fixes.md` pair into `tests/corpus/` showing the rule firing.
+
+5. **Bump the version.** See *Versioning* below.
+
+6. **Run the validator and manual review** before opening the PR:
+   ```bash
+   cd humanizer-classics/
+   python3 tests/validate.py
+   ```
+   The validator catches structural problems (broken references, missing sections, inconsistent rule IDs) automatically. Then run the manual review in `tests/REVIEWING.md` to catch behavioral issues.
+
+## How to add a new book
+
+1. **Copy the template.** `cp references/_template-book-rules.md references/<author>-<short-title>.md`. Use kebab-case: `williams-style-lessons-in-clarity-and-grace.md`, `pinker-sense-of-style.md`.
+
+2. **Pick a single-letter rule prefix not already used.** Currently used: W, Z, S, H. The prefix should be the first letter of the author's last name when possible.
+
+3. **Write 3-7 rules following the format above.** 5 is the sweet spot — enough to be useful, few enough to review thoroughly.
+
+4. **Update `SKILL.md`** — add the new book's rules to the *Craft rules* table and to the references list at the bottom.
+
+5. **Update `references/_rule-index.md`** — add the new rule rows.
+
+6. **Update `README.md`** — add the book to the "Books currently included" list.
+
+7. **Add corpus samples** that exercise the new rules.
+
+8. **Bump the version** (minor bump for a new book — see *Versioning*).
+
+## Rule acceptance bar
+
+A rule is ready to merge when:
+
+- ✅ It traces to a specific page or chapter in a real, citable source
+- ✅ It maps to a real AI-failure mode (not just generic "good writing" advice)
+- ✅ It does something the existing rules don't already cover, or it covers them better with a positive prescription
+- ✅ The before/after example is realistic AI-generated text, not a strawman
+- ✅ The detection cue is specific enough that a reader (or a model) can find the pattern
+- ✅ The "How to apply" gives a mechanical move, not a vague exhortation
+
+A rule is **not** ready when:
+
+- ❌ It restates a Wikipedia detection rule without adding a positive fix
+- ❌ The example is contrived (no real AI writes that way)
+- ❌ The pull-quote can't be traced to a specific edition + page
+- ❌ The rule conflicts with another rule and there's no `context_tags` resolution
+- ❌ It's about taste rather than craft (e.g., "use the Oxford comma" — that's style preference, not AI-failure)
+
+## Pull-quote licensing
+
+Pull-quotes from cited books are short excerpts (10-25 words) used for educational commentary, which is fair use under U.S. copyright law. Don't include longer excerpts. Don't reproduce entire chapters or large structural elements. When in doubt, paraphrase and cite.
+
+If a quote is approximate or paraphrased rather than exact, mark it `(paraphrased)` after the citation.
+
+## Conflict resolution
+
+When two rules disagree (e.g., Zinsser's Z-5 "be present on the page" wants first person, but H-1 "lead with bottom line" governs a memo where first person is wrong), the **`context_tags:`** field on each rule resolves the conflict. The rule applies only when the input's context tag is in its tag list.
+
+Current tags: `all`, `memo`, `email`, `blog`, `book-draft`, `technical-doc`, `dictation`, `meeting-notes`. Propose new tags via PR if the existing set doesn't fit.
+
+## Versioning
+
+This project follows [semver](https://semver.org/):
+
+- **Major bump (e.g., 3.x.x → 4.0.0)** — breaking architecture change. Don't introduce these without an issue and discussion.
+- **Minor bump (e.g., 3.0.x → 3.1.0)** — new book added, or a structurally significant new feature (e.g., a new context tag).
+- **Patch bump (e.g., 3.0.0 → 3.0.1)** — new rule in an existing book, refinement of an existing rule, doc fixes.
+
+Update both `SKILL.md` (frontmatter `version:` field) and `CHANGELOG.md` in the same PR.
+
+## Issue templates
+
+When opening an issue, use one of:
+
+- **`new-rule`** — propose a new rule
+- **`new-book`** — propose a new book
+
+Templates live at `.github/ISSUE_TEMPLATE/`.
+
+## Code of conduct
+
+Be specific. Cite sources. Don't argue from taste — argue from text. The discussion should be about whether the rule actually catches a real AI-failure mode and whether the fix actually helps.
+
+If you disagree with an existing rule, open an issue with the specific text it mishandles. Concrete disagreements move the project forward.