All notable changes to this project are documented here.
- Tier 3 phrases — multi-word boilerplate that's individually unobjectionable but stacks heavily in AI-generated crypto/web3/DePIN/AI-infra content:
emerging sector,the integration of,the intersection of,community-driven,long-term sustainability,user engagement,decentralized compute,sustainable reward emissions,tokenized incentive structures,designed for long-term. Flagged by per-phrase density (≥2 repetitions) or cluster (≥3 distinct phrases in one piece — the LLM-varies-its-own-boilerplate shape). - Generic future-narrative closers — "May become one of the most important narratives of the next market cycle" template family. Modal + "become" + (one of) the most + (narrative / story / trend / theme / chapter / movement).
- Hedge-stacked predictions —
could potentially,may eventually,might ultimately. Modal + hedge adverb stack where each word cancels the next. - "Real/actual" adjective inflation —
real on-chain tokenomics,actual reward sustainability,genuine utility,true product-market fit. The noun-modifier form distinct from the existing sentence-level hollow-intensifier rule. - Hashtag stuffing — trailing blocks of 6+ hashtags on short posts, especially when mixing one project tag with broad category tags (#AI #Crypto #Web3 #Innovation #FutureTech).
- Bullet lists of bare noun phrases — 5+ consecutive bullets where each is a short adj+noun pair with no verb. Detector heuristic excludes genuine list content (verbs in items, ingredient lists, changelog entries).
- Emotional flatline — extended to cover the bare section-header variant: "Interesting part of the project:" / "Interesting thing here:" — same role as "the most interesting part" but as a header opener.
- Severity tiers — all six new categories wired into P0/P1/P2 ladder (hashtag stuffing varies by profile; the rest are P1, with phrase repetition at P2).
- Context profiles tolerance matrix — added rows for all six new categories so the
linkedinanddocsprofiles don't false-positive on legitimate use (e.g., bullet-NP lists relaxed ontechnical-bloganddocssince technical option lists are correctly bare-NP). - "6+" hashtag threshold — added rationale paragraph explaining the empirical floor.
- "Real/actual" inflation — added named-contrast carve-out so honest contrastive writing ("real on-chain settlement, not bridged IOUs") isn't flagged.
- Version bump to 3.4.0.
- A user of the avoid-ai-writing extension flagged two crypto-shill social posts (MineBench reviews) that the v3.3.x wordlist+regex detector scored as "Minimal AI signals" despite being obvious LLM output. Both posts avoided every Tier 1 vocabulary entry by substituting synonyms ("emerging sector," "scalable network contribution," "viability") and used structural shapes (hashtag block, bare-NP bullet lists, hedge stacks, future-narrative templates) the detector had no rule for. v3.4 adds rules for the structures, not just the words.
- "Worth [verb]ing" vague endorsement pattern:
worth reading,worth paying attention to,worth a look,worth exploring,worth checking out,worth your time— broadens existing "it's worth noting that" to the full family - Reader-steering frames:
Here's what's interesting,Here's what caught my eye,Here's what stood out— added to both transition phrases and confidence calibration sections with context on when the pattern is a genuine problem vs. when data-backed usage is acceptable
- Version bump to 3.3.0
- Detect mode: flag-only mode that identifies AI patterns without rewriting. Trigger with "detect," "flag only," "audit only," "just flag," "scan," or similar. Returns issues grouped by severity (P0/P1/P2) plus an assessment of which flags are clear problems vs. judgment calls. Useful when flagged patterns are intentional, when auditing published or third-party content, or when you want a quick scan without a full rewrite.
- Output format section now documents both rewrite (default) and detect mode outputs
- Version bump to 3.2.0
- 3 new Tier 1 words from Pangram AI detection research:
keen(as intensifier),symphony(metaphor),embrace(metaphor) - 2 new template phrases: "Whether you're X or Y" (false-breadth), "I recently had the pleasure of" (review/social AI pattern)
- "In summary" added to transition phrases (alongside existing "In conclusion" / "To summarize")
- Structure-priority note in Rhythm section: structural regularity is the #1 signal AI detectors weight, above vocabulary
- Over-polishing warning: aggressive editing can push writing toward AI statistical profiles by removing natural disfluency
- Total vocabulary: 106 → 109 entries (60 Tier 1 + 38 Tier 2 + 11 Tier 3)
- Template phrases: 2 → 4 entries
- Pangram Labs AI detection research (pangram.com) — decoder-only classifier trained on 28M human documents. Key insight: structural uniformity and pacing consistency are weighted higher than individual word choices.
- Novelty inflation pattern (AI treats established concepts as speaker inventions)
- False concession structure pattern
- Rhetorical question openers pattern
- Parenthetical hedging pattern
- Numbered list inflation pattern
- Severity tiers (P0/P1/P2) for prioritized auditing
- Self-reference escape hatch (exempts quoted examples from flagging)
- Context profiles with tolerance matrix (linkedin, blog, technical-blog, investor-email, docs, casual)
- Auto-detection cues for context inference
- Extended frontmatter: license, compatibility, author, tags, agentskills_spec
- Pattern count: 30 → 35 categories
- OpenClaw compatibility — added
versionandmetadata.openclawto SKILL.md frontmatter - OpenClaw installation instructions in README (ClawHub and manual)
- Skill now works with both Claude Code and OpenClaw from a single
SKILL.md
README.md— broadened description to reference both platforms, reorganized installation into Claude Code and OpenClaw sections
- 5 new pattern categories: reasoning chain artifacts, sycophantic tone, acknowledgment loops, confidence calibration phrases, excessive structure
- New "Rhythm and uniformity" section — checks for sentence length uniformity, paragraph length uniformity, missing first-person perspective, and read-aloud test guidance
- New "When to rewrite from scratch vs. patch" threshold — advises full rewrites when AI density is too high for patching
- 5 rewrite principles in tone calibration section (vary length, be concrete, have a voice, cut neutrality, earn emphasis)
- New "Meta Patterns" group in README pattern table
- Expanded credits: OpenClaw humanizer ecosystem (community patterns)
- Pattern count: 23 → 30 categories
README.md— updated pattern count, added Meta Patterns table, expanded credits with source descriptions- Communication Patterns table in README now includes all communication patterns
- Tiered vocabulary system — words are now organized into three tiers based on AI-signal strength:
- Tier 1 (always flag): 53 entries — dead giveaways that appear 5–20x more often in AI text
- Tier 2 (flag in clusters): 38 entries — legitimate words that signal AI when 2+ appear in the same paragraph
- Tier 3 (flag by density): 11 entries — common words that only flag when the text is saturated with them
- 39 new vocabulary entries across all tiers, including: bustling, intricate, complexities, ever-evolving, daunting, holistic, actionable, impactful, learnings, thought leadership, best practices, synergy, interplay, encompass, catalyze, reimagine, galvanize, augment, cultivate, illuminate, elucidate, juxtapose, paradigm-shifting, transformative, cornerstone, paramount, poised, burgeoning, nascent, quintessential, overarching, underpinning, significant, innovative, dynamic, scalable, compelling, unprecedented, sophisticated, instrumental, world-class
- Credit to brandonwise/humanizer for tiered vocabulary research
- Word/phrase table reorganized from flat list to tiered structure with usage guidance
- Total vocabulary: 58 → 102 entries (53 Tier 1 + 38 Tier 2 + 11 Tier 3)
README.md— updated replacement table description, pattern table, and credits
- 15 new word/phrase replacements: nuanced, crucial, multifaceted, ecosystem, myriad, plethora, deep dive/dive into, unpack, bolster, spearhead, resonate, revolutionize, facilitate, underpin
- New pattern category: "let's" constructions (false-collaborative openers like "let's explore," "let's break this down")
- Skill now covers 23 pattern categories with 58 word/phrase replacements
- Deduplicated filler phrases that appeared in both the word table and the filler section
README.md— updated pattern count (22 → 23), replacement table count (43 → 58), added "let's" constructions row to pattern table
- Em dash detection now catches double-hyphen (
--) in addition to Unicode em dash (—) README.md— updated formatting pattern description to mention--
- New pattern category: emotional flatline (AI claims emotions as structural crutch without conveying them; also flags lazy human writing)
- Skill now covers 22 pattern categories with 43 word/phrase replacements
- 8 new pattern categories: notability name-dropping, superficial -ing analyses, promotional language, formulaic challenges, false ranges, inline-header lists, title case headings, cutoff disclaimers
- 5 new word table entries (nestled, vibrant, thriving, despite challenges, showcasing)
- Skill now covers 21 pattern categories with 43 word/phrase replacements
README.md— expanded full example (6 paragraphs → 4 clean sentences, 40+ tells flagged); added per-pattern before/after table organized into Content, Language, Structure, Communication groups; updated pattern count and replacement table count throughout
SKILL.md— Claude Code skill with 13 pattern categories: formatting, sentence structure, word/phrase replacements (38 entries), template phrases, transition phrases, structural issues, significance inflation, copula avoidance, synonym cycling, vague attributions, filler phrases, generic conclusions, chatbot artifacts- Four-section output format: issues found, rewritten version, what changed, second-pass audit
README.md— installation guide (3 methods), full pattern reference, usage examplesLICENSE— MIT.gitignore— OS/editor exclusions