@cratis/pi 6.3.2 → 6.5.0

package/package/corpus/skills/cratis-writing-voice-and-cadence/references/ai-tells.md

@@ -0,0 +1,419 @@
+# AI-associated writing patterns
+
+This is the full catalog behind the **cratis-writing-voice-and-cadence** skill: patterns that
+make writing read as machine-produced, how a script can find each one, and what a person
+would usually write instead. Sources were checked in September 2026. Bracketed IDs such as
+HV-10 point to the [Sources](#sources) table at the end of this file.
+
+Read the skill's non-detector policy before using any of this. Every check below finds a
+passage for a person to reread. None of them, alone or combined, tells you who wrote a text.
+Readers do poorly at that task (HV-01, HV-03), the cues they rely on are unreliable (HV-02),
+and the studies that do find differences between human and model text find them in aggregate
+over many texts, within the genres and models they tested (HV-05, HV-06). Detector vendors
+publish prevalence figures, but those are the vendor's own classifications of posts its
+searches found, not verified authorship (HV-21, HV-22).
+
+## How much weight each entry carries
+
+Most of this catalog is editorial. The "Basis" line on each entry says which of three kinds
+it is.
+
+"Study" means a verified study found the pattern elevated in model text, in aggregate, in a
+specific genre. It still can't classify one piece. "Guide" means a published editors' guide
+lists it as a possible sign, and that guide says its signs are context-dependent
+observations that prove nothing on their own (HV-23). "Editorial" means no verified study backs it as an AI
+signal. Those entries are here because they recur in generated copy and usually stand in for
+missing substance, so judge them on that alone.
+
+## Running the checks
+
+- Keep the original text of each piece, with file, line and character offsets. Match the
+  lexical, syntactic and tonal patterns against a prose view that leaves out front matter,
+  fenced and inline code, URLs, block quotes and quoted strings, because code and quotations
+  aren't prose the writer chose. Map every match back to its original offset.
+- Some checks need what the prose view leaves out, so they read the original. L4 looks for a
+  link, citation or named source in the original sentence. T3 classifies the original first
+  sentence, backtick spans and calls included. L1, L5 and T5 read the original sentence to
+  decide whether it has a number, identifier or code. N5 compares the complete before and
+  after versions.
+- Treat straight (`'`) and curly (`’`) apostrophes alike. Patterns that depend on a
+  contraction or possessive use `['’]`. A check that normalizes apostrophes does it in a
+  comparison copy and keeps the original text and offsets.
+- Patterns are JavaScript regular expressions. Use the `i` flag unless the entry says
+  otherwise, `g` to count, `m` for line-anchored patterns, and `u` where noted.
+- Report each match with its sentence, file and line. Don't sum matches into a score, don't
+  label a piece "AI", and don't rewrite automatically.
+- The thresholds are starting defaults nobody has calibrated. Before relying on one, run it
+  over material you know people wrote and look at what it flags.
+- A batch means ten comparable pieces: same channel, same content type. With fewer than five,
+  report per-piece results and skip the batch checks.
+- A material limitation is never a cadence problem. If a check flags a sentence that states a
+  real limit, the fix is to move or rephrase it, never to drop it.
+
+## Lexical
+
+### L1. Inflated significance
+
+Words that assert importance without showing any: *pivotal*, *groundbreaking*,
+*transformative*, *game-changing*, *a testament to*, *plays a key role*.
+
+```text
+\b(?:pivotal|groundbreaking|transformative|revolutioni[sz](?:e|es|ed|ing)|game[- ]chang(?:er|ers|ing)|(?:a |stands as a )?testament to|plays? an? (?:key|vital|crucial|pivotal|central) role)\b
+```
+
+Flag two or more in one piece, or one in a sentence with no number, identifier or named
+thing nearby. Replace the word with what it was pointing at: the behavior, the measurement,
+the decision, and how far it reaches. Basis: guide (HV-25), with the HV-23 caveat.
+
+### L2. Abstract scene-setting
+
+Openings that set a stage nobody asked for: *in today's fast-paced world*, *the evolving
+landscape*, *tapestry*, *the realm of*, *paradigm shift*.
+
+```text
+\b(?:(?:ever-?)?(?:evolving|changing|shifting) landscape|in today['’]?s (?:fast-paced|digital|modern|ever-changing) world|tapestry|(?:in|into) the realm of|paradigm shift)\b
+```
+
+Flag any match in the first two sentences, or the same phrase in two or more pieces of a
+batch. Start instead with the engineering question or the symptom somebody saw. Basis:
+guide for "evolving landscape" (HV-25); editorial for the rest.
+
+Fixtures: the pattern matches "In today's digital world" and "In today’s digital world",
+and does not match "Today’s release fixes the importer".
+
+### L3. Polished filler verbs and adjectives
+
+*Delve*, *underscore* and *showcase* rose sharply in one large corpus of biomedical abstracts
+after 2023 (HV-10); the same analysis estimated that at least 13.5% of 2024 abstracts were
+processed with a language model (HV-09). Neither result was measured on developer writing or
+social posts. *Robust*, *seamless*, *leverage*, *unlock*, *elevate*, *empower* and
+*streamline* are unvalidated clichés in the same family.
+
+```text
+\b(?:delv(?:e|es|ed|ing)|underscor(?:e|es|ed|ing)|showcas(?:e|es|ed|ing)|robust(?:ly|ness)?|seamless(?:ly)?|leverag(?:e|es|ed|ing)|unlock(?:s|ed|ing)?|elevat(?:e|es|ed|ing)|empower(?:s|ed|ing)?|streamlin(?:e|es|ed|ing))\b
+```
+
+Count per 1,000 words and show every match in context. Starting default: flag three or more
+per 1,000 words, or one stem that recurs across three or more pieces in a batch. Use the
+direct verb and its object: "runs this query", "fails on replay". Leave a term alone when it
+carries its technical meaning, as in robust statistics or a lever in a physical system.
+Basis: study for the first three (HV-10); editorial for the rest.
+
+### L4. Unattributed consensus
+
+*Experts say*, *studies show*, *developers want*, *everyone knows*.
+
+```text
+\b(?:experts|studies|research|developers|engineers|everyone|many (?:people|teams)|most (?:people|teams|developers))\s+(?:say|says|show|shows|agree|know|want|believe|suggest|suggests)\b
+```
+
+Flag any match whose sentence has no link, citation or named source. Name the study, the
+survey question and its population, or present the point as the author's opinion. Basis:
+editorial.
+
+### L5. Empty assurances
+
+*Easy*, *simple*, *secure*, *production-ready*, *best-in-class*, *blazing fast*.
+
+```text
+\b(?:easy|easily|simple|simply|effortless(?:ly)?|secure|production[- ]ready|best[- ]in[- ]class|enterprise[- ]grade|blazing(?:ly)?[- ]fast)\b
+```
+
+Flag when the sentence has no prerequisite, version, threat model or comparison. Describe the
+path that was actually verified and where it stops working. Any claim of security or
+readiness also has to clear the owning product's claim review; rewording doesn't. Basis:
+editorial; PostHog's voice guide makes the same argument for specifics and candor about
+limits (DP-05).
+
+## Syntactic
+
+### S1. Corrective tail and reflexive contrast
+
+A sentence that ends by correcting itself ("…, not a fresh count."), the "not only X but Y"
+frame, the "It's not X, it's Y" reflex, and "rather than" doing the same job.
+
+```text
+,\s+not\s+[^,.;:!?]{1,40}[.!?]
+\bnot\s+(?:only|just|merely)\b[^.!?]{1,80}\bbut\b
+\b(?:it|this|that)(?:'s|’s|\s+is)\s+not\b[^.!?,;:—–]{1,60}[,;:—–.]\s*(?:it|this|that)(?:'s|’s|\s+is)\b
+\brather\s+than\b
+```
+
+Starting default: flag two or more in a piece under 300 words, four or more per 1,000 words
+in longer ones, or the same construction closing a paragraph in five or more pieces of a
+batch. State what the thing does first. Keep a contrast when the reader would otherwise
+confuse the two things. Basis: editorial.
+
+### S2. Balanced semicolon and em-dash pivot
+
+Two clauses of equal weight on either side of a semicolon, and a dash that swings into a
+short closing clause.
+
+```text
+\b\w+;\s+[a-z]
+(?:\s?[—–]\s?|\s--\s)[^—–.!?\n]{1,40}[.!?](?=\s|$)
+```
+
+Presence means nothing; people use both. Compare density and position across the batch.
+Flag three or more in one short piece, or the same construction in the last sentence of a
+paragraph in more than half the batch. Replace it with the connective that says what the
+relationship is: because, so, unless, which means. Basis: editorial.
+
+### S3. Participial tails and noun stacks
+
+A sentence that trails a ", making it…" clause, and runs of nominalizations: "the
+implementation of the normalization of the configuration".
+
+```text
+,\s+(?!(?:including|according|during|following|regarding|depending|assuming|using|excluding|starting)\b)\w{3,}ing\b
+\b\w{3,}(?:tion|sion|ment|ity|ance|ence|ness)s?\b
+```
+
+Flag a piece where more than a third of sentences end in a participial tail. Count the second
+pattern per sentence and flag sentences with four or more. Put the actor and the action in a
+finite verb and keep the technical terms that matter. Basis: a corpus study found aggregate
+grammatical differences between human and model text (HV-05), limited to its genres and
+models; the specific constructions here are an editorial reading.
+
+### S4. Reflexive triplets and mirrored clauses
+
+Three parallel items where the thought had one or two, and runs of sentences built to the
+same template.
+
+```text
+\b[\w'-]+(?:\s+[\w'-]+){0,2},\s+[\w'-]+(?:\s+[\w'-]+){0,2},?\s+(?:and|or)\s+[\w'-]+
+```
+
+The pattern is noisy by design. Flag three or more triplets in a piece under 500 words, or
+one in each of three adjacent paragraphs. For mirrored clauses, split into sentences and flag
+three consecutive sentences that start with the same word, or whose word counts sit within
+two of each other. Keep every item that is independently useful and cut the ceremonial third.
+Basis: editorial.
+
+### S5. Punchline fragment
+
+A two- or three-word sentence landing a slogan: "They run." "Every time."
+
+No regex. Split paragraphs into sentences on `(?<=[.!?])\s+`, then flag sentences of one to
+three words that end a paragraph or follow a sentence of twelve words or more. Skip headings,
+list items and code. Flag two or more in a piece, or a paragraph-final fragment in three or
+more pieces of a batch. Fold the point into the sentence it punctuates, or cut it if that
+sentence already made it. Basis: editorial.
+
+### S6. Self-answered questions, signposts and pointer sentences
+
+"Why bother? Because…", "Here's the thing", "That's the X. Here's the Y.", "Let's dive in",
+and "That is what X is for" standing in for an explanation.
+
+```text
+\?\s+(?:Because|Simple|Easy|Here(?:'s|’s| is)|The answer|It turns out|Turns out)\b
+\b(?:Here(?:'s|’s| is) (?:the (?:thing|kicker|catch|twist|deal|point)|why|how|what)|Let(?:'s|’s| us) (?:dive|break (?:it|this) down|unpack))\b
+\bThat(?:'s|’s| is) the [\w-]+\.\s+Here(?:'s|’s| is)\b
+\bthat(?:'s|’s| is) (?:exactly )?what [^.!?]{1,40}\b(?:is|are) for\b
+```
+
+The third pattern is case-sensitive. One self-answered question can be fine. Flag two or more
+per piece, or the same signpost in three or more pieces of a batch. Ask a question only when
+it is still open; otherwise give the answer. Basis: editorial.
+
+## Structural
+
+These are mostly batch checks. A single piece can't be a template on its own.
+
+### T1. One shape for every piece
+
+For each piece, record paragraph count, word or character count, and the word count of each
+paragraph. Compute the coefficient of variation of paragraph length (standard deviation
+divided by mean).
+
+Report each piece's figures as a description, not a flag. Don't flag a piece for where it
+falls in a ranking: ranking any batch produces a bottom group, even when the writing is
+varied and suits its purpose.
+
+Flag only concentration across comparable pieces, and report it with its denominator, for
+example "7 of 10 pieces have four paragraphs". The starting defaults are uncalibrated:
+
+- one paragraph count covers seven or more of ten comparable pieces (70% of the batch);
+- seven or more of ten comparable pieces fall within 10% of the batch's median word count
+  while their subjects clearly differ. Flag this one for manual review.
+
+With fewer than five comparable pieces, report the figures and draw no batch conclusion. Let
+the argument decide the shape. Don't pick a new target shape. Basis: study for reduced
+stylistic variation in the genres tested (HV-06); the thresholds are editorial.
+
+### T2. Flat sentence length
+
+Compute the standard deviation of sentence length in words for each piece and report it as a
+description, not a flag. Exclude short reference entries and step lists, which are
+legitimately even. Flag the batch only when flat pieces concentrate. The uncalibrated
+starting default is a standard deviation under five words in seven or more of ten comparable
+pieces, reported with its denominator. There is no correct number. With fewer than five
+comparable pieces, draw no batch conclusion. Basis: as T1.
+
+### T3. One opening for every piece
+
+Classify each piece's first sentence, in this order:
+
+- number: the first token contains a digit
+- question: the sentence ends with `?`
+- identifier: contains a backtick span, a `Name()` call or a CamelCase word
+- direct address: starts with `You`, `Your` or `If you`, or with an imperative verb (manual)
+- abstract assertion: matches `^(?:A|An|The)\s+\w+(?:\s+\w+)?\s+(?:is|are|can|needs?|has|should|will)\b`
+- event or problem: anything else, confirmed by a reader
+
+Flag the batch when one class covers six or more of ten pieces, or when three or more pieces
+share their first two words. Open with a documented event, task or decision. A concrete
+opening doesn't mean an invented scene: Dudycz's article on fixing event-sourcing bugs opens
+on bad data every system accumulates, then a support incident (DA-03). Don't overcorrect
+into density either. In historical headline experiments, added concreteness helped headlines
+that were too vague and hurt ones that were already very concrete (HC-03), and linguistic
+features predicted the winning variant only 54% of the time (HC-05). Basis: editorial, with
+the headline studies as a limit on over-tuning.
+
+### T4. Bold-label bullets, takeaway blocks and emoji bullets
+
+```text
+^\s*(?:[-*+]|\d+\.)\s+\*\*[^*\n]+\*\*
+^\s*(?:#{1,6}\s*|\*\*)?(?:Key takeaways?|Takeaways?|TL;?DR|The bottom line|Why (?:this|it) matters)\b
+^\s*(?:[-*+]\s+)?\p{Extended_Pictographic}
+```
+
+Use `m` for all three and `u` for the third. Flag three or more consecutive bold-label
+bullets where each label is followed by a sentence of prose, any takeaway block in a piece
+under 800 words, and emoji-prefixed bullets as a separate count. Reference lists, glossaries
+and option tables are legitimate uses and are exempt. Use a list for a real procedure or
+comparison and paragraphs for an argument. Basis: editorial.
+
+### T5. Summary or maxim close
+
+Apply to the first line of the last paragraph:
+
+```text
+^(?:In (?:summary|conclusion|short)|To (?:sum up|summarize|conclude)|Overall|Ultimately|All in all|The (?:lesson|takeaway|bottom line)(?: here)? is)\b
+```
+
+Also flag a last sentence of twelve words or fewer that has no digit, identifier or code and
+shares two or more content words with the first paragraph: that is usually a restatement
+dressed as a maxim. Flag the batch when summary or maxim closes appear in five or more of ten.
+End on a result, a real limit, the next step, or a question the author wants answered.
+Basis: editorial.
+
+### T6. Labeled caveat at the end
+
+A limitation parked in its own closing paragraph: "One honest limit:".
+
+```text
+^(?:(?:One|A|The)\s+(?:(?:honest|important|real|big|key)\s+)?(?:limit|limitation|caveat|note|warning)|Caveat|Limitation|Note)s?\b[^:\n]{0,20}:
+```
+
+Flag any match in the last paragraph, and the batch when three or more pieces end this way.
+Move the limit into the sentence it qualifies: "On macOS and Linux, the installer…". Never
+delete it to improve the ending. Basis: editorial.
+
+## Tonal
+
+### N1. Generic optimism and absolute certainty
+
+```text
+\b(?:always|never|guaranteed?|every time|zero[- ](?:risk|downtime|config(?:uration)?)|thrilled|excited to|delighted|incredible|amazing|world[- ]class|next[- ]level|cutting[- ]edge|state[- ]of[- ]the[- ]art)\b|100%
+```
+
+Flag clusters of two or more in a paragraph and every absolute that isn't backed by a source.
+"Never" and "always" are often real limits; check N5 before touching one. State what was
+observed, what is uncertain, and which cases aren't supported. Basis: editorial.
+
+### N2. Synthetic first-person experience
+
+Confessions, lessons learned the hard way, and customer anecdotes.
+
+```text
+\b(?:I|we)\s+(?:was|were)\s+wrong\b|\b(?:I|we)\s+learned\s+(?:this\s+)?the\s+hard\s+way\b|\b(?:a|one of our)\s+(?:customers?|clients?|users?)\s+(?:told|asked|said|called|emailed)\b|\b(?:last|this)\s+(?:week|month|year),?\s+(?:I|we)\b
+```
+
+This isn't a language judgment. Every match goes to the named author, who confirms it
+happened and points to something inspectable: the issue, the commit, the message. If they
+can't, the passage is removed or labeled as a hypothetical. Basis: LinkedIn's stated policy
+that posts should represent the member's own voice and perspectives (HV-18). The rest is an
+ethical rule and makes no evidence claim.
+
+### N3. Engagement bait
+
+```text
+\b(?:comment\s+["“]?\w+["”]?\s+(?:below|if)|agree\?|repost\s+if|tag\s+(?:someone|a\s+(?:friend|colleague))|who\s+else\b[^?\n]*\?|let\s+me\s+know\s+in\s+the\s+comments|thoughts\?)
+```
+
+Flag every match and review the intent. A closing question can be genuine; a request to
+react isn't. Invite a specific counterexample, or ask something the author needs answered.
+Basis: editorial.
+
+### N4. Review language in the copy and writer-protecting hedges
+
+Evidence, verification and approval wording that belongs in the review record, and stacked
+qualifiers that protect the writer without informing the reader.
+
+```text
+\b(?:verified (?:against|by|in)|as (?:approved|reviewed)|(?:the )?evidence (?:shows|suggests|indicates)|per (?:the )?review|this has been (?:checked|validated))\b
+\b(?:may|might|could|potentially|possibly|arguably|somewhat|in some cases|to some extent|it appears|it seems)\b
+```
+
+Flag every match of the first pattern. Count the second per sentence and flag two or more in
+one sentence. Move review residue to the review record. Keep a qualifier when it is true and
+would change what the reader does. Basis: editorial.
+
+### N5. Meaning lost in the edit
+
+This check compares a passage before and after rewriting. Run it on every rewrite, including
+passages nothing flagged.
+
+```text
+\b(?:not|no|never|only|unless|until|without|except|may|might|must|cannot|can['’]t|won['’]t|don['’]t|doesn['’]t|isn['’]t|aren['’]t)\b
+```
+
+Run it on the complete original of both versions, never the prose view: the identifiers and
+quoted values it compares sit in code spans and quotation marks. Extract from each version:
+identifiers (backtick spans, CamelCase words, dotted names, paths), quoted values, numbers
+(`\d[\d.,]*%?`), and the negation, limiter and modal words above. Normalize curly apostrophes
+to straight ones in the extracted words, so a change of apostrophe style alone isn't
+reported. Flag any item present before and missing after, and any number or identifier that
+appears only after. A person then reads each flagged pair for meaning; counts can't show that
+"there is no need to" has become "without needing to". Basis: editorial.
+
+Run these fixtures before relying on the check. Each pair must give the result shown.
+
+| Before | After | Expected |
+| --- | --- | --- |
+| The installer doesn't support Windows. | The installer supports Windows. | Flag: `doesn't` missing after |
+| The installer doesn’t support Windows. | The installer supports Windows. | Flag: `doesn’t` missing after |
+| The installer doesn’t support Windows. | The installer doesn't support Windows. | No flag |
+| The installer doesn’t support Windows yet. | The installer doesn’t support Windows. | No flag |
+| Retries stop after 3 attempts. | Retries stop after 5 attempts. | Flag: `3` missing, `5` only after |
+
+## Sources
+
+Only findings that passed verification are listed. The wording in the second column is the
+verified claim, narrowed where the original summary overstated it.
+
+| ID | What the source supports | Type | Source |
+| --- | --- | --- | --- |
+| HV-01 | In six experiments with 4,600 participants, people could not reliably detect AI-generated self-presentations in professional, hospitality and dating contexts. | Peer-reviewed, 2023 | [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC10089155/) |
+| HV-02 | The same study identifies first-person pronouns and contractions as flawed heuristics for judging human authorship. Applies to the studied self-presentations. | Peer-reviewed, 2023 | [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC10089155/) |
+| HV-03 | In a convenience sample, 140 instructors and 145 students picked the ChatGPT essay from a pair 70% and 60% of the time; experience, subject expertise and confidence did not reliably predict better performance. | Peer-reviewed, 2024 | [Springer](https://link.springer.com/article/10.1007/s40979-024-00158-3) |
+| HV-05 | A study of parallel human and LLM corpora found systematic lexical, grammatical and rhetorical differences, in aggregate. | Peer-reviewed, 2025 | [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/) |
+| HV-06 | In the models and genres studied, LLMs struggled to match human stylistic variation. | Peer-reviewed, 2025 | [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC11874169/) |
+| HV-09 | From excess vocabulary across more than 15 million PubMed abstracts, at least 13.5% of 2024 abstracts were estimated to be processed with LLMs. An inferred lower bound, not a count, and not about social media. | Corpus study, 2025 | [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC12219543/) |
+| HV-10 | The same study reports strong excess 2024 usage of "delves", "underscores" and "showcasing". An aggregate signal, not an individual verdict. | Corpus study, 2025 | [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC12219543/) |
+| HV-11 | In an experiment with 680 US participants, some AI assistance increased engagement and content volume while lowering perceived quality and authenticity of discussion. | Preprint, not peer reviewed, 2025 | [arXiv](https://arxiv.org/abs/2506.14295) |
+| HV-12 | A review of 47 journalism studies found no consistent trust penalty from AI provenance cues; effects depended on topic, baseline trust, source cues and signaled human oversight. Not about developer posts. | Systematic review, 2026 | [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC13183635/) |
+| HV-18 | LinkedIn: "It's ok to use AI to help you write, but your posts and comments need to represent your voice and your perspectives." | Platform statement, 2026 | [LinkedIn](https://news.linkedin.com/2026/keeping-conversations-real-on-linkedin) |
+| HV-19 | LinkedIn: content that appears AI-generated and lacks clear perspective is less likely to be widely distributed beyond the author's immediate network. LinkedIn's own claim, not independently measured. | Platform statement, 2026 | [LinkedIn](https://news.linkedin.com/2026/keeping-conversations-real-on-linkedin) |
+| HV-21 | A detector vendor classified 4,061 of 5,000 sampled public LinkedIn posts of 100+ words as "Likely AI". A vendor classification, not verified authorship or a platform-wide rate. | Vendor, observational, 2026 | [Originality.ai](https://originality.ai/blog/ai-content-published-linkedin) |
+| HV-22 | The vendor limits that estimate to posts found through its topic-and-date searches. | Vendor, 2026 | [Originality.ai](https://originality.ai/blog/ai-content-published-linkedin) |
+| HV-23 | Wikipedia's editors' guide says its indicators are context-dependent, descriptive and not proof that a text was AI-generated. | Editors' guide, living | [Wikipedia](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) |
+| HV-25 | The same guide lists inflated claims of broad significance, including "evolving landscape", as possible indicators. | Editors' guide, living | [Wikipedia](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing) |
+| HC-03 | Across 8,977 historical Upworthy headline experiments, more concreteness raised click-through for too-vague headlines and lowered it for too-concrete ones. Clicks, not reading or trust. | Registered report, 2025 | [Nature](https://www.nature.com/articles/s41598-024-81575-9) |
+| HC-05 | Linguistic features predicted the higher-CTR Upworthy variant 54% of the time across all pairs. | Registered report, 2023 | [PLOS ONE](https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0281682) |
+| HC-16 | Hacker News guidelines: "Don't post generated text or AI-edited text. HN is for conversation between humans." The line sits in the comment guidelines. | Platform rule, living | [HN guidelines](https://news.ycombinator.com/newsguidelines.html) |
+| DP-05 | PostHog's voice guidance favors concrete nouns, directness and honesty about limitations over polish. One company's judgment. | Practitioner guide, living | [PostHog handbook](https://posthog.com/handbook/brand/tone) |
+| DA-03 | Dudycz's article on fixing event-sourcing bugs opens with bad data from integrations, users and shipped changes before describing a support incident. | Case example, 2026 | [event-driven.io](https://event-driven.io/en/fixing-bugs-in-event-sourcing-is-hard/) |
+
+Platform rules and living guides change. Re-read them before citing them in published work.

package/package/corpus/skills/cratis-writing-voice-and-cadence/verification.json

@@ -6,12 +6,24 @@
     { "kind": "skill-contains", "value": "Never trade a claim for a smoother sentence" },
     { "kind": "skill-contains", "value": "The fix for a template is not a better template" },
     { "kind": "skill-contains", "value": "every hedging word" },
-    { "kind": "skill-contains", "value": "First person is not a licence to invent" },
+    { "kind": "skill-contains", "value": "First person is not a license to invent" },
     { "kind": "skill-contains", "value": "Structural sameness is the more damaging half" },
     { "kind": "skill-contains", "value": "Rewrite a tranche, read it, and only then continue" },
     { "kind": "skill-contains", "value": "Punchline fragment" },
     { "kind": "skill-contains", "value": "Labeled caveat" },
     { "kind": "skill-contains", "value": "Hedging that protects the writer" },
-    { "kind": "skill-contains", "value": "A company account is a team talking" }
+    { "kind": "skill-contains", "value": "A company account is a team talking" },
+    { "kind": "skill-contains", "value": "This is an editing aid, not a detector" },
+    { "kind": "skill-contains", "value": "It never means \"AI wrote this\"." },
+    { "kind": "skill-contains", "value": "Never use a flag, a count or a detector score to accuse a person" },
+    { "kind": "skill-contains", "value": "Don't push them toward an arbitrary house style" },
+    { "kind": "skill-contains", "value": "Contractions, \"I\", humor, short sentences and typos prove nothing about who wrote a text" },
+    { "kind": "skill-contains", "value": "Flag concentration in the batch, never one outlier" },
+    { "kind": "skill-contains", "value": "Fix with substance" },
+    { "kind": "skill-contains", "value": "Never fabricate one to make copy read better" },
+    { "kind": "skill-contains", "value": "Copy in the first person needs the named author's own input" },
+    { "kind": "skill-contains", "value": "lacks clear perspective" },
+    { "kind": "skill-contains", "value": "Cratis applies it, as its own stricter policy" },
+    { "kind": "skill-contains", "value": "references/ai-tells.md" }
   ]
 }

package/package/profile-catalog.json

@@ -563,6 +563,8 @@
     {
       "id": "cratis/content",
       "availableTargets": [
+        "cratis-content-length",
+        "cratis-developer-marketing",
         "cratis-release-notes",
         "cratis-social-feed-post",
         "cratis-technical-examples",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cratis/pi",
-  "version": "6.3.2",
+  "version": "6.5.0",
   "description": "Configuration-aware Cratis AI integration for Pi",
   "type": "module",
   "files": [