@ericrisco/rsc 1.1.0 → 1.1.2

package/skills/unslop/SKILL.md

@@ -0,0 +1,114 @@
+---
+name: unslop
+description: "Use when a written text must be audited against a named catalogue of AI tells before it ships: puffery, `not just X but Y`, the rule of three, em dashes, filler, hedging, abstract metaphor nouns. Names each hit and fixes it. NOT the register rewrite in the user's own voice (that is `bro`), NOT a reusable voice guide (that is `brand-voice`)."
+tags: [unslop, ai-tells, editing, audit, slop, publish-check, revisar-texto]
+recommends: [bro, brand-voice, technical-writing]
+profiles: [minimal, core, full]
+origin: risco
+---
+
+# unslop: the named-tell audit a text passes before it ships
+
+`bro` decides how a text should *sound*. This decides whether a text still carries the marks of a
+machine, tell by named tell, on the words that are actually there. Run it on anything about to leave
+the session: a README, a post, an email, a landing page, a commit body.
+
+## The pass
+
+1. **Scan against the tables below.** Every hit gets located, not sensed. If you cannot name the
+   tell, it is not a finding.
+2. **Rewrite in place.** Preserve meaning, facts, numbers, links, commitments and the intended tone.
+3. **Restore the voice** (next section). A text stripped of tells and nothing else reads as sterile,
+   which is its own tell.
+4. **Self-audit once more.** Ask: what still makes this obviously machine written? Fix that too.
+5. **Return the text.** On request, return the list of named hits alongside it.
+
+## Removing tells is half the job
+
+Voiceless writing is as obvious as slopped writing. After the cuts, put something back:
+
+- **Have an opinion.** React to the facts instead of listing balanced pros and cons.
+- **Vary the rhythm.** Short sentence. Then a longer one that takes its time and earns the length.
+- **Admit complexity.** "Impressive and slightly unsettling" beats "impressive".
+- **Use "I" when it fits.** First person is not unprofessional.
+- **Allow some mess.** Perfect symmetry looks manufactured.
+- **Be specific.** Not "this is concerning" but "it churns for six hours at 3am and nobody is watching".
+
+## Content tells
+
+| Tell | Fix |
+| --- | --- |
+| Puffery: "pivotal moment", "testament to", "evolving landscape", "indelible mark" | State what happened |
+| Name dropping outlets with no context | Pick one, quote what it said |
+| Decorative `-ing` clauses: "highlighting...", "ensuring...", "showcasing..." | Delete, or expand into a real claim with a source |
+| Promotional adjectives: "nestled", "vibrant", "breathtaking", "renowned", "must-visit" | Neutral description |
+| Vague attribution: "experts believe", "reports suggest", "critics argue" | Name the source or cut the claim |
+| The formulaic challenge arc: "despite challenges, X continues to thrive" | Specific facts instead of the arc |
+
+## Language tells
+
+| Tell | Fix |
+| --- | --- |
+| AI vocabulary: additionally, crucial, delve, enhance, foster, garner, interplay, intricate, landscape, pivotal, showcase, tapestry, testament, underscore, vibrant | The plain word |
+| Fancy ways to say "is": "serves as", "stands as", "boasts", "features" | "is", "has" |
+| "Not just X, but Y" | State the point directly |
+| Rule of three: forcing every list into three items | The number the facts have |
+| Synonym cycling: protagonist, main character, central figure, hero in one paragraph | Pick one and repeat it |
+| False ranges: "from X to Y" where X and Y share no scale | List the items |
+| Abstract metaphor nouns: substrate, wedge, vector, locus, nexus, primitive, surface, bedrock, scaffolding, paradigm, flywheel, north star, gold-plating, ratchet, evacuate | The concrete word: base, add, way, more than the job needs, move out |
+
+## Style and punctuation tells
+
+| Tell | Fix |
+| --- | --- |
+| Em dashes, anywhere | End the sentence, or use a comma. Swapping in parentheses trades one tell for another |
+| Colons as mid-sentence connectors | Let the point stand as its own sentence. Colons are for lists and examples |
+| Bold on every proper noun and acronym | Bold what the reader must not miss, nothing else |
+| Inline-header lists whose bold label restates the line: "**Performance:** performance improved" | Prose. A bold lead-in that ends in a period and is followed by genuinely new detail is fine |
+| Title Case Headings | Sentence case |
+| Decorative emoji in headings and bullets | Remove |
+| Curly quotes | Straight quotes |
+
+## Filler, hedging and plain speech
+
+| Tell | Fix |
+| --- | --- |
+| "In order to", "due to the fact that", "it is important to note that" | "To", "because", delete |
+| Stacked hedges: "could potentially possibly be argued that it might" | "may", or the claim itself |
+| Generic conclusion: "the future looks bright" | A specific plan, date or number |
+| Chatbot phrases: "I hope this helps", "let me know if", "certainly", "great question" | Remove |
+| Cutoff disclaimers: "while specific details are limited" | Find the detail or drop the sentence |
+| Sycophancy: "you're absolutely right" | Answer directly |
+| Feeling instead of mechanism: "the database stays close at hand" | The mechanism or the number: "`.toSQL()` returns the exact string sent to the database" |
+| A sentence that would fit unchanged in another project's docs | Cut it. It says nothing about this one |
+| Dense sentences the reader has to reparse | One idea per sentence |
+| Passive with a hidden actor: "queries are validated" | Name the actor: "the compiler validates queries" |
+| Adverbs propping up weak verbs: "runs quickly", "significantly improves" | A stronger verb or the measured number |
+| Fancy synonyms: utilize, leverage, facilitate, numerous, in the event that | use, use, help, many, if |
+
+## Anti-patterns
+
+| Anti-pattern | Why it fails | Do this instead |
+| --- | --- | --- |
+| Report the tells and leave the text unfixed | The user asked for a clean text, not a diagnosis | Fix in place, list on request |
+| Cut every tell and stop | Sterile prose is the other AI tell | Restore opinion, rhythm and specifics |
+| Rewrite the meaning while cleaning the style | The edit silently changes the claim | Facts, numbers, links and commitments stay exact |
+| Strip a term because it sounds technical | Precision lost to style is a defect | Keep the term, explain it once |
+| Touch code, commands, URLs, quotes or regulated wording | Those are not prose | Leave them byte for byte |
+| Claim the result will pass an AI detector | Nobody can promise that | Promise named tells removed, nothing more |
+| Flatten a voice the author chose on purpose | Some authors do write long, warm or formal | Match the intended tone, cut only the machine marks |
+
+## Output boundary
+
+Return the cleaned text as an artifact, with no preface. Keep harness commentary, the findings list
+and any compass block clearly outside it, so nothing that is not the text can be copied into an
+email or a document by accident.
+
+## Where this ends
+
+- **`bro`** owns the register pass: make it sound like a person, in the user's language, fitted to the
+  channel and the relationship. A bare "bro" after an answer is always `bro`. This skill is the named
+  audit of a text that already exists, before it goes out.
+- **`brand-voice`** owns the standing voice system across writers and channels.
+- **`technical-writing`** owns structure: what belongs in a tutorial, a reference, a README.
+- **`landing-copy`** owns conversion architecture. Clean its words only after it has chosen them.

package/skills/unslop/evals/README.md

@@ -0,0 +1,25 @@
+# Evals — unslop
+
+The claim under test is the boundary as much as the capability. `bro` sits directly next to this
+skill in the base install, and a description that blurs them costs every user two owners for one
+turn. Half the negatives here exist to prove the split holds.
+
+## Triggering
+
+Load only `unslop` and run every prompt in 3–5 fresh sessions. A positive passes when `unslop` loads
+in a majority of trials; a negative passes when it stays quiet and the declared `route_to` skill is
+the real owner. Require at least 90% accuracy, and treat any failure on the two `bro` negatives as
+blocking rather than statistical: those are the cases the split was designed for.
+
+Run the mirror direction too. `bro`'s own suite carries a negative that routes here; both must pass
+in the same round, or the boundary only exists on one side.
+
+## Capability
+
+Run the scenario without the skill and with it, same post and one grader, three trials. The
+with-skill condition must satisfy at least 80% of the rubric and beat the baseline by at least 20
+percentage points.
+
+Two points discriminate hardest: **naming** the tells by category (a baseline rewrites silently) and
+**preserving** the code block, the price and the URLs (a baseline edits them while tidying). Grade
+both strictly.

package/skills/unslop/evals/cases.yaml

@@ -0,0 +1,60 @@
+skill: unslop
+
+# `unslop` is the named-tell audit of a text that already exists, run before it
+# ships. `bro` is the register pass: sound like a person, in the user's language,
+# fitted to the channel. A bare "bro" is always `bro`; "check this for AI tells
+# before I publish it" is always this skill.
+should_trigger:
+  - prompt: "Revisa este README antes de publicarlo y quítale todo lo que huela a IA."
+    why: "A pre-publication audit of an existing text is the exact contract, in the user's own words."
+
+  - prompt: "Unslop this."
+    why: "The bare skill name over a text in context is the shortest form of the request."
+
+  - prompt: "Can you go through this post and tell me which bits read as AI-written, then fix them?"
+    why: "Named findings plus a fix is what distinguishes this from a general tone rewrite."
+
+  - prompt: "Este texto está lleno de guiones largos y de 'no solo X, sino Y'. Límpialo."
+    why: "The user names two catalogued tells outright; the fix is the enumerated pass."
+
+  - prompt: "I keep getting told my docs sound like ChatGPT wrote them. What exactly am I doing?"
+    why: "The user wants the tells named, which no other writing skill in the catalog does."
+
+  - prompt: "Before this goes to the client, strip the puffery and the hedging but keep every number exact."
+    why: "A ship gate over existing prose with a preserve-the-facts constraint."
+
+should_not_trigger:
+  - prompt: "Bro."
+    route_to: "bro"
+    why: "A bare bro after an answer is the register pass over the last message, not an audit of a supplied text."
+
+  - prompt: "Haz que este correo suene más cercano y menos formal para un cliente de confianza."
+    route_to: "bro"
+    why: "Choosing a warmer register for a specific relationship is bro's job; no tell is being audited."
+
+  - prompt: "Define how our company should sound across the blog, support replies and job ads."
+    route_to: "brand-voice"
+    why: "A standing voice system across channels is brand-voice, not a one-text audit."
+
+  - prompt: "Split this mixed README into a tutorial and an API reference with runnable examples."
+    route_to: "technical-writing"
+    why: "Document structure is technical-writing; unslop edits words, it does not decide document modes."
+
+  - prompt: "Write the hero headline, the offer and the CTA for our pricing page."
+    route_to: "landing-copy"
+    why: "Conversion architecture and net-new copy belong to landing-copy; the audit comes after."
+
+  - prompt: "Translate this announcement into Catalan and adapt the date formats."
+    route_to: "translation-l10n"
+    why: "Translation and locale adaptation are owned elsewhere, even though good prose matters there too."
+
+capability:
+  - scenario: "A 400-word product-launch blog post that opens with 'In today's rapidly evolving landscape', uses four em dashes, three bold-label bullets whose labels restate their lines, 'not just faster, but smarter', 'experts agree', two stacked hedges, and closes with 'the future looks bright'. It also contains a code block, a pricing figure of EUR 39/month and two URLs."
+    must_include:
+      - "names the tells it found by category rather than reporting a general impression"
+      - "removes every em dash without substituting parentheses or en dashes"
+      - "converts the bold-label bullets whose label restates the line into prose, and leaves any bold lead-in that adds new detail"
+      - "replaces 'experts agree' with a named source or cuts the claim"
+      - "keeps the code block, the EUR 39/month figure and both URLs byte for byte"
+      - "leaves the text with a voice: at least one specific, opinionated sentence rather than sterile neutrality"
+      - "returns the cleaned text as the artifact, with the findings list kept outside it"