Writing Quality
The standing quality gate for all human-facing prose in the system. It catches the patterns that make writing read as machine-generated (inflated abstractions, promotional puffery, filler, formulaic structure, the AI vocabulary) and either flags them (DETECT) or removes them (EDIT). Grounded in the public "signs of AI writing" body of knowledge (e.g. Wikipedia's guide), not any private banned-word list.
When to use
- Automatically, as the last step inside
resume-tailor,cover-letter,networking-and-referrals,interview-followup, andlinkedin-presence. Nothing ships before it passes here. - On demand: "does this sound like AI?", "de-slop this", "make this sound like a person wrote it", "clean up this email".
- As stage 0 inside
artifact-eval(the eval runs DETECT before scoring).
Inputs
- Reads: the target text (pasted, or a file path in
workspace/). - Needs from user: which mode, DETECT (report only) or EDIT (rewrite). Default to DETECT when merely asked "does this sound AI?", and default to EDIT when asked to "fix/clean/de-slop." When another skill calls this, it specifies the mode.
Modes
- DETECT. Find and name each tell, quote the offending span, say which category it is, and give the fix. Do not rewrite. Output a flagged list + a verdict (clean / minor / heavy). Use inside evals and when the user wants to learn the pattern.
- EDIT. Rewrite to remove every tell while preserving meaning and every fact, name, number, and claim. Never invent detail to replace slop. Cut it or ask. Return the cleaned text plus a short changelog of what was fixed and why.
The tells: spot it / fix it
Work through all of these. Each entry: what it is → how to spot → the fix.
- Inflated symbolism / grand abstraction. Small things framed as monumental ("stands as a testament to," "a pivotal moment in the ever-evolving landscape"). Spot: cosmic weight on an ordinary point. Fix: state the concrete thing plainly and delete the grandeur.
- Promotional puffery. Unearned superlatives and hype ("rich cultural heritage," "world-class," "cutting-edge," "seamless experience"). Spot: adjectives that praise instead of inform. Fix: replace with a specific, verifiable fact, or cut.
- Vague attribution / weasel words. "industry experts agree," "studies show," "it is widely regarded," with no source. Spot: an authority claim with no named source. Fix: name the source and link/cite it, or remove the claim.
- Em-dash overuse. Em dashes as an all-purpose connector, several per paragraph. Spot: more than ~1 per paragraph, or dashes doing the job of commas/periods. Fix: restructure into clean sentences, using commas, colons, or full stops.
- Rule of three, everywhere. Reflexive triples ("fast, reliable, and scalable", or "we plan, build, and deliver"). Spot: triads used as a default rhythm. Fix: break the pattern. Use one strong item, or two, or a different count. Vary sentence shape.
- AI vocabulary. The tell-tale lexicon: delve, leverage, robust, underscore, testament, navigate, landscape, realm, tapestry, seamless, elevate, foster, harness, unlock, spearhead, streamline, pivotal, multifaceted, holistic, boasts, showcase, embark, ever-evolving. Spot: any of these. Fix: swap for a plain-English verb/noun a person would actually say ("use" not "leverage," "strong" not "robust," "look into" not "delve into").
- Negative parallelism. "it's not just X, it's Y", or "this isn't about A, it's about B." Spot: the setup-reversal contrast used for false profundity. Fix: cut the negation and make the positive claim directly.
- Filler / hedging preambles. "It's worth noting that," "It's important to note," "Needless to say," "At the end of the day," "In today's fast-paced world." Spot: throat-clearing that adds no information. Fix: delete it and start with the point.
- Over-caveating. Stacked qualifiers ("may potentially, in some cases, arguably, to a certain extent"). Spot: three hedges guarding one claim. Fix: make the claim, or drop it. Keep at most one honest qualifier.
- Empty CTAs. "Let me know your thoughts," "Feel free to reach out," "Don't hesitate to contact me," "Looking forward to hearing from you" with no specific next step. Spot: a closer that asks for nothing concrete. Fix: replace with a specific, low-friction ask (a question, a proposed time, a single action).
- Generic openings. "I hope this email finds you well," "I'm excited to apply for," "In this essay, I will," "As an AI language model." Spot: an opener that could head any message to anyone. Fix: open with something only this message could say, such as a specific hook, a shared detail, the actual point.
- Superficial -ing summaries / essay-outro moralizing. Closers like "Ultimately, this highlights the importance of…," "By embracing X, we can…," "In conclusion, it's clear that…." Spot: a tidy moral that restates without adding. Fix: end on the last real point, a concrete next step, or cut the paragraph.
- Section-header parallelism / list bloat. Every bullet padded to the same shape, headers all "Verb-ing the Noun." Spot: uniform, decorative structure. Fix: let content set the length, and cut bullets that only exist for symmetry.
- Repetitive sentence openers / uniform rhythm. Many sentences starting the same way, all the same length. Spot: metronomic cadence. Fix: vary opener and length, then read it aloud.
Workflow
- Read the whole piece first. Some tells (rhythm, triads, repeated openers) are only visible across the full text.
- Pass against all 14 categories. Mark every hit with its span and category.
- Mode branch:
- DETECT → output the flagged list (quote · category · fix) and a verdict (clean / minor / heavy) with a count.
- EDIT → rewrite, removing every tell, preserving all facts, names, and numbers, and the author's intent. Match the target register (résumé bullet ≠ LinkedIn post ≠ email). Don't over-correct into terse fragments. Keep it natural.
- Preserve, don't invent. If cutting slop leaves a gap that needs a real fact you don't have, flag it for the user. Never fabricate to fill it.
- Re-scan the rewrite. EDIT output must itself pass DETECT. Don't trade one tell for another (a classic failure: replacing "leverage" with "harness"). Confirm clean before returning.
- Return the result + a short changelog (what changed, which category each fix addressed).
Output
- Writes to: nothing by default. It returns to chat, or to the calling skill's own output file. Only writes a file if the user explicitly asks.
- Format:
- DETECT → flagged list + verdict + count.
- EDIT → cleaned text + changelog. If facts were missing, an explicit "needs from you" list.
Rules & quality bar
- Meaning and facts are sacrosanct. EDIT never changes a claim, number, name, date, or the author's position. It only changes how it's said.
- Never fabricate to replace slop. Cut or ask. Don't paper over a hole with invented specifics.
- General and publicly grounded. The tells come from the public "signs of AI writing" knowledge base. Do not encode any individual's private brand banned-word list here. Keep it reusable for anyone.
- Register-aware. A cover letter can be warmer than a résumé bullet, so don't flatten voice into robotic minimalism. The goal is human, not sterile.
- The rewrite must pass its own gate. Re-scan before returning.
- Judgment over zealotry. One em-dash or one honest "note that" isn't slop. Flag patterns and density, not every single instance.
Related skills
- Called by:
resume-tailor,cover-letter,networking-and-referrals,interview-followup,linkedin-presence, andoffer-negotiation(on the counter-offer email). All of them run this before finalizing. - Embedded in:
artifact-evalas stage 0 (DETECT feeds the no-slop dimension and hard-gate check).
