Writing Standards Skill for MCP Agent

Writing Standards

Production-grade content creation, editing, and quality assurance for marketing websites. Provides copywriting frameworks, structured editing methodology, content strategy planning, social content creation, and AI writing pattern detection.

This pack is brand-agnostic. It works with any website or product. Brand-specific voice, tone, and constraints are provided by the consumer at runtime via a brand-context.md file.

Brand Context Protocol

Before any content task, the consumer must provide a brand-context.md containing:

• Site identity — name, one-liner, target audience
• Voice rules — tone, personality, formality level (casual / professional / formal)
• Words to use — approved terminology, brand language
• Words to avoid — banned terms, competitor names, off-brand phrases
• Proof points — key metrics, customer quotes, case study data
• Content types — what the site publishes (blog, landing pages, docs, social, email)
• Visual style — design constraints that affect copy (dark theme, minimal, data-heavy)

If no brand context is provided, ask for one before writing. Without it, copy will be generic and miss the brand's voice.

Content Taxonomy

Type	Purpose	Key Constraint
Landing page	Convert visitors → action	One CTA per page, benefits over features
Blog post	Educate, build authority, drive organic traffic	Searchable: keyword-driven, structured for SEO
Case study	Prove value with real outcomes	Data-backed, customer-approved quotes
Documentation	Help users succeed	Scannable, task-oriented, no marketing tone
FAQ	Handle objections, reduce support load	Real questions from real users, concise answers
Email	Nurture leads, retain users	Subject line is 80% of the work, one action per email
Social post	Build audience, drive traffic, establish voice	Platform-specific format and tone
Ad copy	Capture attention, drive clicks	Character limits, clear value prop, strong CTA
Comparison page	Win competitive evaluations	Honest, specific, never disparage competitors

Module Catalog

Module	Purpose	Key Frameworks
copywriting	Write new marketing copy	Headline formulas, CTA patterns, page-type guidance, voice handling
copy-editing	Review and polish existing copy	Nine Sweeps structured editing, quick-pass checks
content-strategy	Plan what to write and why	Searchable vs Shareable, content pillars, buyer-stage keywords
social-content	Create platform-specific social posts	Hook formulas, repurposing system, content calendar
writing-quality	Detect and eliminate AI writing patterns	29-pattern detection, 5-dimension scoring, content-type-aware thresholds

Loading Order

1. This file (_skill.md) — pack overview, brand context protocol
2. Consumer's brand-context.md — site-specific voice and constraints
3. Role file (_roles/writer.md or _roles/editor.md) — task lifecycle
4. Module _skill.md — domain-specific methodology
5. Module references/ — detailed frameworks, loaded on demand

Cross-Module Dependencies

• writing-quality is referenced by both copy-editing (as the Anti-AI-Slop sweep) and by the writer role (as the final quality check before submission)
• copywriting and copy-editing are complementary: use copywriting to draft, copy-editing to review
• content-strategy informs what to write; copywriting and social-content handle the actual writing

Workflow Catalog

Workflow	Module	Purpose
writing_quality_check	writing-quality	Score content on 5 dimensions, detect AI patterns, content-type-aware pass/fail verdict

Provenance & Confidence

Overall: ~85% GEN / 15% RES | Pack total: ~3,700 lines across 22 files

Only writing-quality has external source attribution. All other modules are generated domain expertise without cited research.

Confidence rule: 0.49 max. GEN-only modules set to 0.24 (at least 2x lower than RES-backed modules).

Calibration status: Scoring calibrated against 5 gold-standard landing pages (Stripe, Linear, Basecamp, Vercel, Notion). Content-type-aware thresholds: 28/50 for landing pages, 35/50 for prose.

GEN / RES per Module

                         0%       25%       50%       75%      100%
                         │         │         │         │         │
writing-quality (0.49)   █████████████▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  33% GEN / 67% RES
copywriting     (0.24)   ████████████████████████████████████████  100% GEN
copy-editing    (0.24)   ████████████████████████████████████████  100% GEN
content-strategy(0.24)   ████████████████████████████████████████  100% GEN
social-content  (0.24)   ████████████████████████████████████████  100% GEN
roles           (0.24)   ████████████████████████████████████████  100% GEN
                         │         │         │         │         │
                         ██ GEN (no source)    ▓▓ RES (cited)

Source Attribution

Source	License	Contributes to	Lines
blader/humanizer	MIT	writing-quality: 29 AI patterns from Wikipedia WikiProject AI Cleanup	217
hardikpandya/stop-slop	MIT	writing-quality: 5-dimension scoring, banned phrases/structures, 8 core rules	373
(none)	—	copywriting, copy-editing, content-strategy, social-content, roles	~3,100

Calibration Results

Tested against 5 recognized landing pages (v3, after content-type fixes):

Site	Score	Verdict	Key finding
stripe.com	27	REVISE (-1)	Strong proof but generic promo language
linear.app	22	REVISE (-6)	Unsourced claims, repetitive positioning
basecamp.com	26	REVISE (-2)	Strongest voice, em-dash overuse
vercel.com	25	REVISE (-3)	Unattributed metrics
notion.com	28	PASS	Most named proof points
aictpo.com	28	PASS	Direct, specific, trust is the weak dimension

Later Improvements

Calibration Gaps (from v3 testing)

Scoring threshold: 3 of 5 gold-standard sites fail by 1-3 points. Consider lowering landing-page threshold to 25 or adding a "borderline pass" band at 25-27.

Trust dimension still too strict: Even with the "specific and verifiable" adjustment, the tool flags named metrics (e.g., Stripe's "US$1.9tn") as needing inline sources. Consider a further exception: metrics attributed to the site's own company/product are inherently first-party and need no external citation.

Negative calibration missing: No test against known-bad copy to confirm the tool catches it. Add 2-3 generic AI SaaS pages as negative test cases (expected score <22).

Research Targets (GEN → RES)

High priority (100% GEN, high impact):

• copy-editing: Nine Sweeps methodology — cite editing frameworks (Ann Handley Everybody Writes, Zinsser On Writing Well, AP Stylebook)
• copywriting: headline formulas — cite Copyhackers (Joanna Wiebe), published conversion research, Unbounce landing page studies
• content-strategy: prioritization scoring — cite HubSpot Content Strategy, Animalz research, Orbit Media annual blogging survey
• writing-quality: scoring.md rubric thresholds — anchor to published readability research (Flesch-Kincaid, Hemingway, Contently scoring)

Medium priority:

• copywriting: CTA copy guidelines — cite ConversionXL button copy studies, Unbounce CTA research
• content-strategy: topic cluster methodology — cite HubSpot pillar/cluster model documentation
• social-content: hook formulas — cite LinkedIn algorithm research (Richard van der Blom studies), Justin Welsh methodology
• social-content: platform algorithm behavior — cite platform documentation, Hootsuite/Buffer research reports

Low priority (experiential, hard to cite):

• All Tips and Gotchas sections — domain expertise by nature
• Roles (writer, editor) — process-oriented, not fact-oriented
• Plain English Alternatives reference — widely known substitutions, no single authoritative source