name: Content Hardening description: Defensibility hardening, sourcing audit and Economist Style Guide alignment for content quality

Content Hardening, Accuracy & Style Alignment

Run this skill when the user asks for: "defensibility hardening", "harden content", "defensibility pass", "style guide alignment", "economist style", "apply style guide", "content review", "content audit", "accuracy check", "sourcing audit", "citation check", "fact check", or any combination of the above.

Why This Matters

SafeAI-Aus's mission is to enable safe, ethical and growth-oriented AI adoption through open standards, practical tools and collaborative knowledge-sharing. Our core values include:

• •Transparency & Accountability — fostering trust through openness
• •Safety & Responsibility — protecting people and organisations
• •Alignment with Standards — promoting compliance with Australian regulations

Content quality directly supports these values. Inaccurate claims, overstated capabilities or outdated references undermine trust and could lead organisations to make poor decisions. Defensible, well-sourced, clearly written content helps Australian businesses adopt AI confidently and safely.

Overview

This skill performs a comprehensive content quality review:

1. •Defensibility Hardening — Softens absolute claims to be legally and factually defensible
2. •Sourcing & Accuracy Audit — Flags claims needing citations, outdated references, predictions
3. •Economist Style Guide Alignment — Applies consistent editorial style with Australian English and plain language
4. •Reader Engagement — Makes content more engaging through direct address, question-led openings and actionable framing

Pass 1: Defensibility Hardening

Search Terms

Use Grep to scan for these absolute terms across docs/ directory:

cannot|will|ensure|guarantee|always|never|must|requires|proves|certain|definitely|impossible|perfectly|completely|absolutely|all cases|every case|no exceptions

Replacement Rules

Context-Sensitive Rules

Time-scope claims:

• •Add "at present", "today", "as of [date]", "currently" to technology capability claims
• •Example: "AI cannot do X" → "AI cannot yet do X" or "As of 2026, AI does not reliably do X"

Convert guarantees to design intent:

• •"ensures X" → "is designed to X" or "aims to X"
• •"guarantees X" → "supported by X" or "validated through regular testing"
• •Add qualifiers like "when properly configured", "under normal conditions"

Avoid implying universal capabilities:

• •Replace jargon that implies more than is available
• •"internal states" → "available runtime telemetry" (unless discussing research instrumentation)
• •"full transparency" → "transparency mechanisms" or "available transparency features"

Label aspirations as aspirations:

• •Policy goals: "we recommend...", "should be required...", "best practice is..."
• •Not: stating aspirations as facts about what's possible
• •Example: "AI systems are transparent" → "AI systems should be transparent" or "transparency is a design goal"

Exceptions (Do Not Soften)

• •Direct quotes from legislation or standards
• •Mathematical/logical certainties
• •Definitions
• •Content within ??? note disclaimer blocks
• •Explicit policy statements that are meant to be absolute

Pass 2: Sourcing & Accuracy Audit

What to Flag

1. Unsourced Statistical/Numerical Claims

Search for patterns like:

[0-9]+%|[0-9]+x|[0-9]+ times|studies show|research shows|data shows|according to|statistics|survey

Flag if:

• •Specific numbers without citation (e.g., "70% of projects fail")
• •"Research shows..." without naming the research
• •Comparative claims (e.g., "7x more likely") without source

Fix options:

• •Add citation with verifiable URL
• •Add "illustrative" disclaimer (e.g., "illustrative example based on typical market ranges")
• •Soften to general statement (e.g., "research consistently shows" → no specific number)

2. Predictions vs Scenarios

Search for:

will be|will become|will happen|will require|by 2030|by 2027|in the future|soon|eventually|inevitable

Flag if:

• •Stated as fact rather than scenario/possibility
• •No attribution to a source making the prediction

Fix options:

• •Reframe as scenario: "One possible scenario is..."
• •Add attribution: "According to [source], X may..."
• •Add uncertainty: "X is anticipated" → "X may occur" or "some analysts expect X"

3. Outdated Organisational References

High-risk items to check:

• •Government programs (grants, initiatives) — status changes frequently
• •Regulatory bodies — may be renamed, restructured or abolished
• •Legislation — may be amended, replaced or repealed
• •Policies and standards — may be updated or superseded
• •Technology products/services — may be discontinued

Australian-specific references (high priority):

• •Australian Government department names and programs
• •State/territory programs (NSW, VIC, QLD, SA, WA, TAS, ACT, NT)
• •Regulators: OAIC, ACCC, ASIC, APRA, TGA, Fair Work Commission
• •Standards: Privacy Act 1988, Australian Consumer Law, VAISS, AI6
• •Funding bodies: business.gov.au programs, state innovation funds, CSIRO
• •Industry bodies: AIIA, Tech Council of Australia

Search for:

• •Program names with dates
• •"announced", "launched", "established" + year
• •References to specific funding amounts (especially in AUD)
• •Government department names
• •State/territory government references

Fix options:

• •Verify current status via web search (prefer .gov.au sources)
• •Add "last verified: [date]" or "as of [date]"
• •Add warning note about status changes
• •Update or remove if obsolete

4. Time-Sensitive Content

Check for:

• •Dates in frontmatter (last-reviewed, review-cycle)
• •Explicit dates in content
• •References to "current", "recent", "new", "latest"
• •Regulatory timelines and effective dates

Ensure:

• •last-reviewed frontmatter is current
• •review-cycle is set (quarterly for fast-changing, annually for stable)
• •Content matches the claimed review date

Audit Report Format

For each issue, document:

Risk Levels:

• •High: Legal/reputational risk if wrong; specific numerical claims; regulatory info
• •Medium: Could mislead but lower stakes; general claims; program references
• •Low: Minor accuracy issues; stylistic concerns

Pass 3: Economist Style Guide Alignment

Australian English Spelling

Use Australian English spelling throughout:

Note: Some technical terms retain US spelling when they are proper nouns or industry-standard terms (e.g., "NIST AI Risk Management Framework", "Google Cloud").

Date Formatting

Australian standard: DD/MM/YYYY or written dates

Avoid: MM/DD/YYYY (US format) — causes confusion.

Currency

• •Australian dollars: $1,000 or A$1,000 (when clarity needed)
• •International: US$1,000, €1,000, £1,000 (prefix with currency code)
• •Format: Use comma for thousands separator

Measurements

Australia uses metric exclusively:

Australian Terminology

Government & Regulatory Terms

• •Commonwealth — Federal government
• •State/Territory — Regional government (NSW, VIC, QLD, SA, WA, TAS, ACT, NT)
• •Department — Government ministry
• •Act — Legislation (e.g., Privacy Act 1988)

Oxford Comma

Remove Oxford commas (serial commas before "and"/"or" in lists):

• •"privacy, security, and compliance" → "privacy, security and compliance"
• •"risk, governance, or assurance" → "risk, governance or assurance"

Exception: Keep the comma if removing it creates ambiguity.

Word Substitutions

Capitalisation

Use sentence case for headings (capitalise only first word and proper nouns):

• •"Key Laws That Apply Today" → "Key laws that apply today"
• •Exception: If existing headings are consistently title case, maintain consistency within document

Do not capitalise:

• •"government" (unless part of proper noun like "Australian Government")
• •"federal", "state" (unless part of proper name)
• •job titles when not before a name: "the chief executive said" not "the Chief Executive said"
• •"board" (unless part of proper name like "the Reserve Bank Board")

Do capitalise:

• •Proper nouns and names of specific acts, bodies, frameworks
• •"Privacy Act 1988", "EU AI Act", "NIST AI RMF"

Numbers

• •Spell out one to nine; use figures for 10 and above
• •Use figures for percentages: "5%" not "five per cent"
• •Use figures with units: "3km", "5MB", "$2 million"

Punctuation

• •Use single quotes for quotations (Economist style), or maintain existing convention if consistent
• •Avoid exclamation marks in body text
• •Use en-dashes (–) for ranges, em-dashes (—) sparingly for parenthetical statements

Tone & Plain Language

SafeAI-Aus content should be accessible to businesses of all sizes, including those without dedicated legal or technical teams.

Plain language principles:

• •Prefer active voice over passive
• •Be direct and concise
• •Avoid jargon where plain English works
• •Write with confidence but not arrogance
• •Explain technical terms on first use, or link to the glossary
• •Use short sentences and paragraphs
• •Prefer concrete examples over abstract descriptions

Audience check: Would a small business owner without AI expertise understand this? If not, simplify or add context.

Workflow

Quick Review (Single Page)

1. •Read the specified page
2. •Apply all three passes
3. •Make edits directly
4. •Summarise changes

Full Site Audit

1. •

   Scope: Confirm with user (all docs, specific section, or specific pages)

2. •

   Scan for absolutes:

   code

   Grep pattern="cannot|ensure|guarantee|always|never|must" path="docs/" output_mode="files_with_matches"
   

3. •

   Scan for unsourced claims:

   code

   Grep pattern="[0-9]+%|[0-9]+x|research shows|studies show" path="docs/" output_mode="content"
   

4. •

   Scan for predictions:

   code

   Grep pattern="will be|will become|by 20[2-3][0-9]|inevitable" path="docs/" output_mode="content"
   

5. •

   Check time-sensitive pages:

   • •docs/safety-standards/ (legislation, regulations)
   • •docs/business-resources/ (grants, programs)
   • •Any page with last-reviewed in frontmatter
6. •

   Generate audit report with issues categorised by risk level

7. •

   Apply fixes (with user approval for High-risk changes)

8. •

   Commit:

   code

   docs: Content hardening and style alignment
   
   Defensibility:
   - [summary]
   
   Sourcing/Accuracy:
   - [summary]
   
   Style:
   - [summary]
   
   Files updated: [count]
   
   Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
   

Example Transformations

Defensibility

Before:

This framework ensures complete compliance with all Australian privacy requirements and will never expose personal data.

After:

This framework is designed to support compliance with Australian privacy requirements and aims to minimise the risk of personal data exposure.

Sourcing

Before:

Organisations with effective change management are seven times more likely to meet AI project objectives.

After (Option A - Add citation):

Organisations with effective change management are seven times more likely to meet AI project objectives (Prosci, 2023).

After (Option B - Soften):

Research consistently shows that effective change management significantly improves project success rates.

Style

Before:

We utilize robust methodologies to leverage key insights, facilitating impactful outcomes for all stakeholders.

After:

We use sound methods to draw on important insights, helping to deliver meaningful outcomes for employees, customers and partners.

Pass 4: Reader Engagement

This pass makes content more engaging and actionable for readers, particularly business audiences. Apply selectively based on content type—framework and reference pages benefit most; narrative content (scenarios, case studies) typically needs less work.

4.1 Bold Audit

Problem: Inline bold applied to entire sentences reads as shouting and reduces impact of legitimate emphasis.

Search for: Bold patterns that span full sentences or long phrases:

\*\*[^*]{40,}\*\*

Appropriate bold uses:

• •Key terms on first use: "AI alignment is about..."
• •Structural labels: "Core question:", "Key test:"
• •Table headers and definition list headers
• •Short emphatic phrases (2-5 words): "This is critical."

Inappropriate bold uses (fix these):

• •Full sentences bolded for emphasis
• •Entire paragraphs or bullet points bolded
• •Rhetorical questions bolded

Fix: Remove bold from full sentences. If emphasis is needed, restructure as a callout, use a shorter bold phrase, or rely on sentence position (opening/closing) for emphasis.

4.2 Opening Check

Problem: Pages that open with definitions or abstractions feel academic. Question-led openings ground readers in recognisable situations.

Search for pages that open with:

• •Definition patterns: "[Term] is...", "[Term] refers to...", "[Term] encompasses..."
• •Abstract nouns as subjects: "Containment means...", "Governance includes..."

Fix: Add a grounding question before the definition:

Exception: Glossary/reference pages where users expect definitions first.

4.3 Callout Conversion

Problem: Generic "Australian context" or "Context" callouts don't help readers find relevance. Reader-directed framing is more actionable.

Search for:

Australian context|**Context:**|**Note:**

Fix: Convert to "What this means for you" with role-specific guidance where appropriate:

Role-specific phrases to include:

• •"If you're in government procurement..."
• •"For business..."
• •"If you're deploying AI in public services..."
• •"For policy makers..."
• •"If your organisation operates across borders..."

4.4 Direct Address

Problem: Policy writing defaults to third person ("organisations should...") when second person ("you should...") is more engaging.

Search for third-person patterns in action-oriented sections:

organisations should|organisations must|organisations need to|businesses should|companies should

Fix: Convert to second person where appropriate:

Where to apply:

• •Action-oriented sections ("What to do", "Implementation", "How to use")
• •Questions sections
• •Recommendations and guidance

Where to preserve third person:

• •Formal policy statements
• •When discussing organisations in general (not the reader's organisation)
• •Legal or regulatory descriptions

4.5 Hedging Preservation

Critical: When converting passive to active voice or making other engagement improvements, preserve uncertainty language.

Hedging words to preserve:

• •Temporal: "becoming", "emerging", "developing", "increasingly"
• •Possibility: "may", "might", "could", "can"
• •Likelihood: "likely", "probably", "potentially", "possibly"
• •Scope: "some", "many", "often", "typically"

Bad transformation (loses hedging):

Good transformation (preserves hedging):

4.6 Content Type Awareness

Different content types need different levels of reader engagement work:

Quick test: If the page tells a story with concrete examples, it probably doesn't need engagement work. If it opens with definitions and uses third-person throughout, it's a candidate for this pass.

Maintenance Notes

• •Review this skill quarterly to update examples and patterns
• •Add new problematic patterns as they're discovered
• •Keep word substitution list aligned with current Economist style guide
• •Update Pass 4 examples as new reader engagement patterns are discovered