Accelerating content creation with CARAble AI
Built an AI-powered content assistant that accelerates ideation and structuring—reducing time to review drafts by 50%, improving initial content submission quality, and enabling creators to achieve higher output consistency.
Context
Organisations often struggle to maintain content quality and governance across multiple channels. CARA leverages AI to assist writers and reviewers in producing high-quality, compliant content efficiently.
Problem
Content creation is often inconsistent, time-consuming, and prone to errors, making governance and compliance difficult to maintain.
Content drafts often have structural, tone, accessibility, or SEO issues.
Manual checks are inconsistent and slow.
Goals
To create an AI-assisted tool that guides authors maintain consistent tone, structure, accessibility, and compliance across all content, reducing manual review time and improving content governance.
Ensures content meets style, structure, and accessibility standards uniformly.
Cuts down content review cycles and manual checks, enabling faster publishing.
Reduces risk of non-compliant or outdated content being published.
Key criteria: tone, structure, WCAG compliance, SEO optimisation.
Relies on the Content playbook document for tone, content structure and style rules.
Follows accessibility standards based on the WCAG 2.1 guidelines and incorporates audits inspired by Lighthouse Accessibility Scoring.
Implements SEO best practices guided by Google's SEO Starter Guide, along with audits for descriptive link text, meta descriptions, and font size legibility using Lighthouse SEO audits, meta description checks, and font size guidelines.
Input format: Website URL (gov.sg pages only), paste text directly, uploading a Word Document (.docx)
Output format: JSON with revised content, recommended structure, and report card.
Conceptualise CARAble pipeline
Regardless of input method, the content undergoes the same processing pipeline:
Content is fetched, extracted, or read based on the input type.
The text content is sent to the CARAble pipeline leveraging OpenAI models for analysis.
CARAble returns a detailed JSON response including content quality scores, suggestions, and rewritten content.
The app displays:
Content scorecards for structure, tone, accessibility, and SEO.
Detailed improvement suggestions.
Side-by-side comparison of original and revised content.
Options to download the improved content as HTML or Word document.
Designing the user flow and interactions
Overall structure
Happy flow
Processing interaction
Error scenarios
I developed a functional Streamlit app that accepts input in multiple formats — a gov.sg website URL, pasted text, or an uploaded Word document (.docx) — runs a CARAble review, and outputs structured JSON containing clear, actionable content recommendations for users.
Main build: Python 3.8+, Streamlit, OpenAI API
Features
Content input via URL, text, or document upload.
Integrates content playbook and external guidelines.
Automated content rewriting with tone, structure, accessibility, and SEO guidance.
Provides actionable fixes for tone, structure, accessibility, and SEO.
Governance report card with scores.
Side-by-side comparison of original and improved content.
Download options for revised content.
Below are screenshots of the app with two input–output simulations, along with downloaded samples of the revised content in Word and HTML formats.
Learnt that prompt engineering isn’t just about asking—it’s about asking right to get useful, consistent outputs.
Saw firsthand how small tweaks in phrasing or structure can dramatically improve the clarity and quality of AI responses.
Realised the quality of content you get is only as good as the quality, clarity, and context of what you feed in.
Reinforced that content governance isn’t about control — it’s about enabling clarity and consistency.
