Rapid Prototyping AI Products with Claude Projects

From Hypothesis to Working Prototype in Days

Developing AI-powered products traditionally requires months of development and significant technical resources. Using Claude Projects, we transformed this timeline from months to days, testing five critical AI capabilities and building a fully functional automated lesson plan generation system for South African teachers. This article explores how hypothesis-driven prototyping with Claude Projects enables rapid validation of AI product concepts before committing to full-scale development.

Introduction

The gap between AI capability and practical implementation remains a significant barrier in educational technology. While large language models demonstrate impressive potential, translating that potential into reliable, curriculum-aligned tools requires answering fundamental questions: Can AI reliably research curriculum documents? Can it follow complex multi-step processes? Can it integrate with existing systems? Can it produce professionally formatted outputs?

Rather than spending months building infrastructure to test these questions, we used Claude Projects as a rapid prototyping environment. Within one week, we moved from concept to a working system that generates comprehensive, curriculum-aligned lesson plans for the South African CAPS curriculum, complete with professional formatting, WordPress integration, and evaluation capabilities.

Why This Matters:

Educational technology development often fails not from lack of ambition but from building solutions before validating core assumptions. Rapid prototyping allows educators and developers to test AI capabilities quickly, identify limitations early, and make informed decisions about scaling production systems.

What is Claude Projects?

Claude Projects is a feature available to Claude Pro and Team users that transforms how we work with AI by providing three critical capabilities.

Custom Instructions

Projects allow you to set detailed, persistent instructions that define how Claude should behave. These instructions act as a “programme” for the AI, establishing roles, processes, quality standards, and output formats. Unlike standard chat conversations that require repeating context, Projects maintain these instructions across all interactions within the workspace.

Project Knowledge

Each Project includes a 200,000-token context window, equivalent to approximately 500 pages, allowing you to upload reference documents, templates, evaluation rubrics, and domain-specific materials. Claude can draw upon this knowledge base to inform every response, ensuring consistency with your standards and requirements.

Tool Access and Integration

Projects have access to web search capabilities, document fetching, and, critically, the Model Context Protocol (MCP), which enables Claude to interact with external applications like WordPress, databases, and custom APIs. This transforms Claude from a conversational assistant into an integrated system component.

The Hypothesis-Driven Approach

Before building production infrastructure, we articulated five testable hypotheses about AI capabilities:

Hypothesis 1: AI can search and synthesise curriculum content

Can Claude locate, fetch, and accurately interpret Annual Teaching Plans and CAPS curriculum documents to identify appropriate weekly topics?

Hypothesis 2: AI can follow complex multi-step processes

Can Claude reliably execute a 10-step workflow from initial input to final publication, maintaining quality and consistency throughout?

Hypothesis 3: AI can integrate with external systems

Can Claude publish content directly to WordPress via MCP, handling authentication, formatting, and categorisation correctly?

Hypothesis 4: AI can produce professionally formatted, structured documents

Can Claude generate lesson plans with sophisticated HTML formatting, visual hierarchy, and mobile-responsive design without manual intervention?

Hypothesis 5: AI can evaluate its own output

Can Claude assess lesson plan quality against a detailed rubric and identify areas requiring improvement before publication?

Rather than assuming these capabilities, we designed a real-world challenge to test them systematically.

The CAPS Lesson Plan Case Study

The challenge was ambitious: create a system that could generate comprehensive, curriculum-aligned lesson plans for South African teachers teaching the CAPS curriculum across grades R-9 and multiple subjects.

The Requirements

Each lesson plan needed to include:

Current Annual Teaching Plan (ATP) research to identify appropriate weekly topics
Fully populated 15-section lesson plan template
Professional HTML formatting with colour-coded sections, tables, and responsive design
CAPS-specific pedagogical approaches and assessment standards
Differentiation strategies for diverse learners
Self-evaluation against a 60-point quality rubric
Direct publication to caps123.co.za WordPress site

The Prototype Environment

We created a Claude Project with:

Custom Instructions: A comprehensive 10-step process document (approximately 3,000 words) detailing research methodology, content structure, formatting requirements, and publication protocols
Project Knowledge: CAPS lesson plan template, 60-point quality evaluation rubric, and formatting standards
Tool Configuration: Web search, document fetching, and WordPress MCP connection

Testing the Hypotheses

Hypothesis 1 – Validated: Claude successfully located DBE Annual Teaching Plans by conducting a web search, extracted term and week-specific content, and identified appropriate topics aligned with current school calendar dates. The system demonstrated reliable curriculum research capabilities across multiple subjects and grade levels.

Hypothesis 2 – Validated: The system followed the 10-step workflow consistently, from date analysis through ATP research, topic selection, comprehensive content research, lesson plan creation, quality evaluation, and WordPress publication. Each step built logically on previous outputs.

Hypothesis 3 – Validated: Direct WordPress integration via MCP worked reliably, with Claude handling authentication, content formatting, category assignment, and post publication. The system created properly structured posts with enhanced HTML formatting that rendered correctly across devices.

Hypothesis 4 – Validated: Claude generated sophisticated HTML with inline CSS, creating colour-coded sections, responsive tables, visual hierarchies, and professional typography. Each lesson plan included 15 fully populated sections with appropriate visual formatting for web display.

Hypothesis 5 – Validated: Self-evaluation against the 60-point rubric proved remarkably effective. Claude scored each criterion, identified weaknesses, and, when scores fell below the 75% threshold, automatically revised content before resubmitting for evaluation.

Breaking Down the Capabilities

Search and Synthesis

The system’s ability to research current curriculum requirements proved more reliable than anticipated. By providing clear search query strategies in the custom instructions, Claude consistently located official Department of Basic Education documents, extracted relevant sections, and cross-referenced content across multiple sources.

Key Insight: Success required specifying not just what to search for but how to construct queries and what to validate in results.

Process Following

The 10-step workflow remained stable across dozens of generations. Critical to this reliability was structuring instructions with clear success criteria for each step and explicit transition points between stages.

Key Insight: AI follows complex processes best when each step includes verification criteria and explicit outputs that feed into subsequent steps.

System Integration via MCP

WordPress MCP integration eliminated the manual publication bottleneck entirely. The system authenticated, formatted content with proper HTML structure, assigned categories and tags, and generated draft posts, all without human intervention.

Key Insight: MCP transforms Claude from a content generator into a system component capable of triggering actions in production environments.

Structured Output with Templates

Providing the lesson plan template and detailed formatting specifications ensured consistent output structure. Claude reliably populated all 15 sections while maintaining appropriate depth and pedagogical quality in each.

Key Insight: Templates combined with explicit formatting instructions produce remarkably consistent, structured outputs suitable for classroom use.

Self-Evaluation

Perhaps most surprising was the self-evaluation capability. Claude assessed its own outputs against the 60-point rubric with appropriate rigour, identifying genuine weaknesses and making substantive improvements during revision cycles.

Key Insight: AI can serve as its own quality assurance mechanism when provided with clear evaluation criteria and permission to iterate.

Lessons Learned

Within one week, we proved all five hypotheses and created a functional system capable of generating publication-ready lesson plans in approximately 3-5 minutes per lesson. The prototype demonstrated:

Speed: Several lesson plans generated daily without quality degradation
Consistency: Reliable adherence to CAPS standards and formatting requirements
Scalability: System handled multiple subjects and grade levels without modification
Quality: Self-evaluation maintained standards above 75% threshold consistently

Critical Lessons

1. Instructions as Code: Treating custom instructions as a programme, with clear logic flow, error handling, and quality checks, proved essential for reliable AI behaviour.

2. Templates Enable Consistency: Structured templates combined with detailed specifications produce outputs suitable for production use without extensive post-processing.

3. Verification at Each Step: Building verification and validation into each process step dramatically improved reliability over linear instruction sequences.

4. Real Tools, Real Integration: MCP integration meant we tested with actual WordPress publication, revealing practical integration challenges a simulation would have missed.

5. Prototype Reveals Limits: Rapid prototyping exposed areas where AI struggled, particularly with nuanced pedagogical decisions and context-specific cultural adaptations, informing where human oversight remained essential.

From Prototype to Production

The Claude Projects prototype validated our core hypotheses and demonstrated feasibility of automated lesson plan generation. However, prototypes serve a different purpose than production systems. While the Project successfully generated individual lesson plans, scaling to hundreds of teachers and thousands of lessons required a different infrastructure.

In Part 2 of this series, we’ll explore how we translated the validated prototype into production-grade n8n workflows, adding robust error handling, user interfaces, batch processing capabilities, and monitoring systems. The prototype answered if we could build this system; the production implementation answered how to build it at scale.

Claude Projects enables rapid, low-cost validation of AI capabilities before committing significant development resources. This hypothesis-driven prototyping approach reduces risk, accelerates learning, and ensures production systems build upon proven foundations rather than untested assumptions.

About This Work

This article draws from a rapid prototyping project conducted to validate AI capabilities for automated educational content generation. The CAPS lesson plan system demonstrated how hypothesis-driven development with Claude Projects accelerates learning and reduces risk in AI product development. The approach exemplifies practical applications of AI in education while maintaining rigorous quality standards and curriculum alignment.