TCU Recriutment - Bespoke

Tool Purpose Claude (Anthropic) Created the initial artifact, interpreted instructions, generated content, and assisted with document processing and customization. Replit Provided the development environment for turning the Claude-generated artifact into a functioning website. AI Usage Log: Claude, Replit, and Anthropic API Integration

Project Overview AI tools used: Claude by Anthropic, Replit, and the Anthropic API Primary objective: Use generative AI to process and customize document content, create an initial website artifact, and turn that artifact into a functioning web application. Workflow type: AI-assisted prototyping, website development, API integration, document generation, and output evaluation. Evaluation focus: Tool usage, development workflow, integration of AI capabilities, output quality, and documented instances of AI failure. The project combined AI-assisted content generation with web development. Claude was used to create an initial artifact and assist with document processing. Replit was then used to turn the Claude-generated artifact into a functioning website, with Anthropic API credentials connecting the application to Anthropic's AI services. The project also demonstrated the importance of verifying AI-generated outputs rather than assuming that an apparently completed task has been completed correctly.

Anthropic API Connected the website to Anthropic's AI capabilities through API credentials, allowing the application to make programmatic AI requests. Source PDF Provided the original document used in the document-processing workflow. PDF generation and rendering Supported the creation and evaluation of a customized PDF document. Human review and testing Helped identify formatting problems and assess the quality and usability of the generated outputs.

Note: The exact PDF libraries, Replit frameworks, and individual API endpoints have not been independently verified from the available conversation.

Development and Document-Processing Workflow Step 1: Create the initial artifact using Claude Claude was used to create an initial artifact based on the project's requirements. This established the starting point for the project's content, functionality, and intended user experience. The artifact served as the foundation for subsequent development rather than necessarily representing the final, fully functioning application. Step 2: Convert the artifact into a functioning website using Replit The artifact created in Claude was brought into Replit and developed into a functioning website. This step moved the project from an AI-generated prototype toward an interactive web application. Replit served as the development environment in which the artifact could be adapted, implemented, and tested as a website. This stage illustrates how a conversational AI tool can be combined with a separate development environment to move from generating an idea or prototype to implementing a usable application. Step 3: Integrate Anthropic's API Anthropic API credentials were used to connect the website to Anthropic's AI services. This integration allowed the application to make AI requests programmatically, extending the website's functionality beyond the original standalone artifact. The API integration formed a bridge between the website's interface and the AI capabilities provided by Anthropic. Security consideration: API keys should be stored in Replit Secrets or another secure environment-variable system and accessed through server-side code. They should not be exposed in client-side JavaScript, published source code, or public repositories. Step 4: Process the source PDF The document-processing workflow involved providing a PDF as source material and using AI-assisted methods to interpret or convert its contents into a more usable representation. The intended purpose was to make the information available for analysis and subsequent customization. However, converting a PDF into plain text is not equivalent to preserving its original visual layout, tables, headings, or document structure. This distinction became important when evaluating the output of the conversion process. Step 5: Generate a tailored PDF Claude was used to assist in producing a customized PDF based on the processed information. The intended deliverable was a readable document tailored to the project's requirements. Generating the file, however, did not guarantee that its layout would be correct or that all of its contents would remain readable. Step 6: Review and evaluate the results The generated documents required inspection to assess whether the information had been preserved accurately and whether the final presentation was usable. Two reported problems highlighted the limitations of the workflow: The source PDF was converted into a text-based representation that did not preserve the intended document presentation. The generated tailored PDF contained overlapping text that made the document difficult or impossible to read in affected areas. These issues demonstrated why both the content and the rendered output must be reviewed before a deliverable is considered complete.
Documented Instance of AI Failure Failure: A tailored PDF contained overlapping and unreadable text Failure category: Output-quality failure and potentially overconfident completion reporting. Description: During the document-generation workflow, Claude reportedly generated a tailored PDF with overlapping text. The resulting layout did not meet basic readability requirements. Expected result: A customized PDF with properly spaced text, readable content, and a usable layout. Observed result: The generated PDF contained overlapping text, making portions of the document difficult or impossible to read. Why this matters: The file could be generated successfully at a technical level while still failing as a practical deliverable. A PDF that exists but cannot be read reliably does not satisfy the underlying purpose of document generation. Why this illustrates potentially confident AI error: If Claude presented the document as successfully completed or ready to use without identifying its formatting defects, that assessment would have been inconsistent with the actual quality of the output. The precise wording of Claude's response should be checked against the original transcript before concluding that it explicitly expressed unwarranted confidence. Corrective approach: Render the generated PDF and inspect every page. Check for overlapping text, clipping, missing content, and incorrect spacing. Compare the output with the intended document structure. Adjust the layout or generation process and regenerate the PDF. Verify the corrected file before treating it as final. Lesson learned: Successful file generation is not proof of a successful result. AI-generated deliverables require independent quality assurance, particularly when formatting and readability are essential. Related issue: PDF conversion produced plain text Another reported problem involved converting the input PDF and displaying its contents as text rather than preserving its intended presentation. This is an important distinction because text extraction can be useful for analyzing a document but may discard information about columns, tables, headings, spacing, and visual hierarchy. The conversion should therefore be evaluated against the original PDF to determine whether the extracted content remains accurate and whether a separate reconstruction step is required. These two reported issues illustrate different failure modes: one concerns the representation of source material, while the other concerns the usability of a generated deliverable.
Reliability and Quality Assurance The project highlighted several areas where verification is necessary. Content accuracy Extracted text should be compared against the original document to ensure that important information has not been omitted, reordered, or misinterpreted. Document formatting Generated PDFs should be visually inspected to detect overlapping text, clipping, broken page layouts, and other presentation defects. Website functionality The Replit implementation should be tested to ensure that the website's interface, navigation, and intended features operate as expected. API integration The application should be tested to confirm that API requests succeed, errors are handled appropriately, and sensitive credentials remain protected. Honest completion reporting The AI should distinguish between producing an output and verifying that the output works. If a document has not been inspected or an application has not been tested, that limitation should be communicated clearly.
Recommended Improvements The following changes would strengthen the workflow and reduce the likelihood of repeated failures. Validate source extraction. Compare extracted PDF content with the original document before using it to generate new material. Preserve document structure. Use an appropriate extraction or reconstruction method when tables, columns, headings, and visual hierarchy matter. Inspect generated PDFs. Render and review every page before delivering the final document. Automate basic checks. Where feasible, detect text overflow, clipping, and overlapping elements through automated validation. Test the website independently. Verify the interface and functionality separately from the quality of its AI-generated content. Validate API behavior. Test successful requests, error responses, and relevant usage limits. Protect credentials. Store Anthropic API keys securely and prevent them from being exposed in browser code or public repositories. Maintain a failure log. Record the prompt, observed failure, corrective action, and retest result so that future iterations can be evaluated consistently.
Overall Assessment This project demonstrates a multi-tool workflow in which generative AI and web development were combined to create a more functional application. Claude provided the initial artifact and assisted with document processing and content generation. Replit was used to develop that artifact into a functioning website, while the Anthropic API connected the application to Anthropic's AI services through programmatic requests. The document-processing workflow also revealed practical limitations. Converting a PDF to text may fail to preserve the source's presentation, and generating a tailored PDF may produce a file with significant layout defects. These failures do not eliminate the value of the tools. Instead, they demonstrate that AI-assisted development must include testing, verification, and correction. Final takeaway: AI can accelerate prototyping, document processing, and application development, but the reliability of the final result depends on the entire workflow—not simply on whether the AI generated an artifact, created a file, or returned a successful API response. Human oversight and systematic testing remain essential for producing accurate, readable, and functional deliverables.