Undertext: a scanner for hidden traps in take-home assignments

A free tool that scans a job take-home assignment (PDF, spreadsheet, image, or pasted text) for hidden instructions, invisible text, and other tricks recruiters use to catch candidates who paste the a

@swabhig

undertext.techvisit ↗

Undertext: a scanner for hidden traps in take-home assignments

watch the video↗

process.log7 fields

toolsClaude (chat)Claude API (Sonnet)pdf.jsSheetJSTesseract.jshtml2canvasNetlify FunctionsNetlify DBResendUmami

iterationsMany attempts

dozens of rounds refining detection logic, design, and deployment

timea weekend

%_ai93

questions0

no questions yet

ask how it was made — the dead-ends are usually the good part

ask about the process…sign in

~/slopfolio··

# how_it_was_made

The idea came from a real interview debrief where I was asked to explain a word in my submission I couldn't account for — an assignment PDF had contained white-on-white text instructing AI to insert it. I asked Claude to build a scanner for that exact pattern: tiny/invisible text, suspicious file metadata, hidden spreadsheet rows, and vague instructions that are actually just open-ended judgment tests in disguise. Claude wrote the detection logic and the interface; I tested it against a real assignment PDF and caught a case where the size threshold for "invisible text" was tuned too tight and missed the actual trap, which we then fixed using the real numbers from that file. I made the calls on what to detect, what to leave to human judgment, and how the results should read; Claude handled the parsing logic, the serverless backend, and the UI.

// best_dead_end

I asked for a full visual redesign inspired by a specific cartoon character's cozy, sun-lit aesthetic. Claude built it — new color palette, an original mascot, a lounging-in-the-grass header illustration. Once it was live, it just didn't fit a tool about reading things carefully, so we reverted the whole thing back to the original dark, technical "blueprint" theme.

# prompting_approach

Mostly plain-language, one change at a time — describe a problem or a feeling ("this looks off," "too fast," "I need the numbers back"), let Claude propose a fix, react to the actual result. For the core detection logic, testing against a real document with a known hidden trap mattered more than describing what I wanted in the abstract.