Inspiration

A changed reference document can leave a procedure out of date. A text diff shows edited words, but a procedure owner still has to decide which checklist steps need review.

What it does

What Changed? compares old and new Markdown or plain-text documents against a saved procedure. It places exact source passages beside each checklist step, marks changed or unchanged evidence, and keeps missing or ambiguous evidence visible. Reviewers record decisions and notes, save the procedure locally, and export the review as JSON. A raw text diff remains available.

The 43-second demo follows one fictional shipping-cutoff change from 3 p.m. to 1 p.m., showing the stale checklist instruction, exact source passages, a human revision decision and the exported review record. It combines explanatory graphics with edited screenshots of the running app and locally synthesized narration. All demonstration documents are synthetic. The app records the review decision; a person decides how to revise the procedure. Additional unchanged, missing and ambiguous cases are documented in the repository.

How it was built

Built from an empty folder using an AI coding agent and the Devpost Learn Skill Pack. The public repository includes scope.md, prd.md, spec.md, the build checklist, and workflow/provenance notes. The user delegated implementation and submission preparation; the adaptation from the pack's usual learner interviews is disclosed.

The app uses dependency-free JavaScript, HTML and CSS, with a small local Node.js server. Runtime comparison is deterministic section-anchor and lexical matching. No paid API, cloud storage or model download is required. Documents and review state stay in the local browser unless the user exports them.

Challenges and validation

Exact quotations do not by themselves establish the right interpretation. The implementation therefore preserves uncertainty and leaves decisions with the reviewer. Input edits invalidate stale results, and document text is rendered literally.

Twelve engine tests and eleven browser workflow checks passed. A hand-authored twelve-case diagnostic set matched ten expected routing outcomes. Two retained misses involve unresolved cross-references and subtle hypothetical language. These results are diagnostic examples, not a real-world accuracy estimate. They do not establish superiority over a simple text diff.

What's next

Independent examples, feedback from procedure owners, cross-reference warnings, and better handling of renamed sections. These are future directions, not implemented features.

This project description was prepared with AI assistance.

Built With

Share this project:

Updates

Submission history