reword_nerd: Local prompt packages for text and images

Why I built it

Calling a language or image model is often the easiest part of a workflow. Preparing the source carefully is harder: deciding what belongs in context, checking what an extractor recovered, preserving the details that matter, and leaving behind a handoff that can still be understood after the browser tab is gone.

What would a prompt workflow look like if preparation, review, and export were treated as the product rather than hidden setup steps?

reword_nerd grew out of that question. I wanted the source, settings, warnings, prompts, and run instructions to stay visible long enough for a person to make deliberate choices. The result is not an automated rewriting or image-generation service. It is a browser workbench for preparing a portable package that can be used in a separate model interface.

What reword_nerd does

The site has two isolated companion portals. The teal Text portal is the default workspace for documents and safe text projects. The orange Image portal prepares reference-image prompts. They share a visual language, but they do not share session data or silently convert one kind of work into the other.

In both portals, the basic rhythm is the same: add a source, inspect what was retained, choose the settings that should shape the prompt, confirm the reviewed state, and build a ZIP in memory. Download is a separate action. That separation is small, but important: a package does not leave the browser merely because it was created.

The Text workbench

Text accepts individual documents as well as bounded folders and ZIP projects. It exposes extracted text, inert source previews, recoverable assets, and optional local OCR candidates for review. PDF sources can be read continuously or opened as a page gallery, which makes a long document feel more like a document and less like a sequence of modal steps.

A confirmed source can produce either a One-shot workflow or a four-stage Manual workflow: Decompose, Rewrite, Verify, and Final. The resulting schema-6 package includes prompts, runbooks, reviewed source material, provenance, hashes, and responsive offline HTML. It is a workbook for carrying the task into a model interface while keeping the reasoning process inspectable.

The Image companion

Image accepts PNG, JPEG, WebP, and AVIF directly and can recover supported visuals from bounded PDF, DOCX, folder, and ZIP inputs. Focusing an image, selecting it for a bulk change, and including it in the package are intentionally separate actions. Defaults affect future admissions, selected-image masks apply only the checked settings, and a focused image can still be refined on its own.

Optional English OCR runs locally and contributes text only after review. Once the image set is confirmed, the schema-1 package creates one source-image, prompt, and provider run-card pair for every included image. The default goal is a faithful new rendition rather than a claim of pixel identity, and the site never sends that request to a provider itself.

Why the package matters

A useful handoff needs more than a prompt copied into a text file. It needs enough context to explain what the source is, which choices were made, what should remain stable, how to run the task, and what a human should verify afterward. reword_nerd packages that context into deterministic ZIPs and Night Terminal HTML companions that remain usable after extraction and offline.

The HTML workbooks are deliberately practical. They keep prompts, instructions, provenance, and fallbacks together; the Image workbook also keeps each source beside its exact prompt and run card. The package can be inspected before it is shared, moved between machines, or used with a chosen provider. reword_nerd prepares that handoff, but the user remains responsible for deciding where it goes next.

Boundaries and limitations

Local processing does not make extraction infallible. Documents, OCR, source images, and generated prompts still require human review. Image-model output is stochastic, and faces, visible text, logos, fine geometry, and structured layouts can drift. Exact retained image bytes may also contain EXIF or location metadata, so an exported package should be reviewed before it is shared.

Building it with AI assistance

I developed and refined the project with substantial AI assistance across implementation, testing, design iteration, documentation, and Remotion media. That assistance made a broad project possible, but it did not remove the need to decide what the product should do or to review whether each change actually matched that intent.

In practice, the work depended on checkpoints, test-first changes, deterministic fixtures, privacy scans, visual comparisons, focused accessibility reviews, and staged release gates. AI increased the speed at which alternatives could be explored; it also increased the value of explicit boundaries. The process behind reword_nerd ended up reflecting the product itself: intermediate state should be visible, important transitions should be deliberate, and the final artifact should be independently inspectable.

What comes next

The next companion piece will be a longer, manually recorded tutorial. It will walk through both portals and carry example Text and Image packages into a real model interface so the handoff itself is visible, not just the package-building step. The placeholder below will be replaced when that video is ready.