That source-preserving behavior is what people seized on. For anyone who has inherited years of old TikZ, the killer feature is being able to tweak a figure without exploding it into editor-generated soup. People contrasted it with tools like
Inkscape, which often rewrite an
SVG so heavily that small edits destroy the original file structure. The reaction was strongly positive because this addresses a real pain point for researchers and engineers who want diagrams to stay versionable, reviewable, and idiomatic inside a text-based workflow.
The most useful caveat was about what “visual editing” means for a language like TikZ. The editor can preserve
relative positioning if it already exists in the source, but new elements are generally inserted with absolute coordinates. Several comments pushed on that, arguing that idiomatic TikZ leans on
anchors, relative placement, loops, and higher-level constructs rather than pinning everything to raw coordinates. The author’s answer was blunt and persuasive: once the code uses abstractions, a drag operation often has no single obvious meaning. Moving one node might mean changing a named coordinate, changing an offset, or rewriting a relationship. That tension between clean declarative code and direct manipulation looks like the central design limit of the tool, not a bug that can be polished away.
Readers also spent time on the development process because the author says the app was built almost entirely with
Codex. The concrete numbers got attention: roughly 700 million tokens since February and an estimated API cost around $15,000, though only about $500 was actually paid via ChatGPT subscriptions. The author described a pragmatic workflow: build a narrow
parser-to-
SVG renderer first, expand features incrementally, and validate output by comparing the app’s rendering with LaTeX and
dvisvgm on real TikZ examples, then fixing mismatches manually rather than trusting automated image comparison. That made the “AI-built” claim feel less like magic and more like brute-force leverage on a project that would otherwise be too tedious to attempt. The thread’s broader conclusion was that this is exactly the kind of narrow, annoying, underserved software niche where coding agents can produce something genuinely useful.