A few days ago, when REA (Reverse Engineer Anything) first came out, the internet cheered, and a wave of rewriting every piece of software in Rust began. Adobe suites and CAD all got corresponding implementation versions. After so many years in GIS, I stared at ArcGIS on my desktop and thought: why not try it? First I topped up 100 RMB on DeepSeek, then installed the official REA and got started.

Background

Of course, I did not do this purely for fun. I often need to open some GIS data on Mac and Linux, and the ArcMap document file .mxd has never been well supported. So the goal of this REA attempt was simple: to open .mxd files on Mac and be able to normally turn layers on and off and edit them, etc. For test data, I used the data shared in my previous article "GIS Data: 30m DEM and Provincial/City/County Vector Dataset of the Yangtze River Basin (with .mxd/.aprx templates)".

The original image is as follows:

Development environment

With a goal and data, I got started directly. The tools I used were:

  1. deepseek harness https://www.deepseek.com/en/harness/
  2. REA https://rea.tools/
  3. Cursor

Token configuration:

  1. 100 RMB recharge of the DeepSeek official API
  2. 10 USD Command Code
  3. magpie

The development computer was a 2020 MacBook. deepseek harness worked out of the box. I directly gave it the REA address and let it install according to the documentation by itself. After the environment was set up, we needed to find a REA target. I used ArcGIS's classic version ArcMap 10.2, which can be obtained from "ArcGIS Desktop 10.2 Complete Installation Tutorial (including Win7 32/64-bit + download address + personally tested)".

Practice process

Because this is commercial software, I will not share the prompts here. You can refer to REA's official instructions. My approximate process was like this:

In the early stage, I mainly relied on deepseek harness and REA to develop the basics first: mapping core, GP tools, web layers, geocoding, topology, versioning, MXD parsing, and raster. Then I used the Yangtze River Basin data for actual testing, and gradually added GDB background import and display thinning. This was version 1.0.

In 2.0, I started to refine details: DEM georeferencing, pyramid size, coordinate system, and the zoom misalignment caused by double flipping in reprojection. By this point, my official 100 RMB API credit had run out. Of course, I also did a lot of other development; according to API statistics, the cost was basically around 50 RMB.

Later my wallet could not hold up, so I switched to Command Code's 10 USD plan and connected it to deepseek harness through magpie. As a result, after actual testing with @U87 and other friends, it was a bit slow, and triggered the 5-hour limit after two conversations.

Finally I switched to Cursor, mainly completing symbolization color schemes, provincial capital labels, and restoration of layout elements such as legends, scale bars, and north arrows. The final effect:

Above: Yangtze River Basin data

Above: GS(2024) 0650 Tianditu data

As of when I wrote this article, the implemented features include:

  • Project and interface: support project management, layer control, definition queries, and switching between classic/modern interfaces.
  • Data read/write: support common formats such as Shapefile, GeoJSON, and GeoTIFF, as well as read-only import of FileGDB and MXD.
  • Symbols and labels: support symbol systems such as unique values and graduated colors, as well as field labels and conversion to annotations.
  • Editing and analysis: support feature drawing, vertex snapping, field calculator, and common geoprocessing tools such as buffer, clip, and overlay based on GEOS.
  • Coordinate systems and services: built-in common coordinate systems such as CGCS2000 and dynamic reprojection, and support access to online basemaps such as XYZ, WMS, and WMTS.
  • Layout and map output: support layout of legends, scale bars, north arrows, etc., and export to PDF, PNG, SVG, etc.

Pitfalls I encountered

To be honest, the functions of this small tool can already meet my original needs, and it even implements more than I wanted. The generated code is written in native Mac Swift; I cannot understand a single line anyway.

Along the way I stepped into quite a few pits. First is the wallet issue. I remind you here: if you want to play with this, definitely do not directly use the DeepSeek official API; it burns too fast. Opencode and Command Code are also not recommended, because the context is relatively large and rate limiting is easy. It is best to use a proper subscription. However, many AI models now have safety checks and will refuse this kind of illegal behavior.

Second, AI is actually very good at being lazy. At least that is what I found with several Harness tools I tried. They are even too lazy to decompile; they directly search for the software's features and hand-code them. Several times I had to force it to use REA before it would decompile. Of course, this also fully shows that current open-source solutions are already excellent, and basically have what commercial software has.

Finally, there is the roughness issue. Using AI really can quickly build something, but if there is no good input in the early stage for functional completeness, interaction details, performance, etc., it will end up a mess, fixing one place and changing another. Emmm, maybe the model I used was too low-end; switching to Opus 5.5 might be much better, but then it goes back to the first wallet issue.

Conclusion

Actually, reverse engineering technology has been around for a long time, but REA, as a library, lowered the threshold of this technology for the first time to a level that a novice like me can play with. One reality that must be admitted is that the logic of software development in the AI era has changed. However, although you can achieve some capabilities of the original commercial software, their ecosystems remain irreplaceable.

Finally, I will not share the software I implemented. All the above processes are only for personal learning and tinkering. Please also do not use them for illegal purposes.