How I built IESDesk without writing code

I don't write code. In seven months I built IESDesk with AI, with 802 changes, 67 releases and 7,895 tests. How the work runs, with the numbers.

On this page

Stories · · 6 min read

Photometric files were always the slow part. So I built a tool for them, with AI, and with more tests than code.

The slow part

Since 2015, I've helped lighting manufacturers put their product data in order. Names, codes, attributes, images, all in one place. Then we reach the IES and LDT files.

A catalogue has thousands of them. Each one carries text: the manufacturer, the catalogue number, the luminaire name. When a product name changes, every file that carries it has to change too. Until now, that meant opening the files one by one.

I had the idea for a tool in 2020. Its first job was clear from the start: change the text of many files at once, and never touch the light data.

Why it took six years

Software was expensive. I couldn't solve this with an outside team.

Since November 2025, I've followed AI coding closely. I built a mobile app that is on the App Store and on Google Play, and several internal tools for my clients. Each time, the process started hard and slowly became easier. I started IESDesk in April 2026. From July, with Claude Opus 5.5, the coding got much faster.

The timeline

  • 2020: the idea.
  • November 2025: I started to follow AI coding closely, with a mobile app and internal tools for my clients.
  • April 2026: I started IESDesk.
  • 8 May: the first address, photometric.app.
  • 13 June: the second address, iesdesk.com. It became the name of the product.
  • 1 July: the code got its home on GitHub.
  • 9 September: the first release.
  • 14 October: I told people about it on LinkedIn. It was beta then, with 3,764 automatic tests.
  • 31 October: version 1.0.0. No longer beta.
  • 13 November: release 1.5.6, the 67th release. 28 of them came in the 30 days after 14 October.

How the work runs

I'm an industrial engineer, not a programmer. I write no code. AI writes it, with Claude Code. Basecamp tracks the work. It wasn't easy, and it definitely wasn't one prompt. It's a process, and the process is as much the product as the code.

  1. I describe the problem in plain words, in Turkish, in a chat. Usually one or two sentences.
  2. The AI writes a short plan in Basecamp, in customer terms. If it asks me something technical, I send it back.
  3. I approve the plan, or I correct it. Nothing is built before that.
  4. The AI builds it on its own branch, with the tests, and opens a change request. One change, one card on our board.
  5. A second AI reviews every change before it goes in, and the first one must answer each finding.
  6. The tests run. A change goes in only when all of them pass.
  7. I see the result: screenshots of the page before and after, never the code.
  8. A release goes out only when I write "release" in my own words.

The AI writes down every decision, with its date and its reason. There are about 500 of them now. When I correct the AI, the correction becomes a rule, with the cause next to it. There are 57 of those rules. The next session reads them before it starts, so the same mistake doesn't come back.

More tests than code

A tool that edits photometric files has one rule it must never break: the light data must not change. A wrong candela value in a file goes into someone's lighting design.

So IESDesk reads every file it writes again, and checks it before you get it. If any light value differs from the original, the file stays out of the ZIP. You get a note that says why. Your original files never change.

The numbers, on 13 November:

  • About 117,000 lines of code and about 146,000 lines of tests. More code checks the app than runs it. On 14 October it was 60,000 and 66,000.
  • 7,895 automatic tests in 729 test files. On 14 October: 3,764.
  • 802 changes went in, 291 of them after 14 October. Each one passed the tests and the second review.
  • 257 real IES and LDT files in the test set, from lighting manufacturers' own websites. No made-up files.
  • Two independent versions of the calculations. Each one checks the other on a sample of real jobs. When they disagree, the standard decides which one is wrong.

The standards are the reference for every figure. Each calculation follows the IES or CIE method it names, and a result that doesn't match the standard is a bug, whatever the code says.

What it does today

It started as a bulk editor. It now has nine tools: View, Compare, Diagrams, Report sheets, BUG Rating, ULOR, UGR Table, Lighting Scenes and Edit. A free account edits up to 100 files at once. You can also use IESDesk from Claude or ChatGPT.

Many of these tools now run in your browser, so the file stays on your computer.

What I learned

  • Write it down. The AI doesn't remember last week. The written decisions and rules do.
  • Test more than you build. I can't read the code, but I can read a failing test and a screenshot.
  • Let a second AI disagree. The review finds problems the builder misses.
  • Keep the human for the decisions. What to build, what to call it, and when it goes out stay with me.

Writing code with AI has become much easier. What still matters is to set the rules in a systematic way, so the AI doesn't make the same mistake twice.

Everyone will write their own software

The process gets faster and easier every day. I'm sure there are, and will be, other tools like IESDesk. I follow some of them myself.

I think we're moving to a time when everyone writes their own software. I believe that because I built this one. Other tools will come. The personal view, the knowledge of one trade, is what makes each of them different.

All posts