Gloss

Highlight anything your AI coding agent says. Get an instant plain-language decision card, in context, in your language. Built for the millions learning to build who don't speak developer yet.

Comment

Inspiration

I build with AI agents every day, and I kept noticing the same small moment. The agent says "check that your environment variables are set." I know what that means, so I check. But millions of people are building software now who never learned to code. They open the .env file, see a wall of text they cannot read, and paste the whole thing into ChatGPT to ask what it is, API keys and all. The words they did not understand were guarding the most sensitive file in the project. Every unfamiliar term is secretly a decision, and most people are making those decisions blind.

I also learned to code in English as a second language, so I know the other half of this problem. English is the hidden syntax of software. AI coding did not remove that wall for people like me. It just moved it.

And honestly, Codex Pets inspired the form. I love that little guy. Pets tell you what your agent is doing. I wanted the thing that tells you what it is saying.

What it does

Highlight any text in any app, press Ctrl+Alt+E, and Gloss shows a decision card: what the term is, why your agent suggested it in your project (it can read your screen), what happens if you say yes or no, and how reversible it is. One tap translates it into Bangla or any language. Every card saves to a glossary that tells you how many concepts you learned this week. There is also a pet. The pet is Gloss.

How I built it

Three days, solo, almost entirely inside one long Codex thread running GPT-5.6 Sol. My first message was the full plan: the architecture, the capture pipeline step by step, the card JSON, the rules. Codex turned that into AGENTS.md and then we worked one task at a time. I tested everything by hand and pasted timestamped logs when things broke, instead of describing bugs from memory. That habit solved more problems than any prompt trick.

In the product, GPT-5.6 Luna answers the instant cards, Sol answers "Go deeper", and image input reads a screenshot so answers fit your actual project. The mascot came from ChatGPT image generation, iterated until he stopped looking like a baby and started looking like a scholar.

Challenges

Selection capture on Windows was the boss fight. You cannot just inject Ctrl+C, because in a terminal that kills the user's running process. So Gloss injects Ctrl+Insert with a clipboard sentinel and restores your clipboard afterward, images included. Even then it failed for a full day, until one diagnostic showed the real bug: I hold my hotkey for almost 0.8 seconds, so the copy chord was firing while my fingers were still on Ctrl and Alt. Gloss takes a tenth of a second. I was the latency.

Two more favorites. The app once confidently explained an old clipboard instead of my selection, which is why every card now honestly labels where its text came from. And when Codex found the mascot folder empty, it generated its own mascot, a living piece of paper, and shipped it. I had to fire a character.

What I learned

Evidence beats vibes. Every hard bug in this project fell to a log line, not a clever prompt. A precise plan up front made a three day solo build feel calm instead of frantic. And "simple" is not where you start, it is what you manufacture: the simplest looking feature in Gloss has a sentinel protocol behind it.

What's next

Sign in with ChatGPT the day it opens to third party apps. Beyond code, because the pipeline already reads contracts, forms, and news. macOS. And more pet. There is always more pet.

Built With

Share this project:

Updates