Coding Agents and Smart Glasses: 39% Right From Memory, 89% With Skills

Ask a coding agent for help with smart glasses app development and you get a fluent answer, often based on last year's market. I build Hinty Voice Assistant, which shows live hints during calls and is coming to smart glasses, so I had to work out how to put text on many different pairs. That research became a free, open-source set of Agent Skills, and then I measured whether they help. Without them, agents got 39% of the checklist right. With them, 89%. Facts here were checked on 8 and 9 November 2026.

Why do coding agents get smart glasses wrong?

Smart glasses platforms change every month, and coding agents answer from training data. SDKs open up, stores open or stay closed, and some well-known options are business-only. An agent without current sources fills the gaps with plausible guesses, and on glasses a plausible guess can cost you weeks on an SDK you cannot ship.

The agents in our test often warned that their method names came from memory. The bigger problem was the decisions. Without the skills, they:

built a Rokid app on the CXR-M SDK, which is business-only and cannot run next to the Hi Rokid app every owner has (both runs);

recommended MentraOS for a teleprompter, although MentraOS cannot distribute apps yet (both runs);

described Meta's toolkit as "camera and audio, display coming later" and missed that publishing to all users is not open;

answered "check UTF-8" or "load a custom font" for Arabic on Even G2, where apps cannot ship fonts and the firmware font has no Arabic glyphs;

built a fake SDK plus Playwright for CI instead of using the automation API of Even's own simulator;

suggested USB-C display glasses for iPhone without saying this needs an iPhone 15 or later with DisplayPort output, which rules out 16e, 17e and Air.

None of these are exotic. They are the first decisions of any glasses project: which pair, which SDK, and can I publish.

What are Smart Glasses Skills?

Smart Glasses Skills is a free, open-source set of four Agent Skills that coding agents read before they answer. It works with Claude Code, Codex, Cursor, GitHub Copilot and Gemini CLI, covers 12 platforms and gives a source link and a check date for every fact. Skills and code are Apache-2.0, the platform data is CC BY 4.0.

smart-glasses is the entry point: which glasses to build for, what can ship this month, input, audio, small-screen UI, testing without hardware and publishing.

smart-glasses-live-captions turns partial and final speech-to-text results into lines that fit each display, with a latency budget and notes on non-Latin scripts.

smart-glasses-rokid and smart-glasses-vuzix are full guides for Rokid Glasses (CXR-L SDK) and Vuzix Z100 (Ultralite SDK), which have no official agent skill.

The core idea is the display-text ladder: five ways to put text in front of the wearer, from the vendor SDK down to a web app on the glasses, USB-C display glasses used as an external display, a phone notification the glasses mirror, and voice in the ear. When one rung is not available on a pair, the app falls back to the next.

Even Realities, Meta and Google ship their own agent skills, and we do not repeat them. Ours cover the decisions across brands and tell the agent which vendor skill to install for SDK detail. The repository also has a live-captions example: one Node server and four targets (web page, Even G2 plugin, Android, iOS) on one WebSocket protocol.

How did we measure it?

We gave a coding agent 11 realistic developer requests, with the skills and without, and graded the answers blind against checklists built from official vendor documentation. We ran it twice: once on an early version of the skills, then again after fixing the gaps the first run exposed. The numbers below are the second run, which is the released version. The agent was Claude Code 2.1.318 with the Sonnet model and no web access in either arm. Claude Opus 5.5 graded every answer item by item.

Each checklist had 11 to 14 items with an official source and a date. A critical error, such as building on a business-only SDK, counted only if the grader could quote it verbatim and confirm it against the source.

Average checklist score: 89% with the skills, 39% without.

Answers that pass (70% or more, no critical error): 10 of 11 with the skills, 0 of 11 without.

Critical errors: 0 with the skills, 2 without.

The first run, before the fixes, scored 84% against 38%. The without-skills arm barely moved between runs, which shows how much is run-to-run noise.

The biggest gaps were cross-brand and publishing questions. One double-tap action on four brands scored 100% and 96% with the skills, 23% and 18% without. Publishing on Meta Ray-Ban Display next month scored 88% and 92%, against 21% in both runs. The smallest gap was voice hints on Ray-Ban Meta (100% against 79% and 63%), because Bluetooth audio is well known.

We also checked when the skills switch on. On 20 near misses, such as Meta Quest, Apple Vision Pro and Wear OS, they never triggered. On 20 prompts that should use them, 13 did in the first run and 17 after we sharpened the skill descriptions. In both runs, 3 were Even G2 questions that Even's own official skill answered, which is the split we want.

What doesn't the test prove?

The test shows that agents make better decisions with the skills. It does not show that the code runs on real glasses, because nothing has been tested on hardware yet. It used one model family and two runs, and the same team wrote the skills and the checklists.

No hardware. As of 9 November 2026 nothing has run on real glasses. The Even G2 plugin passed 15 of 15 checks in Even's official simulator 0.9.5, the web page ran in headless Chromium, and the Android and iOS apps compile in CI. Every file states its level: hardware, simulator, browser, compiles or docs-only.

No web search. An agent with search would close part of the gap. We did not measure that.

Fixes between runs. After run 1 we fixed the gaps it exposed. Run 1 alone was already 84% against 38%.

All answers from both runs, the grades, the grading prompt and the command to reproduce the run are in the repository's evals folder.

How do I install the smart glasses skills?

One command installs all four skills for the agents you choose. They are Markdown files: they do not run code on install, send data anywhere or download anything.

npx skills add hinty-eu/smart-glasses-skills

Add -a with an agent name to skip the question, -g to install for your user, or --skill smart-glasses-rokid to install one skill. Plugin installs for Claude Code, Codex, Cursor, Copilot and Gemini CLI are in the README.

Then ask your agent what you would ask a colleague who follows this market: "Which glasses can I ship a teleprompter on this month?" or "How do I show text from my server on Rokid Glasses?"

How can I help verify the skills on real glasses?

If you own smart glasses and ran an example or followed a skill, send a hardware verification report: the glasses model and firmware, the phone and companion app versions, what you tested and what happened. Each report moves a fact from docs-only to hardware.

Two questions we would like answered first: does CXR-L 1.1.2 drive the new Rokid Glasses 2, and what is the real text limit per container on Even G2? The simulator accepted 963 bytes of UTF-8 text and rejected 1,023, and the hardware limit is not documented.

FAQ

Are Smart Glasses Skills free?

Yes. Skills and code are Apache-2.0, the platform data is CC BY 4.0. There is no paid version, no account and no telemetry.

Which coding agents work with these skills?

Claude Code, Codex, Cursor, GitHub Copilot, Gemini CLI and any agent that reads Agent Skills. Other agents can read the AGENTS.md file in the repository.

Do they replace the official Even, Meta or Google skills?

No. Ours tell the agent to use the vendor skills for SDK detail, and add the cross-brand decisions plus full guides for Rokid and Vuzix.

Have the skills been tested on real smart glasses?

Not yet, as of 9 November 2026. The Even G2 example passed Even's official simulator, the web example ran in a browser, and the Android and iOS examples compile.

Which smart glasses do the skills cover?

Even Realities G2, Meta Ray-Ban Display, Rokid Glasses, Vuzix Z100, Ray-Ban Meta and other audio-only glasses, USB-C display glasses such as XREAL and VITURE, standalone Android glasses, Android XR, MentraOS, Brilliant Labs and Snap Spectacles.

Keep reading

Smart Glasses With Display in 2026: 7 Models Compared, Including Which Ones Run Other Apps

Even Hub Apps: How to Install G2 Plugins, the Best Ones and What's Missing

Meta Ray-Ban Display Apps: What's Built In, How Web Apps Work and What's Still Closed

Live hints in your glasses

These skills came out of building Hinty Voice Assistant for glasses. It listens to a conversation and, a second or two after someone asks you a question, shows a short suggested answer, also from your own documents. On glasses, your phone listens and the hint appears in your lens or in your ear. It's coming soon, with no dates yet, and testers get it first.

First access to every glasses version

Free during the test

Use it today in Chrome with Google Meet, Zoom and Teams

👉 Become a tester