All work

Desktop AI tutor

Clicky for Windows

Hold a shortcut and ask about anything on screen. Clicky talks you through it and points to where to look or click.

Windows app
250+ installs
OpenAI-backed PyCon SG · US$3K+ in prizes
1st Runner-Up

I built a Windows version of Farza's Clicky with per-app memory and editable notes.

Making Clicky respond faster

I built the Windows version of Farza Majeed's Clicky, adding per-app memory and a folder for notes about software the model doesn't know.

Clicky captures the screen and loads recent memory alongside transcription, then starts speaking before the full answer is finished. In the benchmark, playback started about 190ms after a sentence was ready. The latency article explains how the parallel work and streaming speech fit together.

Why the memory is plain text

Each Windows app gets a plain Markdown file that the user can open or edit. SQLite tracks the apps and their usage. I skipped a vector database because inspectable memory mattered more than semantic search at this scale.

Checking where the cursor lands

I tested seven "where is X?" questions on screenshots from Granta EduPack, Fusion 360 and GitHub. Gemini 3.1 Pro missed every target. Its answers used a 0 to 1000 coordinate grid, while Clicky expected screen pixels.

Converting those same recorded coordinates into screenshot pixels put six of the seven inside the target buttons. I added that conversion to Clicky's Gemini integration. The updated default models each hit all seven targets in the same test.

Drawing on the screen

With Draw on screen turned on in Settings, Clicky can circle mistakes and add short labels, arrows or underlines.

Clicky circling two swapped signs in a handwritten exergy balance, labelled should be minus and should be plus
Asked "What is the mistake in this question?" on a thermodynamics worksheet.

Testing Windows integrations

Clicky has to coordinate global shortcuts, audio devices and an overlay on top of other apps. Tests cover shortcuts, audio devices, overlay positioning, speech sessions and provider setup.

Next case studyProof