Skip to content
Artificial intelligence. Clearly. RSS feed
AI NEWS AI

News. Evidence. Perspective.

Build & CodeAnalysis

Katsuki Ono builds the research workspace his head switch needs

His software lets an AI agent handle costly mechanical operations while he directs changes from the manuscript itself.

Katsuki Ono works from bed, operating an iPad through one switch beside his head. Each character, and each move between research tools, has a physical cost.[3]

Ono’s paper on development with large language models is listed in the ASSETS accessibility conference’s 28 October programme.[4]

In a Japanese essay published on 31 March, Ono describes customising the Moodle learning platform with Claude Code. He directed the agent to remove unnecessary plugins, lighten the database configuration and deploy the system. Deciding which plugins were dispensable remained his job. He reports doing little manual coding; evaluating and directing the output became the central work.[1]

The distinction matters because typing speed is a poor description of those decisions. Ono’s essay argues that AI changes the combination of abilities a development task requires. It also cautions that one disabled developer’s solution will not automatically suit other people.[1]

Communication provided another concrete design problem. In its 13 May account, Google’s Project VOICE team says Ono fine-tuned Gemini with 50,000 of his own past conversations to personalise predictive text. The project uses context to suggest language with fewer operations. Here, the material supplied by the developer is his own conversational history, shaping what the system predicts he might want to say. The account describes development work, without a controlled measure of communication gains.[2]

Pointing both ways

In TailorTeX, Ono selects a passage or note and sends that context to Codex. The agent can point back to a manuscript passage, file or PDF page. Linked material stays in one vertical research stream.[3]

An early bug emptied his manuscript. His response was specific: recovery copies, a check against replacing a substantial document with near-empty content, and rollback tests for failed saves. The failure changed the workspace’s design.[3]

The public beta runs through desktop or iPad browsers, with a Mac hosting the project and agent.[3]

The design saves more than typing: it reduces the trips between tools needed to tell an agent which passage matters. Reading, pointing and requesting a change share a place.[3]

Follow the evidence

Sources & further reading

  1. Project VOICE: disability-led development and the iF Design Gold Award — Japanese original Project VOICE team / Google Japan · 13 May 2026 · Accessed 28 Oct 2026
  2. TailorTeX — creator’s project account Katsuki Ono / Devpost · Accessed 28 Oct 2026
  3. ASSETS 2026 conference schedule ACM SIGACCESS · Accessed 28 Oct 2026