Meg Kemp
← All work

Open source · AI for Accessibility Toolkit

From what disabled people say to work developers can build

Findings from research sessions too often stay in notes. I built a session kit and a Claude skill that carry participants' needs all the way to scoped GitHub issues, with consent and check-back at every step.

How it works

The kit covers the whole path. Nothing reaches developers until the people who said it have seen how it was written up.

Community session to GitHub issue flow Five stages from left to right: plan and recruit, run the session, draft with AI, check back with participants, and publish issues. Check-back is a gate: private statements stay held until the participant confirms they can be shared. Plan and recruit Layered consent Access and pay Run the session 60–75 min guide Private answers Draft with AI Skill drafts findings, issues Check back Plain summary to participants Publish issues People decide and file issues GATE · Private statements wait until the participant confirms
The AI step drafts. The check-back step is where participants approve or correct how their words are used.

Applying the toolkit's values to the research itself

The AI for Accessibility Toolkit builds software that adapts to each person instead of asking people to adapt. I designed the research process to hold the same values.

The toolkitThe session kit
Records what a person asks for, never a conditionScreens participants and writes findings by needs and tools, never by diagnosis
Lets the person choose every changeGives participants a private way to answer and control over what is quoted
Pauses an AI agent at decisions that belong to the personChecks every finding back with participants before anything is public

Using AI responsibly

The skill drafts; people decide. It never files an issue itself.

Given session notes and each participant's consent record, it returns a findings table, a plain-language check-back summary and draft issues. It works under five fixed rules:

  • Write needs, never conditions.
  • Quote only with consent; otherwise paraphrase.
  • Keep private statements private until the participant confirms.
  • Include no identifying details.
  • Invent nothing.

What's in the kit

PieceWhat it does
Plan and recruitRecruiting, layered consent, access, pay and safety
Discussion guideA 60–75 minute guide mapped to the toolkit's own loop
Synthesis templateFrom session notes to GitHub issues
Claude skillDrafts findings, the participant summary and issues in the toolkit's format
Worked exampleA fictional five-person session traced to 16 prioritized findings and 9 draft issues

Where the practices come from

Layered consent and private response options come from my research for Catholic Relief Services' ReMiND prenatal care app in Uttar Pradesh, India. There I led recruiting, consent, research, design and validation with low-vision and low-literacy residents, working alongside community health workers.

Status

A working draft, open under the Apache 2.0 license so any part can be folded into the toolkit. The example session is fictional; no real participant's words appear in the repository.