rperez030/accessible-agent-skills

A small, curated set of Claude Code skills and hooks built to make working with Claude more accessible for developers who use a screen reader — and useful to everyone else for the same reasons accessibility-first inventions like speech synthesis, audiobooks, and the typewriter ended up benefiting far more people than they were originally designed for.
Current version: 0.2.0.
## Who this is for
The motivating audience is developers who use a screen reader. Every skill here was designed with screen-reader navigation in mind, and the project exists to make working with Claude more accessible. Beyond that, the skills are useful for anyone who:
Wants spoken feedback during long-running agent work.
Wants an audible cue when the agent is waiting on input or has finished its turn.
Wants to summarize or transcribe audio (meetings, voice notes, recorded ideas).
Wants to know what a video shows, or find one fact in it, without watching it.
Reads agent responses in a web client and benefits from heading-based navigation to skim long threads.
Prefers conversational, text-mode workflows for things like PR reviews or jumping to a file.
## What works today
Copy-mode install — skills and hooks are copied into ~/.claude/. Updates require a git pull in the clone followed by a re-run of the installer.
Claude Code only. The installer targets ~/.claude/skills/, ~/.claude/hooks/, and ~/.claude/settings.json.
macOS first. Most skills work on Windows and WSL too; Linux support is partial (TTS gaps in particular). See per-skill READMEs for specifics.
## Roadmap
Symlink-mode install (single source of truth across multiple agents, updates land via git pull).
Multi-agent installer (Codex, Gemini CLI, GitHub Copilot) — each agent gets translated/adapted skill content.
Linux TTS support for the speak-permission hook.
## The skills
Each skill has its own folder under skills/ with a SKILL.md (read by the agent) and a README.md (read by you).
spoken-feedback — Real-time text-to-speech announcements while the agent works (task starts, progress, completion, errors, permission prompts). User-initiated only.
audio-summary — Summarize, analyze, or transcribe audio files of any length via the Gemini API. Requires GEMINIAPIKEY.
describe-video — Describe a video file or YouTube link, or answer a question about what it says or shows, via the Gemini API. Can also save a video's audio or captions with yt-dlp. Requires GEMINIAPIKEY.
show-in-browser — Render the most recent agent response (or any content) as a clean, accessible HTML page and open it in your default browser.
take-me-there — Open the most relevant file and line from the conversation in your editor ($VISUAL, $EDITOR, then VS Code or Cursor).
post-review — Collect a full GitHub PR review in conversation (summary + inline comments + verdict), then post the entire review as a single GitHub submission. A more conversational alternative to the GitHub web UI for PR reviews.
## The hooks
See hooks/README.md for full details.
speak-permission.sh — Plays an attention beep and speaks a one-sentence summary every time Claude Code asks for permission to run a tool. Useful when you're controlling Claude Code remotely (for example, from claude.ai) — there's currently no screen reader feedback for permission requests in that setup, and remote control is itself a more accessible alternative to running the agent in the macOS terminal. Also useful any other time you'd rather hear that the agent is waiting on you than have to look for it.