/ReadIt

name: ReadIt
metadata:
  short-description: "Read articles, files, pasted text, or the last answer aloud in the DA voice; pause, resume, or stop. Use Teach for interactive spoken lessons."
version: 1.0.9
evals:
  manual:
    reason: "spoken delivery must be heard for pacing, intelligibility, and whether the requested content was read"
    falsifier: "the listener hears omitted or garbled content, unacceptable pacing, or a control action that does not match the requested state"
description: Reads content aloud in the DA's Fish Audio voice (an article, a file, pasted text, or what the DA just wrote) as one continuous read that starts within seconds and can be paused or stopped. USE WHEN read it to me, read this to me, read that aloud, read it, ri, /ri, play this, say this out loud, listen to this, read the article, read your write-up, stop reading, pause the read, resume the read. NOT FOR spoken tutorials (Teach), narrated video essays (the video-essay skill), one-line status notifications (the voice endpoint), or the phone and text lines.

ReadIt

One command turns content into the DA's voice reading it. The read starts within a few seconds regardless of length: the first chunk synthesizes, a detached player starts, and the rest synthesizes while earlier chunks play. `stop`, `pause`, `resume`, and `status` act on the running read.

Ideal State

• The principal hears the requested content in the DA's voice within seconds, whatever its length, and the running read answers stop, pause, resume and status.
• What plays is a faithful spoken version of the requested piece: not a summary, and not raw markup read out.
• Nothing else speaks over the read, and a headless channel never speaks at all.

Intent Routing

| Intent | Trigger | File | |--------|---------|------| | **Read** | "read it to me", "ri", "/ri", "read this", "read your write-up", a URL or file to read | `Intents/Read.md` | | **Control** | "stop", "pause", "resume", "where are you in the read" | `Intents/Read.md` § Control |

References

| File | Read when | |------|-----------| | `Setup.md` | first-time setup: the Fish Audio account and key, the voice profile, and a read-only check |

Tool contract

bun ~/.claude/skills/ReadIt/Tools/ReadIt.ts read --file <path> [--title "<title>"] [--voice <profile>]
bun ~/.claude/skills/ReadIt/Tools/ReadIt.ts read --text "<short text>"
bun ~/.claude/skills/ReadIt/Tools/ReadIt.ts stop | pause | resume | status

• Voice: `LIFEOS_CONFIG.toml` `[da.voices.reader]` when it has a `fish_reference_id`, else `[da.voices.main]`. `--voice <name>` picks any `[da.voices.<name>]` profile that has a `fish_reference_id`; `--voice principal` uses the voice id in `FISH_PRINCIPAL_VOICE_ID`. Speech goes through `LIFEOS/TOOLS/FishTTS.ts` (Fish Audio, Drama 3 model).
• Key: `FISH_API_KEY` in the env file. Never printed, never in a URL.
• Pronunciations: `LIFEOS/USER/PRINCIPAL/PRONUNCIATIONS.json`, through the one LifeOS pass (`LIFEOS/TOOLS/Pronounce.ts`) that FishTTS runs on every chunk, homographs like "live" included.
• Audio and state: `~/Library/Caches/LifeOS/ReadIt/<run>/` (tmpdir on other platforms). Playback is macOS `afplay`.
• A headless channel (`LIFEOS_NOTIFICATION_CHANNEL=headless`, or no terminal identity) prints `skipped: headless` and exits 0. Scheduled jobs never speak.

Data & Config

User customizations live at `<configRoot>/LIFEOS/USER/CUSTOMIZATIONS/SKILLS/ReadIt/` (`configRoot` is `$CLAUDE_CONFIG_DIR` when set, else `~/.claude`). When it exists, its PREFERENCES.md, configurations and resources override the defaults; otherwise the defaults stand.

First-time setup (the Fish Audio key and voice, macOS playback) is in [Setup.md](Setup.md).

Gotchas

• **This skill sends no `/notify` voice line.** The read itself is the voice; a notification would queue in front of it and delay the start.
• **The tool strips markdown as a backstop, but a raw article is not a script.** Stripped, tables read as comma lists, headers become bare fragments, and acronyms get spelled out letter by letter.
• **A new `read` stops the running one.** There is no queue.
• **Chunks are capped at 2,400 characters and split at paragraph, then sentence, boundaries.** A single-paragraph wall of text still works; it just splits by sentence.
• **Pace is the profile's `fish_speed` (0.5 to 2.0), sent to Fish as prosody speed.** Unset means 1.0. Fish accepts at most 2,500 characters per request, which is why chunks stay under 2,400. Drama 3 tags such as `[sigh]` and `<emphasis>` pass through to the voice.
• **Source tree only:** On hosts with a low open-file limit, `bun test` from the repo root can fail with EMFILE. In the source tree, run this skill's test path instead: `bun test ./skills/ReadIt/Tools/ReadIt.test.ts`. Tests do not ship in the public release.

Examples

**Example 1: read what I just wrote**

User: "ri"  (after the DA produced a write-up)
→ Read intent: authors a spoken script from the write-up (no tables, no URLs, no headers), writes it to a temp file
→ `ReadIt.ts read --file <tmp> --title "<subject>"`
→ Reply: title, chunk count, estimated minutes; audio is already playing

**Example 2: read a URL**

User: "/ri https://example.com/some-article"
→ Fetches the article, authors a spoken script of its prose, reads it

**Example 3: control**

User: "pause"  /  "ok keep going"  /  "stop reading"
→ `ReadIt.ts pause`  /  `resume`  /  `stop`