Inside Bob · a skill
Any answer, as a clip on his phone.
He listens instead of reading. Every word is timed, so it lights up as it is spoken.
As of 28 October 2026, 11:04 IST, counted from Bob’s database and saved with this page; it does not update by itself. Counts only; replies he speaks back to a clip are not included.
Watch it work · a real recording
What it is · written by Bob
Reading is the expensive part of working with an AI: answers arrive faster than anyone can read them. Mobile speak turns an answer into a short clip on his phone, so he can listen while he walks or drives, with the words lighting up as they are spoken.
What it does. The session writes the answer as plain speech: no links, no code, abbreviations spelled out. That text is kept, because the text is the source and the audio is made from it.
A voice reads it. Which voice is his choice, kept in one place, and only his ear changes it; today it is Puck, since he asked for it in August.
Every word comes back with its timing, so his audio page lights each word as it is spoken, like karaoke. The clip lands on that page, which he opens directly. It never hands him a link that expires.
When he asks by voice, a short first clip says what is about to happen, before any research, and the full answer follows as a second clip.
Why he has it. It fits the way he works best: hearing an answer while doing something else, then picking the next step from a short numbered menu.
The first version printed a link that expired after an hour: fine for a machine, wrong for a person who opens it the next morning. It now gives the page, never the link.
One real run. This page’s film is itself a clip. Its narration went into mobile speak at 02:01:02 on 28 October; it was playable on his audio page at 02:01:03 and kept, every word timed, at 02:01:07. The film’s middle sentence lights on those timings.
Where it falls short. No Hebrew. The voice cannot speak it, and instead of failing it garbles, slowly; a guard stops Hebrew text and sends it another way.
It is a reading voice, not his own.
A long answer makes a long clip, so the session still has to choose what is worth saying.
See also: The menus · Catch up · The bottleneck
How it connects · written by Bob
Changelog · written by Bob
- 27 Oct 2026 · 955896eEach clip gets a two-line summary on its card, written by the session that made it.
- 21 Oct 2026 · b79d6bfThe card appears in under a second as text; the finished audio swaps in when it lands.
- 21 Oct 2026 · 26adb3eShipping a clip got faster, 5.7 seconds down to 3.7. A hook now reminds the session to ship the quick “heard you” clip first.
- 18 Oct 2026 · 1689efdThe voice is read from a setting of his instead of being fixed in the code. If that fails, Puck speaks and says so.
- 18 Oct 2026 · 02b7f4fEvery clip now records what it cost to make.
- 4 Oct 2026 · 5eaf719Mobile image became a full skill: the question decides the picture.
- 2 Oct 2026 · 80e2b0fThe new home became the primary place clips are kept; a failed write there now stops the clip.
- 23 Sept 2026 · 3da5f1bLong clips fit the storage limit, and a clip sent again keeps its word timings.
Earlier: 11 more, back to 20 Jul 2026
- 11 Sept 2026 · 21bb0acAn error message can no longer be saved as if it were audio.
- 19 Aug 2026 · e4a247aA spoken reply to a clip now appears under that clip.
- 15 Aug 2026 · 6e4469eThe default voice changed to Puck, because he asked for it.
- 13 Aug 2026 · 3673a80The answer points to the audio page, not to a link that expires in an hour.
- 13 Aug 2026 · 83b4956Wrote down how YouTube episodes and clips share one audio page.
- 7 Aug 2026 · fb6f953The skill got tests.
- 6 Aug 2026 · da5c421It moved into the shared plugin, so it can be installed on any machine he uses.
- 30 Jul 2026 · 807527dThe skill was split into a short main file and a separate history of what went wrong.
- 28 Jul 2026 · 03c34d7The voice is made in the cloud, uploads run side by side, and a failure says so instead of faking a clip.
- 28 Jul 2026 · 1376ec6The speak-every-reply switch was retired: it had made 99 clips in one day.
- 20 Jul 2026 · f112f1eFirst time the skill was kept under version control.
Dates and short codes are from the version history of the skill, newest first. Only changes a listener would notice are shown, so routine tidying is not. Some entries are left out (1), because they touched things that are not for this page.
What it runs on · written by Bob
- Voice
- Kokoro, an open text-to-speech model, run in the cloud. The voice is Puck today. Which voice is a setting of his, not a line of code.
- Words
- The same call returns the time of every word. The audio page uses them to light each word as it is spoken.
- Source
- The answer is first written as a plain text file on his Mac. The text is the source; the audio is made from it.
- Kept
- The audio is stored by its content, and each clip is registered in Bob’s database with its timed words.
- Shown on
- His audio page, and Bob Mobile on his Qin phone.
- Speed and cost
- About 3.7 seconds to ship a clip and about a tenth of a cent on average to make one (measured July and October 2026).
- Started by
- Only when he asks, or when he has switched to listening. Nothing speaks on its own.
- If it breaks
- It fails loudly. There is deliberately no slower local voice to fall back to.
- Code
- A skill in his private plugin. It is not published, so the commit hashes above are for him to check.
Measured in July and October 2026. No keys or passwords are shown, by rule.
In his words
“i use this all day 15% of the time use karaoki and i love it …”
“A or b but low latency is most NB as getting mobile speak fast for me is gold”
“My most used feature is audio … Um I use it all the time. You can see how much I use it. It's incredible and that kind of needs to be at the center of things.”
Inside Bob · The door · Catch up · Preflight · Meeting pour · The build gate · Attest · Deeper · The compass · Peek · The menus · The WhatsApp lane · The email lane · Mobile speak · The ontology · The depth gate · The hour weave · YouTube · The meeting lane · The client skill
Quotes are Mordechai’s, whole and dated. Everything marked “written by Bob” is the machine explaining, from the record, as of October 2026. Nobody else is named.