agentleFS
matthewdonsemail-labmatthewdonsemail-lab/log/.claude/skills/log-inbox-voice-note/SKILL.md

log-inbox-voice-note

Skill63 starsChanged 2 days agoMITView on GitHub

Sending a voice note. A log play (docs/useCases/inbox/voiceNote.md): Send someone a voice note in the owner's own cloned voice, as the platform's own voice note, on Discord, Instagram or Slack. It is made from the owner's words by the voice provider ([Inworld](https://inworld.ai/signup?ref=FSNLMJ2H) by default, [Fish Audio](https://fish.audio/?aff=Q6W4DNSLJCSXG) too) and goes through the platform's engine and the account's one session, never the browser (rule 142). A voice note only goes out when the owner asks for one.

The file

We split this skill into the rules it gives an agent

1 rule. About 0 are common, 1 is distinctive, and 0 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈1

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 0

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Write it the way the owner talks, loose and short.

    line 41

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Sending a voice note

This is the log play inbox/voiceNote. Before anything else, call the log MCP tool getGuide({ useCase: "inbox/voiceNote" }): it returns this page and makes it the session's play, so every log tool's answer says which step you are on and a send with no play loaded is refused (rules 157, 158). Then follow the steps in order.

Draft every message from that person's own thread and matt's real replies (voice with no text, search), never from memory, and never send two that read alike (sameAsSent).

Send someone a voice note in the owner's own cloned voice, as the platform's own voice note, on Discord, Instagram or Slack. It is made from the owner's words by the voice provider (Inworld by default, Fish Audio too) and goes through the platform's engine and the account's one session, never the browser (rule 142). A voice note only goes out when the owner asks for one.

How the owner asks for it: "send <person> a voice note saying …", "voice note him that …", "send a vn to the group saying …"

Run it from a fresh context

Say: "go on docs/useCases/inbox/voiceNote.md for <person or chat>". The agent reads this page and runs the steps.

The steps

#StepToolDone when
1Read the chat the voice note goes todiscordRead { channel }, instagramDms { with } or slackRead { chat }, with jsonthe last messages both ways are known, and nothing new is waiting on the owner that the note would talk over
2Take what he asked to say, in his words, only cleaned for spelling; when he asks you to write it, read the whole chat and write it at the length he asks (a 40 s note is about 100 to 110 words), building on what they are talking about and never re-explaining what he already told them (rule 142)the agent, from his message and the whole chatthe words to say, as text, at the length asked; not every note three short lines
3Gate it as a voice note, unless the owner dictated the exact wordsvoice { text, to, with, context, spoken: true }the gate passes; if it refuses words the agent drafted, rewrite them; when the owner dictated the exact words, skip this step and use mattsWords in steps 4 and 5 (rule 137)
4Dry run itdiscordVoice, instagramVoice or slackVoice { account, chat, text, mattsWords when they are the owner's dictated words }, or voiceNotesSend { account, chat, text } when they aren't (the platform is the account's)the dry run names the chat, the words and that the clone is ready
5Send itdiscordVoiceLive, instagramVoiceLive or slackVoiceLive { account, chat, text, mattsWords when they are the owner's dictated words, approvedBy: "<owner>", confirmLive: true }, or voiceNotesSendLive { account, chat, text, approvedBy: "<owner>", confirmLive: true } when they aren'tthe platform answers sent with the message or file id
6Tell the owner what went outthe reply to the ownerthe words said, how long it ran, and where

On Discord the chat flag is channel, on Instagram with. If there is no clone yet on the provider, cloneVoice lists the owner's own voice notes and cloneVoiceLive (confirmLive: true) makes it once; cloneSpeak says a test line. provider picks inworld or fishAudio (default LOG_VOICE_PROVIDER, else inworld). When the owner dictates the exact words ("say this: …", or the words straight from the owner's message), they are the owner's own speech: pass mattsWords, which skips the voice gate's read (it grades speech as typed text and refuses it) and still refuses AI artifacts (rule 137). Words the agent drafted are never sent with mattsWords. The voice note is made from text only, never from an audio file. A refusal stops it, and nothing is tried again until the owner says so (rule 94). A send refused with accountBehind means the account's catch-up was refused: tell the owner, and only on the owner's word run inboxResume { platform, account, by: "<owner>" }, then send again once it has caught up (rule 94). On X it is twitterVoice / twitterVoiceLive { account, with: "<handle>", text }: an X Chat voice message, sent the way X Chat's own app sends one (the m4a encrypted with the conversation key, uploaded to the conversation's media store, one AUDIO attachment), with X Chat opened by the account's saved PIN; with no PIN it is refused xchatPinMissing, and the owner saves it with log twitter chat pin in a terminal. An X Chat group takes its id as with (g and digits), the same as a dm to the group takes it as userId. A 401 from X while uploading is the upload built the wrong way, not a refusal: it is fixed and sent again in the same turn, never left for later. The parts go to the resume_upload_url X's Initialize answers, under the same message id the voice note is then sent with. Only a rate limit, a block or the governor's cap is a refusal that stops (rule 94); when the owner says to reset the cap, it is reset and the catch-up resumed with inboxResume.

A voice note after a cold opener goes one step further than the opener, never back over it. The opener already forwarded their tweet and said what was mad about it, so the voice note doesn't name the tweet again. It asks what comes next ("are you building that for yourself or are you gonna sell it?"). It also never talks as if they had answered: no "then" or "though", as in "is that something you're building yourself then". In one run, every voice note that restated the tweet was held for repeating the opener (buildsOn 0.60), and every one that read as a reply was held for talking as if the chat were live (fitsTheRoom 0.37 to 0.49). Gate it inside the send, where the gate sees the opener: a voice check run on its own scores it as a cold first message. A queued voice item that is held is rewritten and put back with twitterQueueEdit, which needs the queue's run stopped first (jobCancel), then run again.

Fast, and in his spoken words. The live send says and converts the note while the account catches up and the gate reads it. A note that passes goes straight out, without waiting for it to be made (about 20 s saved on a 12 s note); one the gate refuses is never sent. Most of the wait is a refused draft, so get it right first time:

  • Answer what they last said. When someone in a group asked whether anyone else was making content, the note that passed opened on that question.
  • Write it the way the owner talks, loose and short. A tidy list of three things ("connect your Twitter, your Discord, your Instagram"; "finds them, messages them and replies to them") is flagged as an AI tell (VerbTricolon).
  • Explain log only with what the owner has said about it: you plug Claude into your socials, it reads everything, you tell it to go find people and it messages them for you.

The first send after a restart can wait up to a minute while the account catches up from its gap.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.