Build Domain Dict Skill — Development skill for Claude Code

build-domain-dict

[](https://github.com/humaniods/build-domain-dict-skill) [](https://github.com/humaniods/build-domain-dict-skill) [](LICENSE)

Built by **Sanyam Sharma** — Voice AI Engineer, [Unpod AI](https://unpod.ai)

A skill for authoring **supervoice Domain Dictionaries** — the per-domain STT vocabulary, TTS pronunciation, and latency filler sets that a voice agent uses on a call.

You give it a domain name, a prose brief, or an existing playbook/dictionary YAML. It produces one `-dictionary.yaml` file to paste into the playground's Dictionaries pane.

Works identically in **Claude Code and Codex** — both load the same `SKILL.md` + `references/` format natively, no conversion needed.

flowchart LR
    input(["domain name / prose brief / playbook YAML"]) --> dict[Domain Dictionary YAML]
    dict --> voc[vocabulary]
    dict --> pron[pronunciation]
    dict --> fill[fillers]
    voc -->|to_keyterms| stt["STT keyterm list, fixes mishears"]
    pron -->|to_pronunciation_entries| tts["TTS respelling, fixes mispronunciation"]
    fill -->|to_filler_pool| lat["filler phrase pool, fills the dead air"]
    stt --> live(["live call"])
    tts --> live
    lat --> live

The problem it solves

Three separate call failures come from the same missing artifact:

• The agent mishears domain jargon (IFSC → "if sec", rider → "writer") because nothing told the STT which terms to expect.
• The agent mispronounces acronyms, reading IRDAI as a word.
• The agent leaves dead air while the LLM generates, because it has no filler pool in the caller's language.

Hand-written dictionaries then fail a second way: they get padded with plausible-sounding domain words. That is not