always

name	always
description	Manipulate the configuration and custom vocabulary of the Always voice activation daemon. Use when the user wants to view or change Always preferences (API keys, energy thresholds, silence timeouts, auto-enter behavior), add domain-specific terms to the glossary so Whisper transcribes them correctly, import vocabulary from existing STT software, or analyze conversation patterns to extract meaningful vocabulary corrections from context.
slash_command	always-config

Always — Configuration & Vocabulary Skill

Always is a voice activation daemon that transcribes speech via Groq's Whisper API. Two things make it usable day-to-day:

1. Preferences — stored in a SQLite DB managed by the always config subcommand.
2. Glossary — a glossary.json file at the project root that biases Whisper toward domain-specific terms and tells the LLM post-processor how to fix common mistranscriptions.

This skill covers both.

When to use this skill

• The user asks to set, view, or reset an Always preference (API key, threshold, etc.).
• The user wants Whisper to recognize a specific term ("add Kubernetes to the vocabulary", "Whisper keeps hearing 'cuber netties' instead of Kubernetes").
• The user wants to import vocabulary from another STT tool already on the system.
• The user wants to extract vocabulary corrections from conversation context - analyze patterns where something was transcribed incorrectly and then corrected, adding only meaningful terms (not nonsense).

Configuration — always config

always config show                                    # Show all current preferences
always config set <key> <value>                       # Set one preference
always config reset                                    # Reset all preferences to defaults

Known keys (check always config show for the authoritative list):

Key	Type	Purpose
groq_api_key	string	Groq API key (required for transcription)
deepgram_api_key	string	Optional Deepgram fallback key
stt_energy_threshold	float	Mic energy gate for STT (lower = more sensitive, default 0.05)
hear_energy_threshold	float	Mic energy gate for the always-listening loop (default 0.002)
stt_cooldown_ms	int	Cooldown between transcriptions, ms (default 1500)
stt_silence	float	Silence timeout to end a phrase, seconds (default 2.0)
stt_auto_enter	bool	Press Enter automatically after pasting (default false)
always_log_path	string	Custom log path

Examples:

always config set groq_api_key "gsk_..."
always config set stt_silence 0.4
always config set stt_auto_enter true

Vocabulary — glossary.json

Located at the project root (or set ALWAYS_GLOSSARY_PATH to override). Each entry:

{
  "term": "Kubernetes",
  "mistranscriptions": ["cuber netties", "kubernetics"],
  "frequency": 100
}

• term (string, required) — canonical spelling.
• mistranscriptions (array of strings, optional) — common phonetic misreads; the LLM post-processor will rewrite these to term.
• frequency (int, optional, default 100) — higher values get priority in Whisper's 224-token bias prompt.

Adding a term safely

Use jq so the file stays valid JSON. Always write to a temp file and move:

jq '. += [{"term":"Kubernetes","mistranscriptions":["cuber netties"],"frequency":100}]' \
  glossary.json > glossary.json.tmp && mv glossary.json.tmp glossary.json

If the term already exists, update it instead of duplicating:

jq '(map(select(.term == "Kubernetes")) | length) as $n
    | if $n > 0
      then map(if .term == "Kubernetes"
              then .mistranscriptions += ["cuber netties"] | .mistranscriptions |= unique
              else . end)
      else . += [{"term":"Kubernetes","mistranscriptions":["cuber netties"],"frequency":100}]
      end' \
  glossary.json > glossary.json.tmp && mv glossary.json.tmp glossary.json

Removing a term

jq 'map(select(.term != "Kubernetes"))' \
  glossary.json > glossary.json.tmp && mv glossary.json.tmp glossary.json

Listing terms

jq -r '.[].term' glossary.json

Importing from installed STT software

always vocab import

This scans for Dragon NaturallySpeaking, macOS Dictation, Windows Speech Recognition, Google Speech Recognition, and OpenAI Whisper, then imports their vocabularies.

Fuzzy pattern analysis from conversation

When analyzing conversation context to extract vocabulary corrections:

1. Identify correction patterns - Look for instances where:

   • A transcription was followed by an immediate manual correction
   • The user explicitly stated "I meant X" or "that should be X"
   • The same term was consistently mistranscribed multiple times

2. Filter for meaningful terms - Only add terms that:

   • Are domain-specific or technical (e.g., Kubernetes, TensorFlow, specific product names)
   • Have consistent phonetic patterns that explain the mistranscription
   • Are not common words or phrases that would naturally vary
   • Appear in multiple correction instances (not one-off typos)

3. Extract the correction pair:

   • term: the corrected, canonical spelling
   • mistranscriptions: the common phonetic misreads observed
   • frequency: based on how often the correction pattern appeared

4. Add to glossary using jq:

jq '. += [{"term":"<corrected_term>","mistranscriptions":["<mistranscription1>","<mistranscription2>"],"frequency":<count>}]' \
  glossary.json > glossary.json.tmp && mv glossary.json.tmp glossary.json

Example: If the conversation shows "cuber netties" → "Kubernetes" appearing 3 times:

jq '. += [{"term":"Kubernetes","mistranscriptions":["cuber netties"],"frequency":3}]' \
  glossary.json > glossary.json.tmp && mv glossary.json.tmp glossary.json

Token budget

Whisper's prompt field has a 224-token hard limit. Always uses ~120 words to leave headroom. Entries are sorted by frequency descending — high-frequency terms make it into the bias prompt; low-frequency ones still help the LLM post-processor.

After changing the glossary

The glossary is loaded once per daemon run via OnceLock. Restart the daemon for changes to take effect:

always stop && always start