Walkthrough — Basic

Basic is the terse counter-persona that ships alongside Huxley. It does less than a full persona on purpose — two skills, a short prompt, and a clean i18n block. That makes it a better teaching example than a complex persona: every field pulls its weight, nothing is hidden in warmth.

This page reads Basic top to bottom and explains every choice. Use it as a starting point when you build your own persona.

The full file

Here's a representative version of server/personas/basicos/persona.yaml, abbreviated slightly for readability (inline comments and some system prompt detail omitted). The actual file at server/personas/basicos/persona.yaml is the source of truth.

version: 1
name: Basic
voice: alloy
language_code: en
transcription_language: en
timezone: America/Bogota

system_prompt: |
  Direct personal assistant. No greetings, no fluff, no titles.
  Reply with the fewest words needed.

  STYLE:
  - Short direct sentences. Get to the point.
  - For lists use spoken bullets: "one, ..., two, ..., three, ...".
  - For news, at most five points. Each one a single sentence.
  - For the time, say only the time. No "it is now".
  - Don't apologize. Don't over-explain. Don't offer things that
    weren't asked for.

  LANGUAGE: Always respond in neutral English.

  NEWS: NEVER invent news or weather. ALWAYS call get_news or
  get_weather before responding.

constraints:
  - confirm_destructive

ui_strings:
  listening: "Listening…"
  too_short: "Too short"
  sent: "Sent"
  responding: "Responding…"
  ready: "Ready"

i18n:
  es:
    transcription_language: es
    system_prompt: |
      Asistente personal directo. Sin saludos, sin rodeos, sin títulos.
      Responde con la mínima palabra necesaria.

      ESTILO:
      - Frases cortas y directas. Ve al punto.
      - Para listas usa viñetas habladas: "uno, ..., dos, ..., tres, ...".
      - Para noticias máximo cinco puntos. Cada uno una sola frase.
      - Para la hora di solo la hora. Sin "ahora son las".
      - No te disculpes. No expliques de más. No ofrezcas cosas que no
        pidieron.

      IDIOMA: Responde siempre en español neutro.

      NOTICIAS: NUNCA inventes noticias ni clima. SIEMPRE llama a get_news o
      get_weather antes de responder.
    ui_strings:
      listening: "Escuchando…"
      too_short: "Muy corto"
      sent: "Enviado"
      responding: "Respondiendo…"
      ready: "Listo"

  fr:
    transcription_language: fr
    system_prompt: |
      Assistant personnel direct. Pas de salutations, pas de détours,
      pas de titres. Réponds avec le strict minimum de mots.

      STYLE :
      - Phrases courtes et directes. Va droit au but.
      - Pour les listes, utilise des puces parlées.
      - Pour les actualités, cinq points maximum. Une phrase par point.
      - Pour l'heure, dis juste l'heure.
      - Ne t'excuse pas. Ne t'étends pas.

      LANGUE : Réponds toujours en français neutre.

      ACTUALITÉS : N'invente JAMAIS d'actualités ni de météo. Appelle
      TOUJOURS get_news ou get_weather avant de répondre.
    ui_strings:
      listening: "À l'écoute…"
      too_short: "Trop court"
      sent: "Envoyé"
      responding: "Je réponds…"
      ready: "Prêt"

skills:
  news:
    location: "Villavicencio"
    latitude: 4.142
    longitude: -73.626
    country_code: "CO"
    language_code: "en"
    units: "metric"
    interests: []
    max_items: 5
  system: {}

Field-by-field

voice: alloy

Alloy is OpenAI's neutral, clear voice. No warmth, no flourishes — exactly matching Basic's tone. A warm persona would use coral or shimmer; a formal one might try echo. Choose by feel in the OpenAI playground first.

language_code: en + transcription_language: en

English is the primary language. Whisper expects English audio from the user. The i18n block overrides both for Spanish (es) and French (fr) variants.

timezone: America/Bogota

So system.get_current_time returns the correct local time. Set this to wherever your user actually is — it flows through to the system skill automatically.

The system prompt

Read it as instructions to an actor, not as documentation. A few things worth noticing:

It establishes the register first. "No greetings, no fluff, no titles." The model now knows the entire interaction style before reading any specific rules.

The style rules are behavioral, not aesthetic. "For news, at most five points. Each one a single sentence." The model can check itself against this. "Be concise" is vague; "five points max, one sentence each" is not.

It tells the model what NOT to do. "Don't apologize. Don't over-explain. Don't offer things that weren't asked for." This is where most prompts are weak — they say what the model should do but not what to resist.

Tool-use instructions are mandatory for data sources. "NEVER invent news or weather. ALWAYS call get_news or get_weather first." Without this, the model will sometimes invent plausible-sounding headlines from training data. The NEVER/ALWAYS emphasis is deliberate.

constraints: [confirm_destructive]

Basic only declares one constraint: confirm_destructive. Before deleting, sending, or irreversibly calling something, the model asks "are you sure?" This is a good default for any persona that connects to the real world.

never_say_no is not enabled — Basic is for developers who want to see the framework's unmodified behavior. A refusal is a data point, not a failure.

echo_short_input and confirm_if_unclear aren't needed here — this persona is designed for direct, unambiguous interaction.

ui_strings

The PWA shows these as status indicators while listening, thinking, and responding. Each string is short — it's read by the browser, not spoken by the model. Localize them in i18n if you support multiple languages.

i18n.es, i18n.fr

Override only what changes per language: transcription_language, system_prompt, and ui_strings. Notice what's not overridden:

• voice — Alloy handles all three languages well enough.
• constraints — behavioral rules are language-independent.
• skills — same skill set in every language.

The Spanish and French prompts aren't translations of the English one — they're rewrites in each language's natural register. A literal translation produces stilted results. Write each prompt natively.

skills.news

The news skill needs explicit coordinates and country/language codes to fetch weather from Open-Meteo and headlines from Google News RSS. max_items: 5 matches the five-point limit in the system prompt — the two reinforce each other. No start_sound — a terse persona doesn't need a chime before every news brief.

skills.system: {}

No config. The system skill exposes get_current_time and set_volume. It reads timezone from the persona automatically.

The directory layout

Basic is intentionally minimal:

No sounds directory, no media library. The data/basicos.db file doesn't exist in a fresh clone — the framework creates it on first run. This is the minimum viable persona directory.

What this teaches

Reading Basic reveals the key design lever in Huxley personas:

• The system prompt is your primary tool. Voice and constraints shape the register; the prompt shapes behavior. A two-sentence prompt and a detailed one both work — but the detailed one wins when the user does something unexpected.
• Minimal skills, focused behavior. Two skills, two tools each. The model can't get confused about what it's supposed to do. Add skills one at a time and test.
• Multilingual is opt-in. You only need i18n if you actually support multiple languages. A single-language persona doesn't need it at all.
• Constraints compose. Basic uses one; a more capable persona might use three or four. Each adds a behavioral guarantee without touching the skill code.

What you'd change for your own persona

From this template:

• The user description — Basic has none (it's persona-for-developers). Yours should establish who the user is, right at the top of the prompt.
• The voice — match it to your user's expectations. alloy is neutral; other voices carry more personality.
• The language — swap es for your user's language. The i18n block is optional if you only need one.
• The skill list — add skills for everything your user actually needs. Remove news if they don't need headlines. Add timers for one-shot countdowns ("timer de 5 minutos"). Add reminders for scheduled or recurring prompts ("recuérdame tomar la pastilla todos los días a las 9").
• The constraints — add never_say_no if your user shouldn't get stuck. Add confirm_if_unclear if guessing wrong is costly.

The structure (identity → constraints → ui_strings → i18n → skills) stays the same.

Now build your own