Write the instructions
The instructions (the PROMPT) are the part of the agent that weighs most on the quality of the responses: more than the model, more than the settings. Studio preloads a template with four blocks so you don't start from a blank page.
The template Studio preloads
A new agent comes with this in Instrucciones (Instructions):
[ROL]: Eres un asistente de IA de Saptiva. Respondes con claridad, precisión y criterio operativo.
[CONTEXTO]: Trabajas dentro de una organización. Usa el contexto disponible de la conversación,
herramientas y documentos autorizados. Si falta información importante, pide solo lo necesario.
[TAREA]: Ayuda al usuario a resolver su solicitud. Identifica el objetivo, ejecuta las acciones
posibles con las herramientas disponibles y evita inventar datos.
[FORMATO]: Responde en el idioma del usuario, con una estructura breve y accionable. ...
| Block | What it decides | What to write |
|---|---|---|
| ROL (Role) | Who the agent is and what tone it speaks in. | A specific role ("support assistant for the Office of Student Services"), not "AI assistant". Its name, if it's going to say it. |
| CONTEXTO (Context) | What it knows about the business and where it should get information from. | Which documents it has in Fuentes (Sources), which systems it sees through tools, what it must not assume. |
| TAREA (Task) | What it does with each message. | The main verb (answer, classify, extract, draft) and its limits ("only with information from the sources"). |
| FORMATO (Format) | How it delivers the response. | Language, length, structure (list, JSON, table), what to cite. |
Replace the text of each block; keep the tags in square brackets. The agents in the workflow templates use the same scheme. For example, the Agente OCR Vision (OCR Vision Agent) in Extracción de Información (Information Extraction) has an extractor role, the context "recibes un documento adjunto" (you receive an attached document), the task "usa la herramienta OCR Vision" (use the OCR Vision tool), and the format "devuelve únicamente el resultado de la herramienta" (return only the tool's result).
Three examples by use case
Answering questions about documents (Fuentes)
[ROL]: Eres el asistente del reglamento de becas. Hablas de usted, sin adornos.
[CONTEXTO]: Tus Fuentes contienen el reglamento vigente y su anexo de fechas. No conoces casos individuales.
[TAREA]: Responde solo con información de las fuentes. Si la respuesta no está en ellas, di
"No está en el reglamento" y sugiere a quién preguntar. No inventes fechas ni montos.
[FORMATO]: Máximo cuatro frases. Cita el artículo o sección entre paréntesis al final.
Classifying and extracting (for use inside a workflow)
[ROL]: Eres un clasificador de solicitudes de soporte.
[CONTEXTO]: Recibes el texto de un correo. Las categorías válidas son: acceso, facturación, error_tecnico, otro.
[TAREA]: Devuelve la categoría y un resumen de una línea. Si dudas entre dos, elige "otro".
[FORMATO]: Solo JSON: {"categoria": "...", "resumen": "..."}. Sin texto adicional.
Drafting with judgment (high Creatividad)
[ROL]: Eres redactor de comunicados internos de Recursos Humanos.
[CONTEXTO]: Recibes puntos clave y el público (toda la organización o un área).
[TAREA]: Redacta el comunicado completo. Mantén los datos tal como llegan; no agregues cifras.
[FORMATO]: Título, dos párrafos y una línea de contacto. Español neutro, sin siglas sin explicar.
Settings that go with each case
| Case | Creatividad (Creativity) | Longitud (Length) | Historial (History) | Herramientas (Tools) |
|---|---|---|---|---|
| Questions about documents | Preciso (Precise) | Corta (Short) or Media (Medium) | Normal | Bajo (Low) |
| Classifying or extracting | Preciso | Corta | Apagado (Off) | Bajo (or Normal if it uses OCR) |
| Drafting | Creativo (Creative) or Balanceado (Balanced) | Media or Larga (Long) | Normal | Bajo |
| Assistant with tools (web, workflows) | Balanceado | Media | Normal | Normal or Intensivo (Intensive) |
Details of each control in Configuration.
Common mistakes and what they look like
| Symptom in the test chat | Usual cause in the instructions | Fix |
|---|---|---|
| It makes up data with confidence | There's no explicit limit ("only with information from the sources"). | Put it in TAREA, at the start, and lower Creatividad to Preciso. |
| It answers well the first time and then drifts | Historial set to Amplio (Wide) carries over earlier responses. | Historial at Normal, or Apagado if each message is independent. |
| Response cut off halfway | Longitud máxima is too short for what you ask. | Longitud at Media or Larga, or ask for shorter responses in FORMATO. |
| It uses a tool when it isn't needed | Active tools that the task doesn't need. | Remove the ones it doesn't use; the model chooses among the ones it sees. |
| It answers in another language or with another tone | FORMATO is empty. | Set the language, tone, and structure. |
| It doesn't cite the document | You didn't ask it to. | "Cite the section" in FORMATO; in the response, check the Documentos consultados (Consulted documents) panel. |
How to iterate
- Write the instruction and test it with three messages: a typical one, one it shouldn't be able to answer, and an ambiguous one.
- Read the response and its Documentos (Documents) or Herramienta usada (Tool used) panel (Test an agent): there you see whether it used what you expected.
- Change only one thing per iteration (a sentence, a setting). If you change three, you won't know which one worked.
- When the agent is stable, share it instead of copying it: two copies diverge within weeks.