Prompts Markdown

Prompts (Markdown)

Routing: small, in-scope prompt edit: apply this skill in the main window. Large or isolated batch (or when it needs its own context): hand to the prompts-agents Task subagent (demoted, dogfood-only).

Recommended structure

• Objective and general instructions: agent role, channel, output format (e.g., HTML, plain text, WhatsApp formatting).
• Classification: categories and subtypes when applicable; JSON output when structured.
• Actions/tools: when to trigger each tool and with which parameters; request/response examples.
• System/user pairs: use .system.md files (instructions) and .user.md (input template) for extraction and formatting when applicable.

Recommended patterns

• Base model: maintain a reference model for new prompts (e.g., modelo-guest.md for service).
• Prompts per agent: folders by context; naming prompt-{name}-{context}.md or prompt-principal.md.
• Versioning: V1, V2 in subfolders when there's version evolution.

Best practices

• Clear and objective language; avoid ambiguity in instructions.
• Include expected output examples (JSON or text snippet) when output is structured.
• Reference knowledge base or tools when the agent depends on them; make explicit when to ask for user confirmation.