ai-email-secretary

name	ai-email-secretary
description	Generates a new AI prompt module with matching JSON schema, typed parser, and unit test. Use when adding a new AI capability (e.g., a new classifier, a new extraction prompt, or a new channel intent).
author	sekpriAI team
version	1.0.0
user-invocable	true

AI Email Secretary Skill

When to use

Invoke this skill when you need to create a new AI prompt module under features/ai/prompts/.

What it generates

1. features/ai/prompts/<name>.ts containing:
   • System prompt string
   • User prompt template function
   • JSON schema for expected output
   • Typed parser with zod validation
   • Retry logic for malformed JSON
2. features/ai/prompts/<name>.test.ts with:
   • Happy path test (valid model output)
   • Malformed JSON test (retry behavior)
   • Edge case tests (empty input, missing fields)

Template structure

// features/ai/prompts/<name>.ts
import { z } from 'zod';
import { getAIClient } from '../clients';

const SYSTEM_PROMPT = `...`;

const outputSchema = z.object({
  // define expected fields
});

export type <Name>Output = z.infer<typeof outputSchema>;

export async function run<Name>(input: <Name>Input): Promise<<Name>Output> {
  const client = getAIClient();
  const response = await client.chat({
    messages: [
      { role: 'system', content: SYSTEM_PROMPT },
      { role: 'user', content: buildUserPrompt(input) },
    ],
    jsonSchema: outputSchema,
  });
  return outputSchema.parse(JSON.parse(response.text));
}

Rules

• Follow the global system prompt in specs/005-ai-agent-spec.md §2.
• Never invent facts. Use only provided context.
• Every prompt must define a JSON schema for structured output.
• Every prompt must have a typed parser that validates the response.
• Every prompt must log to ai_actions table via the orchestrator.
• Prompts must NOT import UI components or React.
• Prompts must NOT call provider SDKs directly.

Safety rules (always include in system prompt)

• Never send email without user confirmation.
• If context is missing, say what is missing.
• Store extracted memory as pending, not active.

Usage

/ai-email-secretary <prompt-name>

Example: /ai-email-secretary sentiment-classifier