# Prompt Taxonomy And Prompt Package Requirements

## Purpose

This document defines how system prompts and customer prompt documents should be structured, stored, and assembled for the marketing agent platform.

The goal is to make prompts:

- modular
- reviewable
- versioned
- tenant-safe
- reusable across services and channels

## Prompt Package Model

Every request should be resolved into a prompt package with four layers:

1. `Platform policy`
   - global non-negotiable rules for safety, privacy, approvals, and tool behavior
2. `Service expertise`
   - reusable marketing strategy, process, and quality guidance owned by the agency
3. `Customer context`
   - tenant-specific brand and workflow instructions
4. `Session context`
   - transient work product and clarifications created during the active session

These layers should be assembled in that order so the most stable rules are applied first and the most situational context is applied last.

## Recommended Prompt Asset Structure

```text
src/prompts/
  system/
    policies/
    services/
    channels/
    tools/
  customer/
    templates/
    document-types/
    approval-states/
```

The source-of-truth documents may live in Google Drive or Google Docs, but the application should normalize them into a consistent internal prompt asset model.

## System Prompt Documents

System prompt documents represent private agency expertise and operating rules that apply across customers.

### Required System Prompt Categories

#### `policies/`

These are global rules that should apply to all requests.

- privacy and data-handling policy
- approval and escalation policy
- claims and compliance guardrails
- provenance and citation policy
- tool invocation constraints

#### `services/`

These describe how the agency performs marketing work.

- discovery and intake method
- positioning and messaging frameworks
- audience and ICP analysis methods
- campaign planning and optimization standards
- content strategy and editorial rules
- creative brief and review standards
- research and synthesis expectations

#### `channels/`

These specialize behavior by interface and workflow.

- Teams chat interaction rules
- Copilot response shape guidance
- short-form versus long-form answer conventions
- clarification question patterns

#### `tools/`

These explain when and how the platform should use tools.

- retrieval rules
- search and research tool policies
- image and video generation constraints
- file access and tenant-boundary rules

### Required System Prompt Metadata

Every system prompt asset should include:

- `promptId`
- `title`
- `owner`
- `status`
- `version`
- `lastReviewedAt`
- `appliesToServices`
- `appliesToChannels`
- `priority`
- `sourceUri`
- `changeSummary`

### System Prompt Authoring Rules

- write durable instructions, not one-off advice
- keep policy separate from examples
- prefer many small documents over one large prompt
- make conflicts resolvable through explicit priority
- state when the system must ask a question instead of guessing

## Customer Prompt Documents

Customer prompt documents represent brand, business, governance, and historical delivery context for a specific tenant.

### Required Customer Prompt Categories

#### `brand-profile`

- company overview
- business model
- target markets
- product lines or service lines

#### `brand-voice`

- tone and style
- preferred vocabulary
- banned phrases
- reading level or formality expectations

#### `visual-guidelines`

- design principles
- color and typography references
- asset usage constraints
- image style rules

#### `messaging`

- value propositions
- proof points
- positioning statements
- approved claims and required disclaimers

#### `audiences`

- ICPs
- segments
- buying triggers
- objections and desired outcomes

#### `workflow-rules`

- stakeholder map
- approval chain
- legal or compliance review requirements
- business-unit-specific handling rules

#### `approved-work`

- approved campaign assets
- approved content examples
- approved creative references
- approved performance narratives

### Customer Prompt Metadata

Every customer prompt asset should include:

- `tenantId`
- `brandId`
- `documentType`
- `approvalState`
- `owner`
- `sourceUri`
- `version`
- `lastSyncedAt`
- `businessUnitScope`
- `roleScope`
- `campaignScope`
- `assetType`
- `retentionClass`

### Customer Prompt Authoring Rules

- keep brand instructions separate from campaign-specific instructions
- keep evergreen guidance separate from expiring launch material
- explicitly tag approval state
- avoid mixing two brands in a single reusable document
- store examples with enough metadata to explain why they were included

## Prompt Assembly Rules

The prompt assembler should:

- load required platform policy fragments first
- select service expertise based on the inferred job type
- retrieve only customer documents in scope for the user and task
- include generated session context only when it improves the current response
- exclude unapproved generated work from reusable future context
- attach provenance metadata for every included prompt asset

## Prompt Precedence

When instructions conflict, the system should resolve them in this order:

1. platform safety, privacy, and tenant-isolation policies
2. customer compliance and approval rules
3. customer brand guidance
4. service-specific execution guidance
5. session clarifications and working context

## Retrieval Boundaries

### System Prompt Retrieval

- source from a dedicated internal Google Drive or Google Docs collection
- cache aggressively because these materials change infrequently
- re-index on version or content hash change

### Customer Prompt Retrieval

- scope retrieval by tenant and brand before semantic relevance
- include only approved reusable materials for long-term context
- allow session-generated materials only within the active conversation or task workspace

## Suggested Normalized Prompt Asset Schema

```yaml
promptId: system.messaging.frameworks
title: Messaging Framework Guidance
owner: strategy-team
status: approved
version: 3.2
lastReviewedAt: 2026-04-01
appliesToServices:
  - messaging
  - copywriting
appliesToChannels:
  - teams
  - copilot
priority: 80
sourceUri: gdrive://marketing-ip/system/messaging-frameworks
changeSummary: Updated differentiation guidance and proof-point rules.
body: |
  Use the customer's approved positioning before introducing new framing.
  Ask clarifying questions when the target segment is not specified.
```

## Review Workflow

Prompt assets should move through a lightweight governance path:

1. draft
2. review
3. approved
4. retired

Only `approved` prompt assets should be available to production prompt assembly.