BrandGuard Standard Operating Procedure

This skill defines the complete workflow for generating brand-safe marketing content.

Workflow Overview

Follow these steps in order:

1. •Retrieve Context - Ground yourself in the knowledge base
2. •Draft Content - Write content for each channel
3. •Critique Each Draft - Score quality and identify issues
4. •Verify Claims - Ensure all claims are traceable to sources
5. •Fix Issues - Revise content based on feedback
6. •Generate Images - Create marketing visuals
7. •Save Output - Package and export final deliverables

Step 1: Retrieve Context

Before drafting ANY content, retrieve relevant context from the knowledge base.

Use the MCP tool retrieve_context:

• •Query for "brand" context to get voice/tone examples
• •Query for "product" context to get facts and features

Save the chunk IDs - you'll need them for claim verification later.

Example:

retrieve_context(query="brand voice guidelines", context_type="brand")
retrieve_context(query="product features security", context_type="product")

Step 2: Draft Content

For each requested channel, write content that:

• •Matches the brand voice (see brand-voice skill)
• •Fits channel constraints (see channel-rules skill)
• •Includes a clear CTA with the provided URL
• •Explicitly lists all factual claims you're making

When making factual claims, cite the source chunk ID:

• •"Our platform supports SSO and MFA [source: chunk_abc123]"

Step 3: Critique Each Draft

Call critique_draft for EVERY channel's content.

The tool returns:

• •brand_voice_score (0-10)
• •cta_clarity_score (0-10)
• •length_ok (boolean)
• •issues (list of problems)
• •feedback_string (natural language guidance)
• •passed (boolean - true if all scores >= 7 and length_ok)

If any draft fails: Note the feedback and plan to revise.

Step 4: Verify Claims

Collect ALL factual claims from ALL drafts.

Call verify_claims with:

• •claims: List of factual statements
• •source_chunk_ids: List of chunk IDs from your retrieval step

The tool returns for each claim:

• •supported: boolean - is this claim traceable to a source?
• •source_id: the matching chunk ID (if supported)
• •similarity: confidence score
• •quoted_span: exact text from source (if found)

For any unsupported claims:

• •REMOVE the claim entirely, OR
• •SOFTEN to opinion: "We believe..." or "designed to help..."

Step 5: Fix Issues

If critique scores < 7 OR any claims are unsupported:

1. •Review the feedback_string from critique
2. •Review unsupported claims from verification
3. •Revise the content
4. •Re-critique to confirm improvement

Maximum 3 revision cycles - if still failing, proceed to export with flags.

Step 6: Generate Images

After content is finalized, call generate_images:

• •channels: List of channels needing images
• •event_title: The event title
• •visual_style: Optional style guidance (default: "modern, professional")

Images are saved to output/images/{channel}.png.

Step 7: Save Output

Use the Write tool to save final content to output/bundle.json:

{
  "event_title": "...",
  "generated_at": "2026-01-04T...",
  "content": {
    "linkedin": {
      "headline": "...",
      "body": "...",
      "cta": "...",
      "hashtags": ["#...", "#..."]
    },
    "email": {
      "subject": "...",
      "body": "...",
      "cta": "..."
    }
  },
  "scorecard": {
    "linkedin": {"brand_voice": 8, "cta_clarity": 9, "length_ok": true},
    "email": {"brand_voice": 7, "cta_clarity": 8, "length_ok": true}
  },
  "claims_table": [
    {"claim": "...", "source_id": "chunk_abc", "supported": true}
  ],
  "images": {
    "linkedin": "output/images/linkedin.png",
    "email": "output/images/email.png"
  }
}

Quality Gates

Content MUST meet these requirements:

• •Brand voice score >= 7
• •CTA clarity score >= 7
• •Length within channel limits (see channel-rules)
• •ALL factual claims verified

Iteration Rules

• •Maximum 3 revision cycles per generation
• •If still failing after 3 cycles, export with "flags": ["quality_gate_failed"]
• •Log all iterations in the output for observability