# Brand-Voice Content Pipelines at Cadence

Named brand-voice writer Skills we run for a handful of clients. Each one produces research-grade, Generative Engine Optimization (GEO) optimized content in the client's voice at a cadence traditional agencies cannot match.

**Stack:** Claude · Claude Skills · Airtable · Firecrawl · DataForSEO · Perplexity API
**Time:** ~2 weeks to build a brand Skill · ongoing to operate
**Cost:** ~$0.30–$1.20 per article in API fees
**Shipped:** Running against live client brands today

## Context

Volume is not the hard part of AI content. Volume in an actual brand voice, with research behind every claim, and with GEO chunking baked in. That is the hard part. Most AI content pipelines we audit fall apart at the voice layer. The output reads generic inside a week.

The fix is a brand-specific writer Skill. One per brand. The Skill encodes the voice, the house style, and the chunking rules, and it reruns against new topics without drift.

## The approach

Seven stages, one required human.

```
1. Intake. Notion brief or Airtable row (topic, keywords, brand rails)
         ↓
2. Research agent. Firecrawl + Perplexity + DataForSEO → source pack
         ↓
3. Outline agent. Structured outline chunked for GEO
         ↓
4. Brand writer Skill. Full article in the brand's voice
         ↓
5. Schema agent. FAQ / HowTo / Article schema injected
         ↓
6. Human editor (REQUIRED). 10-point rubric, 10–15 min per article
         ↓
7. Publisher agent. Formats and pushes to CMS
```

The editor is non-negotiable. Skip step 6 and the output drifts off-voice within a handful of articles.

## The 10-point quality-control rubric

1. Does the first sentence of each H2 directly answer the implied question?
2. Are there at least 3 citable chunks (300–500 words, self-contained, no pronoun dependencies)?
3. Is there a numbered list or table for at least one key concept?
4. Does the FAQ section have 3–5 genuine questions, not filler?
5. Is the brand voice consistent with the voice guide?
6. Are claims backed by named sources, not vague "studies show"?
7. Is the schema payload valid and aligned to the article's actual structure?
8. Is the meta description 150–162 characters and genuinely descriptive?
9. Are internal links included to the three nearest pillar pages?
10. Would you personally share this on LinkedIn without being embarrassed?

Anything scoring below 8 of 10 goes back for revision.

## What we shipped

- Brand-specific writer Skills for a handful of clients. Each one encodes that brand's voice, house style, and chunking rules
- The general-purpose `winston-geo-article` Skill we use for new brands before we invest in a brand-specific writer
- Zapier or Make workflows + Airtable template
- Editor rubric and training doc

## Lessons

1. Brand voice is downstream of a written voice guide. Without one, output regresses to generic LLM voice within three or four articles. The voice guide is load-bearing infrastructure.
2. Skip the human editor = skip the quality. No amount of prompt engineering fixes that.
3. Pipeline velocity is rate-limited by editor throughput, not API throughput.
4. A brand-specific Skill beats a general-purpose Skill with a long system prompt. Build the Skill once. Reuse forever.

## Related

- [See the service](/services/content-creation/)
- [Read the methodology](/playbooks/agentic-content-pipeline-10-articles-a-day/)