# Building an agentic content pipeline that produces 10 articles a day without sounding like slop

<span class="byline">by John Morabito · April 18, 2026 · 14 min read</span>

**TL;DR**
- Volume is not the hard part. Quality at volume is.
- Five stages: research, outline, draft, edit, publish. Agents do 1, 2, 3, 5. Humans own 4.
- The edit stage is the moat. Skip it and you ship slop.

## The architecture

```
  [ topic brief ]
        v
  1. RESEARCH  (Research Skill + Firecrawl + Perplexity API)
        v  source map, quotes, stats
  2. OUTLINE   (Outline Skill + voice guide + chunk rubric)
        v  structured outline
  3. DRAFT     (Drafter Skill + branded prompt library)
        v  draft with inline citations
  4. EDIT      (human senior editor + 10-point rubric)  <- the moat
        v  approved manuscript
  5. PUBLISH   (Publisher Skill + CMS API + schema validator)
        v  live URL
```

## Stage 1: Research (agent)

Research is where volume gains compound. An agent can crawl 100 sources in the time a human reads 10.

```
---
name: content-research
description: Triggers on "research {topic}".
---

1. Search Perplexity API for the 10 most-cited sources.
2. Crawl each with Firecrawl; extract supporting or
   contradicting quotes.
3. Output source_map.json with quotes, statistics,
   credibility tier.
4. Flag contradictions explicitly.
```

## Stage 2: Outline (agent)

The chunk rubric gets enforced here. Every H2 must open with a direct-answer sentence.

```
---
name: content-outline
---

1. Read source_map.json + voice-guide.md.
2. Produce:
   - H1, TL;DR (3 citable bullets)
   - 5-8 H2 sections with direct-answer openings
   - FAQ with 5 Q&A
3. Output outline.md.
```

## Stage 3: Draft (agent)

The drafter writes against the outline. It does not decide structure. The outline already did that.

- Short sentences, active voice.
- Inline citations with markdown links.
- No banned clichés.
- Every H2 opens with a direct answer.

## Stage 4: Edit (human)

This is the stage most pipelines try to skip. They fail.

### The 10-point rubric

1. Every claim is cited.
2. Every source is credible.
3. Every H2 opens with a direct-answer sentence.
4. No banned clichés.
5. Voice matches the brand guide.
6. Structure beats prose. Lists and tables where appropriate.
7. TL;DR bullets are each citable.
8. FAQ questions are real, not filler.
9. Schema matches visible content.
10. The piece earns its length.

The editor's job is not to rewrite. It is to apply the rubric, cut what does not pass, and ship what does.

## Stage 5: Publish (agent)

```
---
name: content-publisher
---

1. Generate JSON-LD: Article + FAQPage + HowTo if applicable.
2. Validate with Google Rich Results Test.
3. POST to CMS API.
4. Scan last 30 posts for internal-link opportunities,
   propose edits as a PR.
5. Close task.
```

## The quality-control loop

Every correction the editor makes is a training signal.

- Cliché cut → added to banned-clichés list.
- H2 opening rewritten → original tagged as negative example in the Skill's references.
- Source deemed non-credible → removed from future research-Skill allowlist.

After 100 edited pieces, the drafter should need half the corrections it needed in the first 10.

## How to debug

- **Prose feels generic** → update voice-guide.md with new "this is us / this is not us" examples.
- **Facts wrong** → tighten research Skill's credibility threshold.
- **Structure mushy** → rework the chunk rubric in the outline Skill.
- **Editor cutting 50%+** → schedule weekly sync where editor cuts feed back into the Skill.

## FAQ

**Can you do more than 10 articles a day?**

Yes, but the editor bottlenecks before the agents do. 10 is sustainable for one senior editor. 20 needs two. 50 needs a pod.

**Can we skip the human editor?**

You can. You should not. That is how slop ships.

**How do you handle fact-checking?**

Research stage outputs cited quotes. Draft cites inline. Edit stage verifies every claim against its source. If unsupported, it is cut.

**Does this work for product content?**

Yes, with a modified research stage that reads specs and reviews instead of the open web.

**What's the cost per article at volume?**

Single digits of dollars in model cost for a 1,500-word piece. Real cost is editorial salary, and it must not be cut.