Most AI content production stacks promise scale and deliver slop. The trade-off feels inevitable. Quality drops the moment volume goes up.

That trade-off is real, but the cause is rarely the AI. The cause is the workflow around it. Bad inputs to a great model still produce bad outputs. Great inputs to a great model, with the right hand-offs and the right quality gates, produce work that ships.

RZLT runs AI content production at scale across multiple client brands and its own blog at a cadence most agencies cannot match. The system is not a secret tool. It is a stack of small, opinionated choices about where AI is allowed to drive and where the human drives. This is what that AI content production scale playbook actually looks like.

The 60-in-6 reality

A recent six-week sprint shipped 60 pieces of content across formats. Long-form SEO articles, listicles, case studies, LinkedIn posts, social calendar content, and event recaps. One person running the entire pipeline. Writer, editor, and operator of the system.

Sixty pieces in six weeks is not a vanity number. It is what the stack outputs when it is running properly. The number is interesting only because of what it implies about the workflow: the writer is not typing 60 first drafts. The writer is operating a system that produces 60 first drafts, and then deciding which ones ship, which ones rewrite, and which ones get killed.

The AI content production stack

Four tools do most of the work. Each one has a specific job.

Claude (Opus). The reasoning layer. Drafts long-form articles, structures arguments, runs quality gates, audits keyword placement, rewrites for voice. Most of the visible output passes through Claude at some point.

Claude Code. The execution layer. Custom skills (small markdown files with brand voice rules, structural templates, banned-phrase lists, and quality gates) sit in a folder Claude Code reads automatically.

n8n. The orchestration layer. Connects Claude to client data, CRM signals, content calendars, and publishing endpoints. Handles the boring connective tissue: pulling research, routing drafts to reviewers, posting to channels on schedule.

ClickUp. The accountability layer. Every piece of content has a task with acceptance criteria, time estimate, and assignee. The writer does not start work without a brief. Nothing gets approved without checking against the criteria.

Nothing in that stack is novel on its own. Most teams have access to all of it. The output difference comes from how the four pieces are wired together.

Where AI drives

The honest map of which parts of the AI content workflow are automated.

First drafts. Claude writes them. The writer never starts a blog post from a blank page. A skill-loaded prompt with a topic, target keyword, and competitor URLs produces an 80% draft in under five minutes.

Research collection. Web search, source verification, competitor analysis, keyword research synthesis. Claude pulls and structures. The writer reviews and selects.

Quality gates. Em-dash checks, banned-phrase checks, keyword placement audits, link verification. All scripted. A bash command runs them in seconds against the draft.

Repetitive variations. Social posts in three lengths. LinkedIn version, X version, Threads version. Same core argument, different shape. Claude handles the reformatting.

File generation. Docx outputs for client review, formatted to spec. Pptx decks built from outlines. Pdf exports. All scripted.

Where the human drives

The parts that look automated in other people's "AI content" articles but actually are not.

Angle selection. Every piece needs a defensible point of view. Why this article exists, what argument it makes, what the reader walks away believing. That is a writer decision, made before Claude touches a keyboard.

Source verification. Every claim, every stat, every linked agency. The writer verifies. Claude can suggest sources, but the writer confirms each one is real, current, and accurate. This is non-negotiable. Hallucinated citations are the fastest way to lose credibility in any vertical RZLT writes for.

Voice calibration. The skills handle 80% of voice. The remaining 20% is judgement: when a sentence sounds too literary, when a paragraph drifts into consultant-speak, when the argument has gone soft. The writer fixes those by hand.

Final cuts. Whether a piece ships at all. Whether it ships now or after a rewrite. Whether the angle is strong enough to justify the post. Editorial judgment.

The last read. The final pass is part of the same role. After the draft is polished, the piece gets read against the original brief, the keyword target, and the broader content strategy. Tone drift across the piece. Redundant arguments. Missed internal linking opportunities. Brand-voice violations the skills did not flag. Short in time, high in impact. Skipping this pass is the fastest way to ship work that should not have shipped.

What this stack does not do

A few things this stack is openly bad at. Worth naming so the rest of the system has credibility.

It does not produce original research. The "we tested 50 queries across 12 tools" type of piece still requires running the test. Claude cannot fabricate primary data, and any team that tries to shortcut that is publishing fiction.

It does not replace strategic thinking. The decision about which 60 pieces to write, in what order, with what target keywords and what internal linking architecture, is a planning problem. Claude helps execute the plan. It does not make it.

It does not work without skills. Without the skill files (brand voice rules, structural templates, banned phrases), Claude defaults to its own voice, which sounds like a confident consultant who has never met your audience. The skills are the thing that make output usable. Building them takes weeks. Maintaining them takes ongoing edits.

The principle underneath

Speed and quality are not opposites. They are both downstream of the same input: clarity about what each layer of the system is supposed to do.

A writer who knows exactly when to let Claude draft, when to step in, and when to switch into editor mode produces better work, faster, than either a writer working alone or a team trying to fully automate. The 60-in-6 number is the output of getting those handoffs right. Anyone telling you to remove the human entirely is selling you slop. Anyone telling you AI cannot do the heavy lifting has not built the skills yet.

The system works because the boundaries are honest. AI content production at scale is not a tool problem. It is a workflow problem with a tool inside it.