Claude Code Max 5x is a flat-rate $100/month plan that runs my entire content factory at 3-5 articles per day. Compared to per-token Anthropic API billing for the same volume, it saves roughly 70% — and, crucially, it caps the unpredictable token spikes that happen when an agent loops on a hard sub-task. This article walks you through the exact 9-stage pipeline I use, the prompts at each stage, and the rules that keep the output safe from Google’s Helpful Content Update.
After 18 months running content factories at the $45M agency Kreators and 9 months shipping this exact stack on 500k.io, the numbers are stable: 90-150 articles per month, 92% pass the quality gate on first run, and average per-article human time sits at 8 minutes. If you want to skim the pipeline without reading the prose, jump to the diagram. Everything else is the explanation behind it.
Why Claude Code instead of the API?
Three reasons.
- Flat-rate economics. Anthropic API at Sonnet 4.6 rates costs roughly $0.06–$0.12 per 1,500-word article when you include the inputs (system prompt + SERP context + outline). Five per day = $9–$18/day = $270–$540/month. Max 5x is $100 flat. Once you’re past ~30 articles/month, Max wins.
- Agentic-loop safety. Real factories occasionally have a sub-task that loops 8 times before settling (e.g., “write the FAQ; the FAQ failed schema validation; write it again with stricter constraints”). On the API that can quietly burn $5 in 30 seconds. On Max it’s just minutes.
- Filesystem + git native. Claude Code can read your repo, write articles to
/content, commit, push, and trigger CI. The API can’t — you’d glue it to a custom orchestrator. That orchestrator is most of the engineering work.
What does “at scale” actually mean here?
I’m not talking about 100 articles a day. That’s a spam-farm pattern, and it’s the one Google’s Helpful Content Update was built to nuke. At scale in this article means 3-5 well-edited articles a day, which compounds to 90-150 a month, which is the sweet spot the 2024–2026 case study data points to: sites editing AI-assisted content perform substantially better than fully-unedited operators. Ahrefs’ 2024 case study on AI content put the spread at +40% to -90% on traffic between edited and unedited operators, and a follow-up Search Engine Land analysis confirmed the gap held through 2025 core updates.
The factory below does the heavy lifting. The human keeps the steering wheel.
The 9-stage pipeline
1. Keyword pick → 2. SERP scan → 3. Outline draft
↓ ↓
9. IndexNow ping ← 8. Publish (git) ← 7. Schema injection
↑ ↑
6. Quality audit ← 5. Visual enrich ← 4. Article draftEach stage is a Claude Code subagent with a tight scope and a strict input/output contract. The orchestrator is the main Claude Code session, holding the article ID and routing between agents.
Stage 1 — Keyword pick
Input: the cluster you’re writing for this week (e.g., claude-code).
Output: one keyword, one format, one priority — written to a single file /.factory/today.json.
The agent reads keyword-map.csv (your version of the 215-row map), filters by cluster + priority, picks the highest-value unwritten one, and tags it with the format the map prescribes (tutorial, listicle, etc.).
This stage is boring on purpose. Don’t let the agent invent keywords — it will, and they will be the same five generic phrases every other AI site is also writing.
Stage 2 — SERP scan
Input: the keyword from stage 1.
Output: /.factory/serp.md — top-10 organic results, top-3 PAA boxes, top-3 AI Overview citations (where present), shared headings, and one hot-take counter-position.
I use Bright Data MCP because their MCP server is first-party and the free 5K-req/month tier covers a single-author factory. Apify is the drop-in alternative if Bright Data hits friction.
The hot-take counter-position is the single most important field. “Everyone is saying X. Here’s the case for ¬X.” That’s the one paragraph that earns the article its citation in ChatGPT instead of ranking-but-invisible.
Stage 3 — Outline draft
Input: SERP scan. Output: outline as a YAML tree.
The outline must:
- Lead with the direct answer in the first paragraph (≤100 words).
- Include a TLDR with 5 bullets at the top.
- Use H2 question patterns (“What is X?”, “How does X work?”, “Is X worth Y?”) — these mirror PAA queries and boost AI Overview citation chance per the 2026 research.
- Have ≥1 comparison table for any tool/topic comparison.
- End with a FAQ section (5+ Q&As).
I keep this prompt in prompts/outline-tutorial.md (one per format).
Stage 4 — Article draft
Input: outline, SERP scan, voice rules.
Output: MDX file under /content/blog/[slug].mdx with frontmatter + body.
Three guardrails make or break this stage:
- Voice rules. A 50-line
voice.mdwith the do’s (specific, opinionated, founder-to-founder) and don’ts (banned phrases, hedge words, “in today’s fast-paced world”). - 30%-original rule. At least 30% of the article must be your own data, screenshots, opinion, or counter-position. The agent flags itself when it can’t hit this — that’s a queue item.
- Length tolerance. Tutorials 1,500-2,500 words. Listicles 2,500-4,000. Comparatifs 2,000-3,500. The agent rewrites if it under- or over-shoots by more than 25%.
Stage 5 — Visual enrichment
Input: finished draft.
Output: hero image at /og/[slug].webp, plus any in-article diagrams.
Replicate Flux Schnell at $0.003/image is the cheapest reliable option. The prompt is fixed style across the site — see our brand guidelines. For diagrams, Mermaid renders server-side at build; the agent just writes the Mermaid code into the MDX.
Stage 6 — Quality audit
Input: finished draft. Output: score 0-100 + reasons.
A separate Claude Opus call (the Auditor) scores the draft against a rubric:
| Dimension | Weight |
|---|---|
| Direct answer present in first 100 words | 10 |
| TLDR present, 3-5 bullets | 8 |
| ≥3 specific numbers / cited sources | 12 |
| ≥5 H2 sections, question-format where natural | 10 |
| FAQ section with ≥5 Q&As | 10 |
| At least one strong opinion / counter-position | 12 |
| Format-specific rules (table count, schema type, etc.) | 14 |
| Voice compliance (forbidden phrases, length variation) | 12 |
| Originality estimate (≥30% original) | 12 |
≥85 auto-publishes. 75-84 goes to a queue. Below 75 returns to Stage 4 with the audit reasons fed back as repair instructions.
Stage 7 — Schema injection
Input: approved draft.
Output: JSON-LD blocks injected into the MDX <Layout schema={...}> prop.
Per page type:
| Format | Required schema |
|---|---|
| Tutorial | Article + HowTo + BreadcrumbList + Person + FAQPage |
| Listicle | Article + ItemList + BreadcrumbList + FAQPage |
| Comparatif | Article + ItemList + Product (×N) + BreadcrumbList + FAQPage |
| Tool review | Article + Product + Review + AggregateRating + FAQPage |
I run every page through validator.schema.org’s API in CI — any invalid block fails the build.
Stage 8 — Publish
Input: approved + enriched draft. Output: a git commit, a push, a Cloudflare Pages deploy.
The agent commits with a structured message: factory: publish [slug] (score 87, format: tutorial). That makes it trivial to audit factory output via git log.
Stage 9 — IndexNow ping
Input: the new URL. Output: Bing + Yandex + Naver indexed.
POST https://api.indexnow.org/indexnow with the URL list and your IndexNow key. Bing turns this into ChatGPT Search visibility and Perplexity’s Brave-backed index. This is a 10-line script that earns days of ranking time on AI search.
How to set this up in 4 steps
Step 1 — Static site + content collection
If you don’t already have one, start with Astro + the astro:content collection. This article is published from one. The schema for blog frontmatter is a 30-line zod object that catches malformed agent output before it ships.
Step 2 — CLAUDE.md master brief
A CLAUDE.md at repo root tells Claude Code:
- What this site is.
- The voice rules.
- The locked decisions (no re-debating brand, color, format inventory).
- The pipeline scope.
- The Definition of Done checklist (everything that must be true before a publish counts).
Keep it under 600 lines. Mine is ~580.
Step 3 — Subagents
Each stage is a Claude Code subagent with its own system prompt. Define them in .claude/agents/ as Markdown files with frontmatter. Names map 1:1 to pipeline stages.
---
name: outliner-tutorial
description: Builds a tutorial outline from SERP scan + keyword. Strict YAML output.
tools: [Read, Write]
---The orchestrator calls them with Agent(subagent_type: "outliner-tutorial", prompt: "...").
Step 4 — Cron
I trigger the pipeline via Cloudflare Cron at 06:00, 11:00, 14:00, 17:00, 20:00 UTC. Each cron call is a tiny webhook that posts to a queue; a worker pops the queue and runs Claude Code.
Locally, you can just npm run factory:run from the terminal. It works the same way.
“I had three ChatGPT-API content factories blow up in 2024 — runaway loops, $400 surprise bills, articles that all sounded the same. Switching to Claude Code Max with a real audit stage was the difference between a $300 spam factory and a $100 keeper. The flat rate alone bought me back 4 nights of sleep a month.” — A founder peer who shipped $500K solo in 2025
What are the common errors and how do you fix them?
| Error | Cause | Fix |
|---|---|---|
| Agent picks the same keyword twice | The factory writes-but-doesn’t-mark | Maintain keyword-map.csv status column; updater agent flips to published after Stage 8 |
| FAQs are generic | Outline didn’t seed the FAQ Qs from PAA | Stage 3 must require ≥3 PAA-anchored Qs; reject outline if missing |
| Audit score gets stuck at 80-84 | Voice rule bias | Add explicit examples of good paragraphs to voice.md (positive, not just negative) |
| Schema fails Rich Results Test | Date format | All dates must be ISO 8601 in JSON-LD; the schema-injection agent must re-format |
| IndexNow returns 422 | Bad host key | Generate a fresh IndexNow key, host the txt file at /[key].txt |
How do you verify the factory ran end-to-end?
After your first end-to-end run, you should see:
- One MDX file in
/content/blog/. - One image in
/public/og/. - One commit on
mainwith the structured message. - Cloudflare Pages deploy success.
- Bing Webmaster Tools showing the URL within ~10 minutes.
- (For approved articles) Article live at
/journal/[slug]with valid schema (run a Rich Results Test).
If any of those is missing, the failure is in the corresponding stage. The factory log at /.factory/log.jsonl tells you which one.
Going further
- Best AI tools for solopreneurs in 2026 — the rest of the stack that pairs with Claude Code.
- Claude Code vs Cursor in 2026 — pick the right base tool before scaling.
- Beehiiv review 2026 — the newsletter platform we wire to the factory’s “publish” stage.
FAQ
Can Claude Code generate SEO content at scale without getting penalized?
Yes — if you (1) auto-publish only drafts scoring ≥85 in a quality gate, (2) keep at least 30% original analysis per article, and (3) rotate through 8 formats. The penalty story is correlation; the cause is thin, repetitive output.
Why Claude Code Max instead of the API?
Flat $100/mo covers ~5 articles/day. Equivalent API spend at Sonnet 4.6 rates would be $300-700/mo for the same volume. Max also caps token spikes during agentic loops.
Do I need any other tools beyond Claude Code?
Bright Data MCP for SERP scraping (free 5K req/mo), a static site (Astro/Next), and Bing/Google indexing APIs. Optional but recommended: Replicate for hero images, IndexNow for Bing pings.
How long does the setup take?
About 4-6 hours if you already have a content site. The factory itself runs in 8-15 minutes per article once the pipeline is wired.
What's the realistic output ceiling per day?
5 articles/day with Max 5x is comfortable. Past 5/day, you start hitting the 5-hour usage windows; spread runs across the day. Max 20x scales to 15-20/day if you ever need it.