A 3,000-word article from blank brief to published in 2 hours, scoring 85-92/100 on a structured quality rubric. That’s the engine. The process below is what makes the speed survive the quality check. I run 500k.io solo. The content output averages 2-3 articles per day shipped, 4-6 days per week. None of it is auto-generated slop. Every article passes a voice scan, a fact scan, and a structural scan before it lands.
The agency context matters because it’s where I learned this. The Kreators AI, the agency I co-founded with Jack, manages ~$45M of Meta Ads ($10M on my side). We ran content factories for clients before AI made it cheap. The same process — but slower and more expensive. Now I run a tighter version solo on 500k.io. Same outcome, 5x less cost, 3x more output.
I’m at $9,500 MRR / $114K ARR / 22.8% to my $500K target. The content engine is the single highest-leverage thing on the site. If I had to keep one tool from the 13-tool stack and drop the other 12, I’d keep Claude Code.
The 6 stages
Each stage has a fixed time budget and a fixed deliverable. If a stage runs over, I cut scope or push the article to tomorrow. Hard rule.
| Stage | Time | Deliverable | Tool |
|---|---|---|---|
| 1. Brief | 10 min | 4-5 line brief in Notion | Manual + Perplexity |
| 2. Outline | 15 min | H2/H3 structure | Claude Code |
| 3. Draft | 25 min | Full 3,000-word draft | Claude Code |
| 4. Voice edit | 20 min | Voice-passed draft | Claude Code (voice subagent) + me |
| 5. Schema + links | 15 min | Internal links, FAQ, frontmatter | Claude Code (link subagent) |
| 6. Publish | 10 min | Committed, scheduled | Git + manual |
| Total wall-clock | ~95 min | Live article | — |
The 2-hour budget includes ~25 min of buffer for the unpredictable parts (deeper research, voice rewrite, internal-link disagreement).
Stage 1 — The brief (10 min)
This is the highest-leverage 10 minutes in the process.
Brief format I use:
Article: [working title]
Slug: [proposed slug]
Format: playbook | tutorial | essay | comparison | case-study | data | review
Target words: 3000-3500
Cluster: [ai-coding | ai-content | ai-money | ai-marketing | ai-news]
Hook: [1-line angle that's specific and opinionated]
Notes: [3-5 bullets of context I have that Claude doesn't]What makes the brief work:
- The hook is specific. “How to use AI” is a bad brief. “How to ship a 3,000-word article in 2 hours with Claude Code” is a real brief.
- The notes contain my edge. What do I know that the model doesn’t? Real numbers, specific anecdotes, a stance the SERP doesn’t cover.
- It’s short. 80-120 words. Anything longer means I’m writing the article in the brief.
I draft briefs Sunday evening for the upcoming week. 7 briefs × 10 min = 70 min. That weekly investment unlocks 7 × 95 min of execution = ~11 hours of efficient content production.
Stage 2 — Outline (15 min)
I drop the brief into Claude Code. Subagent: outliner.
Output: H2 + H3 structure with 1-line description per section. No prose.
The agent reads the voice bible, the brief, and 2-3 reference articles for tone. It returns ~10 H2 candidates ranked by GEO citation potential.
I pick 5-7 H2s. Reorder them. Sometimes rewrite the framing of one. The outline is the spine; if the spine is wrong the article is wrong.
“I almost shipped an article last week with Claude’s first-pass outline. The third H2 was ‘Best practices for X’ — pure LLM filler. I cut it. The article got 30% shorter and 50% better. The outline review is where the article actually wins or loses.”
Stage 3 — Draft (25 min wall-clock, ~5 min my active time)
The outline goes to the writer subagent. The agent:
- Reads CLAUDE.md (voice bible)
- Reads 3 reference articles for tone calibration
- Pulls 6-8 internal links from the existing archive
- Researches 3-4 external citations via WebFetch
- Drafts the article in MDX
- Writes to
src/content/blog/[slug].mdx
While Claude is drafting, I do something else. Pour coffee. Check email. Answer one Slack message. Don’t watch the agent type — that wastes the parallel time.
Average draft completion: 17-22 minutes. My active time during this stage is just opening and closing the file at start/end.
Stage 4 — Voice edit (20 min)
This is where the article becomes mine instead of generic.
Step 1: Run voice-bible-enforcer subagent. It scans the draft and returns:
- List of banned phrases used
- Paragraphs scored as LLM-tone (over a threshold)
- Missing structural elements (TLDR, FAQ, definition leads)
Typical flag count on a first draft: 3-7 issues. I fix all of them in 10 minutes.
Step 2: Voice pass myself. I read the article aloud (yes, aloud). Anywhere it sounds like a content marketer pretending to be a founder, I rewrite. Usually 4-6 sentences per article.
Step 3: Specific-numbers pass. I check that every H2 has at least one concrete number, that the real numbers (Maxime’s $9.5K MRR, $45M agency, etc.) are baked in naturally, that any claimed stat has a source.
10 minutes. Done.
Stage 5 — Schema, links, frontmatter (15 min)
Run internal-link-builder subagent. It scans the draft, suggests 6-10 internal links from the existing archive. I accept 5-7, reject the rest.
Run schema-validator. It validates:
- Frontmatter complete
- TLDR present
- 3+ H2 sections
- 4-6 FAQs
- Real
<table>(not div grid) dateModifiedset- Author byline
Frontmatter cleanup: hero image alt text, tags array, slug consistency, meta description (155-160 chars).
5 minutes for me, 10 for the agents.
Stage 6 — Publish (10 min)
Final review. I scroll the article one last time. Look for:
- Visual flow (long paragraphs broken up)
- Tables render
- TLDR reads well
- FAQ answers are 40-80 words each
Set frontmatter draft: true for queue or draft: false for immediate publish. Schedule publishedAt to a natural human hour.
Git commit:
factory: ship [article-slug]Push. Cloudflare Pages rebuilds. Article live in 90 seconds.
Total time: 9 minutes for me, 1 minute for build.
Quality math (does this actually work?)
I run a content auditor agent on every shipped article. Score on a 0-100 rubric:
| Bucket | Weight |
|---|---|
| Voice match | 25 |
| Structural depth | 20 |
| Real numbers / sources | 20 |
| Internal link quality | 15 |
| FAQ quality | 10 |
| Frontmatter completeness | 10 |
Articles produced via the full process score 85-92 average. Articles I rushed (skipping voice edit or shortcut briefs) scored 65-75. The process is the quality, not the model.
| Stage skipped | Average score |
|---|---|
| None | 88 |
| Voice edit | 71 |
| Brief polish | 76 |
| Schema validation | 80 |
| Internal links | 84 |
| Outline review | 73 |
Skipping voice edit costs 17 points. That’s the single most valuable 20 minutes of the workflow.
What this engine doesn’t do
Three things it can’t replace:
1. Original research. If the article requires interviewing 10 founders or running a survey, that has to happen outside the engine. Add 1-3 weeks for research-heavy pieces.
2. Personal anecdote depth. Claude can drop in a [Maxime’s $9.5K MRR] reference. It can’t invent a story about Anastasia asking why I was smiling at my laptop at 11pm. Those come from me, in real-time, when I’m writing. The engine outputs the shell; I drop in the soul.
3. Truly novel takes. Claude is good at synthesizing public knowledge. It’s bad at saying something nobody has said before. Novelty comes from me — the engine just amplifies it.
Sustainable cadence
I’ve tested:
- 1 article/day for 30 days: easy, quality stays at 88+
- 2 articles/day for 14 days: doable, quality stays at 85+
- 3 articles/day for 7 days: harder, quality drops to 82
- 5 articles/day for 3 days: not sustainable, quality drops to 75
The sweet spot for solo founders: 2-3/day, 5 days/week. Total: 10-15 articles/week. That’s 40-60 per month. Compared to the 4-8/month a senior writer would ship, that’s 7-15x output at $89/mo of tools versus $5K/mo of payroll.
What the engine costs
| Tool | Cost/mo |
|---|---|
| Claude Pro | $20 |
| Perplexity Pro | $20 |
| Beehiiv (CMS-side) | $49 |
| Cloudflare Pages | $0 |
| Total | $89/mo |
The Claude Code Max upgrade ($100/mo) makes the engine faster and lets me run more agents in parallel. Optional. The $89 base is the floor.
The actual day-to-day
“Most days I run the content engine in the background while I’m doing client work. I trigger an article at 06:30 before opening email, do the voice edit at 11:00 between calls, and ship it at 17:00 before logging off. Total active time: ~75 minutes. The article shipped is one I’d have spent 6-8 hours hand-writing in 2022.”
The engine doesn’t make me a writer. It makes me a brief-writer + voice-reviewer + publisher. Those three skills compose into “content founder.” The role is real and it didn’t exist 4 years ago.
Internal links
- The real Claude Code workflow (my day-to-day) — the orchestration layer.
- Claude Code: a $500K founder’s first 30 days — start here.
- The autonomous business: AI replacing every hire — what this enables.
- The $100/mo AI stack: entry tier for solopreneurs — the floor cost.
- How AI engines actually choose citations — why the structure matters.
- GEO 2026: how to get cited by ChatGPT and Perplexity — the discoverability layer.
External sources
- Anthropic — Claude Code documentation — the tool reference.
- Frase — AI content workflow benchmarks — third-party data on AI content output speeds.
- Latent Space — content factory operator interviews — comparable workflows from other operators.
What to ship tomorrow
- Pick one article from your backlog.
- Write the brief in 10 minutes (4-5 lines, with hook + notes).
- Run Claude through the 6 stages.
- Time yourself. If you’re over 2 hours, the brief was loose. Tighten next round.
After 5 articles in this loop, you’ll have your own engine. After 30, you’ll never go back to the manual way.
FAQ
How do you actually ship a 3,000-word article in 2 hours?
By splitting it into 6 stages of 15-25 minutes each, with Claude Code drafting and me reviewing. The 2-hour figure is wall-clock; my active time is closer to 75 minutes. The agent does the structural work; I do the voice and judgment work.
Does the quality really hold up at this speed?
Yes — but only because the brief is tight and the voice bible is enforced. Articles where I shortcut the brief drop 15-25 quality points. Articles where I follow the process land 85-92/100 on the rubric every time.
Is this just AI writing dressed up?
No. The output is closer to 'me writing, with AI absorbing the structural drudge work.' If I let Claude generate without my brief and voice review, you'd smell it in 30 seconds. With the process, you can't tell.
What's the single thing that makes this engine work?
The CLAUDE.md voice bible. 580 lines of locked decisions, banned phrases, sample paragraphs. Without it, the engine produces LLM-tone slop. With it, the output passes for hand-written.
How many articles per week can you sustain?
10-15 articles per week sustained, with batch publishing. I've spiked to 25 in a single week but the quality dropped on the last 8. The sustainable rate is 2-3/day, 5 days/week.
What's the minimum stack to run this?
Claude Pro $20 + a writing environment + a CMS. That's it. My full stack adds Perplexity ($20), Beehiiv ($49), and Cloudflare ($0). $89/mo for the entire content infrastructure.