The Two-AI Editorial Pipeline
I didn't plan to build this. It just happened.
Over 40 months of worldbuilding, I developed a habit: build with one AI system (Claude), then hand the document to a second AI system (ChatGPT) and ask it to judge the work. Then I'd give ChatGPT's judgment back to Claude and ask for a counter-review.
One of the AI systems looked at this process and said something that made me pause:
"You accidentally built a two-AI editorial pipeline. That is basically how AAA game lore teams operate."
And I realized: that's exactly what happened.
What the Pipeline Looks Like
Author Decides
What to build, what matters, what the world needs next
Claude Builds
Full access to source files. Extracts, cross-references, synthesizes into comprehensive document
ChatGPT Reviews
No source files. Judges the document cold. Flags overreach, checks consistency, scores quality
Claude Counter-Reviews
Defends or corrects, citing specific source files. Identifies what the reviewer missed vs. what's genuinely new
Author Decides Canon
Reads both assessments. Checks manuscript. Declares what is real.
The Data: Convergence in Action
I ran this pipeline on a character called Elder Maelis — a being made of living magic. Here's what happened:
ChatGPT Initial
8.5
Flagged 3 items as
"possible overreach"
Claude Counter
9.0
All 3 flags confirmed
as existing canon
After Context
9.0
Both systems converge
0.0 divergence
What happened: ChatGPT, reviewing the document cold (without source files), flagged three claims as potentially invented. Claude, with access to the source files, showed that all three were established canon. When ChatGPT saw the source references, it revised its score upward. Both systems converged to 9.0/10.
The lesson: Initial disagreement was caused by access disparity, not analytical error. Once both systems had the same information, they reached the same conclusion.
Why This Works
Each AI system has a different analytical strength:
Claude (The Builder)
Narrative integration. Source-backed synthesis. Cross-referencing databases. Building documents that connect to everything else in the world. Depth.
ChatGPT (The Reviewer)
Canon validation. Boundary-checking. Structural evaluation. Asking "is this established or invented?" Fresh perspective without assumptions. Skepticism.
Together they replicate a professional editorial team:
- The builder = the writer who creates the content
- The reviewer = the editor who checks the content
- The counter-review = the fact-checker who verifies disputes
- The author = the showrunner who makes the final call
In AAA game studios, this is a team of 5-10 people. In my process, it's one person and two AI systems.
The Numbers Across 12 Documents
This wasn't a one-time experiment. I ran the pipeline across all 12 Elder documents:
85%
Agreement rate
between systems
100%
Disputes resolved
by checking sources
0.5
Average initial score
divergence (points)
0.0
Final divergence
after context sharing
The pattern: Every time the systems disagreed, the disagreement was traceable to either (a) access disparity or (b) a genuine documentation gap. Once resolved, they converged.
That convergence is the proof that the pipeline works. If two independent systems with different architectures reach the same conclusion about your worldbuilding, the worldbuilding is structurally coherent.
What the Author Does That Neither AI Can
The human role in the pipeline:
Decides what to build. Neither AI chose to work on Elder Maelis. I did.
Catches errors both miss. Both AIs missed that a character was seated, not standing. I checked the manuscript.
Declares canon. When they disagree, neither AI's opinion becomes truth. Mine does.
Provides the creative insight. "What if magic could become alive?" is a human question. The 298-line answer was collaborative. But the question started the whole thing.
One More Lesson: Conversation Is Not Content
One thing I learned the hard way: AI treats everything in a conversation as available for output. Personal details, evening routines, private asides — all of it can end up in a blog post if you're not careful.
The pipeline needs a filter: the author must review every word before it goes public. Not skim. Read. Because sometimes the AI will faithfully reproduce something that was meant to stay between you and the screen.
How to Build Your Own Pipeline
If you're a solo creator working with AI:
Step 1: Choose a builder AI — the one with the most context about your project. Give it full access to source files.
Step 2: Choose a reviewer AI — use a different system, or the same system in a fresh conversation with no context.
Step 3: After building, send the document to the reviewer and ask: "Score this. Flag anything that seems invented or overreached."
Step 4: Send the review back to the builder and ask: "Defend or correct, citing sources."
Step 5: You decide. Check the manuscript. Declare canon.
What you get: A professional-grade quality assurance process for creative work. Not perfect — but vastly better than one brain reviewing its own output.
The Bottom Line
Two AI systems. Different strengths. One human deciding.
Claude builds. ChatGPT checks. The author decides.
That's the pipeline. It works because disagreement between systems reveals gaps, agreement confirms coherence, and the human provides the one thing neither AI has: creative authority.
Part of the devlog for The Ethereal Web. Pipeline data from 12 Elder documents built and reviewed across a single 40-month creative session.
— Jorge