At ~400 pieces, one bad rule can repeat faster than an editor can catch it. Speed without control simply moves the workload into correction.
Operating on real content
AI-assisted content production: 5 materials in ~2 hours with editorial QA
When hundreds of pieces move through a system, mistakes scale too. I did not build a “text generator”; I built a workflow where sources, structure, visuals and editorial review are one cycle, and recurring feedback changes the next iteration.
Case at a glanceDesigned the workflow and quality criteria, set up the review loop and refined the rules using real editorial feedback.
About 400 materials have gone through the system; a five-article batch with visuals can be prepared in roughly two hours through the manual-placement stage, and about 95% pass the first editorial review.
~5%Needed another cycle
~95%Accepted first pass
About 95% of pieces pass the first review.
Automation scales mistakes as quickly as throughput
The strongest part of the system is not text generation but feedback: an editorial issue changes the rule for the next iteration, local defects are repaired locally, and the final material is checked after placement. That is what made faster production useful instead of simply moving the saved time into manual rework later.
Key decision06
Feedback improves both the current page and the next iteration.
How feedback improves the next material
Generalized examples from the real production loop.
| Check | Expected | Observed |
|---|---|---|
| Headings | Natural structure. | Awkward heading → editorial correction → new rule. |
| Applicability | Claim fits the actual scenario. | Over-broad advice → evidence check → narrower wording. |
Why text generation alone is not enough
A publishable page needs a reader outcome, sources, structure, visuals, links and rendered QA. Faster text generation alone can simply move the bottleneck into editorial cleanup.
How one material moves through the system
The pipeline is organized around page readiness.
Page contract
Reader, question, angle and acceptance criteria.
Sources and evidence
Support for claims and applicability limits.
Copy and visuals
AI-assisted preparation against the contract.
SEO / AEO / GEO
Search intent, direct-answer clarity and extractability.
Editorial acceptance
Human review and targeted repair.
CMS and rendered QA
Verify the actual published page.
What actually changed in production time
This compares the preparation of a five-article batch with visuals through the manual-placement stage, not raw text-generation speed.
Before: several weeks
A five-piece batch moved through research, writing, visuals, and review in the normal manual cycle.
Now: ~2 hours for five pieces
The system prepares research, structure, copy, and visuals; humans remain responsible for editorial review and site placement.
What the ~2 hours include
The old full cycle and the new preparation stage have different boundaries.
Before
A full batch could take several weeks.
Current system
Five articles with visuals are prepared in roughly two hours before manual placement.
Why speed did not require giving up quality
Rules were changed from real editorial feedback, not only from lab tests.
~95% first pass
Roughly 95% of outputs pass the first editorial review without a major return cycle.
Targeted repair
A local defect is repaired locally instead of giving the system permission to rewrite an already-correct page.
Humanization as real feedback
Remaining edits are usually about natural language and communication angle rather than rebuilding the whole article.
Rendered QA
The system checks not only the file but how the material appears after placement on the real page.
What ~95% first-pass acceptance means
It is the share of materials accepted without another editorial revision cycle. Repeated issues are used to improve the next iteration.
Publishing is a separate stage
Text and visual preparation are more automated than publishing. Placement and final page review remain a separate stage.
Context & measurement
- Context
- Operating content system · real website
- Scale / period
- ~400 materials · ~2 hours / 5 articles · ~95% first-pass editorial acceptance
What was measured
The case measures production volume, preparation time and editorial first-pass acceptance.
Content automation
Facing a similar challenge?
Tell me where SEO is blocked by product, engineering or data.