Why We Automated 3 Drafts/Day but Refused to Automate Publishing

Problem Statement#

The MARIA OS team needed a mechanism to produce daily technical blog drafts that reflect recent code changes, while preserving editorial control over final publication. The goal was to increase content velocity without compromising credibility or introducing unvetted titles.

Options Considered#

1. Full automation: generate draft and automatically push to the public site after passing minimal syntax checks. 2. Semi‑automation (chosen): generate draft artifacts (.md and .blog.json) with metadata, enforce a title quality gate, and require human review before publishing. 3. Manual process only: continue writing drafts by hand using existing templates.

Decision#

The team adopted the semi‑automated approach (Option 2). Draft generation is fully automated; publication remains a manual step performed by an editor.

Rationale#

We rejected the obvious full‑automation approach because it would eliminate the human editorial checkpoint that safeguards title relevance and overall credibility. A skip policy was introduced as a trust mechanism, allowing editors to bypass draft creation on days when changes are trivial. The title gate prevents weak or misleading titles from reaching publication even if the generation pipeline succeeds. Reproducibility artifacts (.md + .blog.json) were retained despite increasing file count, because they provide traceability between repository changes and generated content.

Trade‑offs

• Control vs. speed: Manual publishing adds latency (estimated 15–30 minutes per draft) but preserves editorial oversight.
• File system overhead: Storing paired artifacts raises the number of files in the repo, modestly increasing diff size and storage cost.
• Complexity: Implementing a skip policy and title gate introduces additional configuration and testing effort compared to a single‑step pipeline.
• Risk exposure: The system relies on repository change evidence; external context (CI results, production telemetry) is not considered, which may lead to drafts that miss broader impact signals.

Known Risks

• Human bottleneck could become a point of delay if editorial resources are limited.
• Title quality gate thresholds may be set too conservatively, causing excessive rejections and reducing draft throughput.
• The reproducibility artifacts increase the chance of merge conflicts in high‑frequency change environments.

This concludes today’s record of self-evolution. The interpretation of these observations is left to the reader.