How does the approval flow work?

Topic proposals run at 2am CT on batch days. The proposer queries Nexus for site performance data, generates seven topics per site, and syncs them to the Nexus Publish UI. Joe marks approved topics in the UI. A polling writer picks up approved topics within thirty minutes and produces full drafts automatically.

What does the Cleanup Crew validator check?

The validator runs blocking and warning tiers. Blocking conditions include featured image present, three or more internal links, word count between 1,500 and 4,500, and no placeholder artifacts. Warnings cover H1 and H2 structure, SEO title and description lengths, atomic answer blocks, and FAQ blocks. Posts that pass are deployed. Posts that fail are held with a Slack alert.

How are hero images generated?

Hero images run on Bedrock Stable Image Ultra (stability.stable-image-ultra-v1:1 in us-west-2). The writer generates the image after the post draft and before internal-link injection. Sites that use SVG-based heroes are skipped, and a per-site aesthetic is applied so images stay on-brand rather than defaulting to a generic stock-photo look.

How does the pipeline avoid duplicate content?

The topic proposer checks each generated topic against the live site before syncing it for approval. A stale topics.json guard was added so the pipeline asserts freshness before proceeding. Internal-link injection runs inline during the write step to build cluster coverage across the network.

Content Pipeline

Abstract

The content pipeline is a fully automated, multi-site content production system that runs on a Monday/Thursday schedule. It proposes topics, writes AEO-optimized posts at 1,500+ words, generates hero images, validates output against a quality gate, and deploys — with one human approval step.

Problem

Agency content production at scale is expensive and inconsistent. Hiring writers per site, maintaining editorial calendars, and keeping pace with AEO requirements across multiple domains is operationally heavy.

Batch AI generation without editorial controls produces content that fails quality gates and doesn't compound organically.

Without routing and cost discipline, the economics of AI-generated content at network scale erode margin before the content earns anything.

Approach

Topic proposal

The pipeline runs on a scheduled worker that calls Claude for topic proposals and post generation. Topic proposals run at 2am CT on batch days: the proposer deletes any stale cache, queries Nexus for site performance data, generates seven topics per site, checks each topic against the live site to avoid duplicates, and syncs proposals to the Nexus Publish UI.

Writing and imagery

After a single human approval gate (Joe marks topics approved in the Publish UI), a polling writer picks up approved topics within 30 minutes. It writes each post, generates a hero image via Stable Image Ultra on Bedrock (skipping sites that use SVGs), commits the repo locally, runs internal-link injection, and calls the Cleanup Crew validator.

Quality gate

The validator checks each post for blocking conditions (featured image, ≥3 internal links, 1,500–4,500 word count, no placeholder artifacts) and warnings (H1 + H2 structure, SEO title and description lengths, atomic answer blocks, FAQ blocks). Posts that pass are pushed and deployed via GitHub Actions + SST.

Posts that fail are held with a Slack alert listing the specific issues.

Status

Running on seven sites on the Monday/Thursday schedule.
Cleanup Crew validator (blocking + warning tiers) shipped 2026-04-16.
Topics stale-guard fixed 2026-04-16 — propose-topics now asserts freshness before proceeding.
Hero images run on Bedrock Stable Image Ultra (`stability.stable-image-ultra-v1:1`, us-west-2).
Pipeline health monitor runs as a remote agent each batch morning, posting a pass/fail report to Slack.