Case study
One source in. A month of publishable content out.
A content engine that turns one source document into platform-native posts, where every line passes a codified voice audit before it ships.
Built in Python on Claude agents. Clean-room version public on GitHub. By Amber Lan, AI marketing engineer. Last updated 20 August 2026.
Back to selected work
The problem.
One person cannot hand-write platform-native content at volume, and raw model output reads generated.
Quality control at scale
The tells are recognisable, the voice drifts, and the volume meant to build trust starts eroding it. The problem is not generation.
Voice as law, not vibes
Rules a machine can check, written down. Zero tolerance for recognisable AI patterns. Each platform gets its native shape, not one post cross-pasted five ways. The pipeline runs solo inside a bigger operating week.
What was built.
Source document in, a packet of platform-shaped posts out, ready for a human yes or no.
Plan
The engine plans a skeleton per artefact and generates platform-native drafts.
Audit
Every line runs through a voice gate: banned words and patterns, sentence rhythm, claim discipline, and the brand's specific rules.
Rewrite
Artefacts that fail are rewritten against the specific violations, not regenerated blind.
Key decisions.
Four calls that moved the engineering to where it matters.
The audit is the product
Generation is cheap. The gate makes the output shippable, so most of the engineering lives there.
Voice as law, not prompt
Style guidance in a prompt decays. Codified rules checked line by line do not.
Human approves, machine produces
The operator's job collapsed from writing to judging, the only part that needed a human.
Packets, not posts
One source becomes a coordinated set across platforms, so the calendar fills in one pass.
The result.
In production, one source document reliably becomes a month of on-brand, platform-native content.
In production
The operator's time goes to judgment instead of drafting. Packaged as a reusable skill: a new brand means codifying its voice rules, not rebuilding the pipeline.
Evidence
The clean-room version is public: read the code on GitHub. The distribution half is public too: Social Media OS.
What I learned
Generated content fails at the last metre, in the lines a reader flags as machine-made. Fixing that at the gate, line by line, separates a content engine from a content faucet.
Want a system like this pointed at your business?
Tell me what you do and where the pipeline stalls. You get a written read within the first week.