Pack

AI Content Machine

The design document for running two brands through one pipeline without voice bleed.

An LLM for generation (local Ollama or any API), a scheduler, Postiz if you want multi-platform publishing, and Python or similar for the glue code you write from the spec.

// what it is

What this is: the full design document of a production content pipeline — every stage specified in implementable detail (data models, routing algorithm, compliance scoring weights, prompt structures), plus ready-to-edit YAML config templates and real worked examples.

What it is not: a pip-installable app. There is no pip install content-machine. You — or your coding agent, which this pack is very good input for — build the pipeline from the spec. The hard part of content automation is not the code, which is thin glue. It is knowing what to build: the routing rules, the compliance weights, the brand-separation discipline.

The pipeline: source intake → brand router → strategist → writer (2–3 drafts) → compliance → human approval → publish, with a memory engine feeding metrics back into future prompts. Compliance is a hard gate with five scored dimensions and specific weights — banned phrases 0.35, brand drift 0.40, generic phrases 0.15, length 0.20, repetition 0.30 — and a pass is 0.0, not "close enough".

// why it works

Separation enforced by structure, not prompts — per-brand configs, partitioned memory, cross-brand tone flags
Score, gate, approve — compliance weights are explicit, not vibes
Approval-first: automation handles the mechanical 80%, a human owns the editorial 20%
Anti-repetition via the last 10–20 hooks pulled from memory