Multi-Agent Content Pipeline: Brief → Research → Draft → QA

The problem: Marketing content workflows that depend on copy-pasting between LLM chat tabs are slow, inconsistent, and break the moment voice or style needs to change. I wanted a structured pipeline where each stage takes the previous stage's output as JSON and produces JSON for the next stage — so the same brief produces the same shape of output every time, and voice/style is a configuration knob, not a prompt rewrite.

Stage 1 — Brief Interpreter: converts a free-text content brief into a JSON blueprint (audience, goal, constraints, voice, expected structure).

Stage 2 — Research: adds structured research notes to the blueprint based on the brief's topic and scope.

Stage 3 — Outline: produces a structured outline (sections, key points, callouts) from the enriched blueprint.

Stage 4 — Draft: writes the draft against the outline.

Stage 5 — QA: checks alignment, completeness, and voice consistency against the original blueprint.

Every stage reads JSON, writes JSON. The handoff IS the contract — no prompt-chaining duct tape, no hidden state. Voice and style live in config/.env; prompts live in prompts/. Swapping a brand voice doesn't require touching the pipeline code.

The stack: Python virtual environment, OpenAI API (extensible to other providers), Streamlit UI for non-CLI users, JSON output to data/output/result.json. Version-controlled, prompts and config separated from execution.

What this demonstrates: Multi-agent orchestration with handoff contracts. Production prompt engineering with separation of concerns. Dual-mode deployment (UI for demos, CLI for batch). A reusable pattern for AI-augmented content ops that works across brand voices and content categories — built solo, reusable across clients.