This is the capstone. You've studied anatomy, frameworks, memory, security, and dialogue. Now you build — a real, working Content Assistant agent that you plan, construct, test against live inputs, measure, and document. By the end you'll have a prototype and the process to iterate it into production.
You're building ContentPilot — a Content Assistant agent that helps creators, marketers, and entrepreneurs generate post ideas, write summaries, draft LinkedIn posts, build Twitter/X threads, and craft content hooks. It receives a topic and audience, thinks about platform and tone, and produces structured, actionable output.
Every agent runs on a pipeline — a sequence of stages from receiving user input to delivering output. The order matters: if you try to call the LLM before validating input, bad data poisons the response. Arrange the 8 stages below in the correct execution order.
Follow the guided checklist for your platform of choice. Complete each step — the most important artifact you'll create is the System Prompt. Everything else configures the environment; the system prompt is the agent's brain.
Below is a fully functional version of ContentPilot, powered by Claude. This is what you just built. Test it like a real user would — try all 5 output modes with different topics and audiences. Your testing unlocks the performance data in the next section.
You've built and tested. Now look at the data. Below are performance metrics from the ContentPilot beta — 247 test runs across all 5 modes. One metric stands out as the highest-impact improvement opportunity. Your job: find it, diagnose it, and prescribe the fix.
A prototype without documentation is a dead end. The spec document is what makes ContentPilot reproducible, improvable, and handoffable. Fill out all 6 sections — this is the artifact you'd give an engineering team to implement and a QA team to test against.
Covering agent design, workflow planning, system prompts, evaluation, and iteration. 70% to pass.