Building a Multi-Agent AI for Company LinkedIn Pages - Part 7: Building the Hook Agent

A developer built a Hook Agent for a multi-agent AI system that generates LinkedIn posts, which produces five structured hook variations from a content brief. The agent uses a system prompt to enforce fixed ordering and JSON output, with retry logic for malformed responses. The project is part of an open-source repository on GitHub.

Our system can now classify a topic, conduct research, back it with examples, identify weak assumptions and unsupported claims, and structure everything into a content brief before handing it over to the writing agents. Before the writing agents generate a full post, we need an opening that's compelling enough to stop people from scrolling. That's why we built the Hook Agent. Unlike the previous agents, it doesn't consume the topic, research, examples or critique directly. Instead, it only takes the ContentBrief generated by the Brief Agent. Content Brief │ ▼ Hook Agent │ ▼ 5 Hook Variations Most of us aren't satisfied with the first hook we get. Instead of generating just one opening, I generate five different hook variations. The Draft Agent can later use any one of these to generate complete posts. One hook format doesn't work for all posts, so we use different types of hook variations like Bold Claim, Statistic, Confession, Question, One-liner. Before we start writing a single line of code, let's start with the system prompt. I explicitly instruct it to generate exactly five hooks, define the hook formats with examples, enforce a fixed ordering, and return valid JSON. SYSTEM PROMPT = """ Generate exactly 5 hooks. Formats: - Bold claim - Stat - Confession - Question - One-liner Return valid JSON. """ The production prompt is much longer, but I've shortened it here to highlight the core rules. So the JSON response returns hooks in the exact order which makes them easy to compare and lets downstream agents reference them deterministically. For example, hooks 0 is always the bold claim. php 0 - Bold Claim 1 - Stat 2 - Confession 3 - Question 4 - One-liner Instead of manually constructing the prompt, I convert the ContentBrief object into a dictionary using model dump and serialize it into a structured user message. brief dict = brief.model dump user message = "\n".join f"{k}: {v}" for k, v in brief dict.items Just like the Critic Agent, the Hook Agent retries generation up to three times if the model returns malformed JSON or fewer than five hooks. for attempt in range max retries : Before returning the response, I validate that exactly five hooks were generated. I'd rather return no hooks than fewer than five. The Draft Agent expects all five hook formats, so partial output is treated as a failure. if len parsed == 5: return parsed Before writing code, let's breakdown what happens in the function. It takes a ContentBrief object as input, converts it into a dictionary using model dump , builds the user message, calls the LLM, strips markdown fences, parses the JSON response, and validates that exactly five hooks were returned. php def generate hooks brief: ContentBrief, max retries = 3 - list str : brief dict = brief.model dump user message = "\n".join f"{key}:{value}" for key, value in brief dict.items for attempt in range max retries : raw response = call llm SYSTEM PROMPT, user message refined = strip json fences raw response try: parsed = json.loads refined if len parsed == 5: return parsed except json.JSONDecodeError: print f"Hook attempt {attempt + 1} failed, retrying..." return At this point, our system can understand a topic, gather supporting research, find real-world examples, critique them, structure everything into a content brief, and generate five structured hook variations before writing a single post. In the next article, we'll build the Draft Agent, which takes one of these hooks and turns it into complete LinkedIn post drafts for the user to choose from. Github Repo: https://github.com/Manav-N4/linkedin-agent linkedin-agent https://github.com/Manav-N4/linkedin-agent linkedin-agent