How to create a custom AI agent for a recurring marketing job

To create your own AI agent for a marketing job, you write down the job the way you'd brief a new hire, give the agent what it needs to do that job, and decide when it runs. You don't need to code or train a model. Today's agents run on general-purpose language models, and what makes one yours is the brief and the setup around it.

Every custom agent has the same parts:

• One job. A recurring task with one kind of output, such as "write the monthly customer newsletter from what we shipped and published that month."
• Instructions. The agent's role, the facts it works from, and the rules for doing the job well, with the most important decisions first.
• Knowledge. Your positioning, audience, brand voice, and examples of past work you liked.
• Tools. The apps and data sources the job reads from and writes to, such as your email product, your analytics, or a research source.
• A trigger. A schedule, such as the first Tuesday of every month, or an event, such as a new form submission.
• A delivery point and an approval rule. Where the finished work goes, and which steps need a person to say yes.

You can assemble those parts yourself with several products. A custom GPT in ChatGPT holds your instructions and uploaded files, and a Claude Project does the same for Claude. No-code agent builders such as Zapier Agents, Make, and n8n add triggers and connections to your apps, so the agent can run on a schedule and act in other tools.

Those products give you a capable agent, and you do the setup and upkeep. You write the instructions and test them. You upload the company context, and when your pricing or positioning changes, you update it in every agent that uses it. You connect each app, set each schedule, and decide what happens when a run produces something you don't like.

A good result is an agent you stop thinking about. It produces the same kind of work on time, sounds like your company, uses current facts, and gets better when you correct it. The hard part is scoping and briefing it well, because a vague brief produces generic work no matter how good the model is.

Kite does that scoping and setup for you. You describe the job in plain language, and Kite builds the agent, connects it to what your team already knows, and puts it to work on the schedule you want.

Start by telling Kite what the job is, in the Kite web app or in Slack:

"Create an agent that writes our monthly customer newsletter from the product updates and blog posts we published that month."

Kite is an AI marketer. It researches your business, builds a growth strategy, and ships the work that brings you new customers. By the time you ask for a custom agent, Kite has already researched your company and recorded your positioning, products, customers, brand voice, and competitors in your team's knowledge base. Your agent starts with all of that, so your request only needs to cover the job itself.

Kite needs three answers before it builds anything: what the agent's job is, what kind of task it will be handed and when, and what it should produce, including the result that counts as success. If your request already answers all three, Kite goes straight to building. If it doesn't, Kite asks the missing questions together in one message. For the newsletter, that might be who receives it, how long it should be, and whether you judge it by opens, clicks, or replies.

If the job needs one of your own accounts, such as the email product your newsletter goes out from, connect it through OAuth, which is a sign-in screen where you approve Kite's access. Kite's built-in research sources and tools need no account of your own.

Kite checks that the job is recurring and narrow enough for one agent before it builds anything. A custom agent is worth creating when the same kind of work comes back again and again and you want it done by a worker with a fixed role. For a one-time request, Kite simply does the work.

Kite proposes a tighter scope in three situations:

• The job and its output can't be stated in one sentence.
• One of your existing agents already covers the same job and output.
• The instructions would add nothing beyond the agent's one-line description.

A request like "an agent that handles all our content" usually splits into several agents, such as one for the newsletter, one for social posts, and one for blog briefs. Each one then has a clear standard to meet, and Kite can tell them apart when it hands out work.

Good fits for a custom agent include:

• Writing the monthly newsletter from what the company shipped and published.
• Researching a set group of competitors every week and reporting what changed.
• Turning each new blog post into posts for LinkedIn and X.
• Summarizing last week's website and campaign performance every Monday.

Kite gives the agent a short name that says what it does, such as newsletter-writer or competitor-research, and a one-line description of its job and output. Names use lowercase letters, numbers, and hyphens, and a name can't repeat one of Kite's built-in specialists, so a newsletter agent gets a name more specific than "writer."

The description carries more weight than the name. When work comes in, Kite reads the descriptions of your agents to decide which one should take it, so the description has to be specific enough to separate this agent from every other one on your team. "Writes the monthly customer newsletter from product updates and published posts" routes the right work. "Helps with content" doesn't.

Kite writes the instructions in three parts: the agent's role, the context it works from, and the rules for doing the job. These instructions are the agent's job description, and they shape everything it produces.

• The role is one sentence that sets the agent's responsibility, such as producing the monthly customer newsletter.
• The context covers the facts and inputs the agent works from: where it finds the month's product updates and posts, which tools it can use, and who reads its output.
• The instructions are the rules, ordered with the most important decisions first. For a newsletter, that means how to choose what goes in, how long each section runs, what the subject line should do, and what to leave out.

Kite writes the rules as positive directions that name what the agent should do. It leaves out anything your knowledge base already covers, such as your brand voice, so there's one place to update when that changes.

Kite chooses the skills the agent can use and keeps the list to what the job needs. A skill is a set of instructions for one kind of work, such as researching a competitor, planning an email campaign, or reading analytics. A newsletter agent might get the skills for email campaigns and copywriting, while a competitor agent gets competitor research and web research. A shorter list keeps the agent focused on its job.

Every custom agent can read and add to your team's knowledge base, and it can create images and PDFs and hand smaller pieces of work to other specialists. The knowledge base is what makes the agent sound like your company from its first run. It holds your positioning, products, pricing, customers, competitors, brand voice, and visual identity, along with findings from earlier campaigns. Because Kite keeps it current as it works, a pricing change recorded once reaches every agent that reads it.

When the job follows a specific procedure, Kite can also save that procedure as a custom skill and give it to the agent. If your newsletter always opens with one customer story, then lists three product updates, then ends with an event, that order belongs in a skill. The agent loads it every time it writes a newsletter.

Kite creates a workflow that starts the job on a schedule in your timezone, or whenever another system calls the workflow's webhook. A webhook is a private web address that starts the work when another tool sends it a message, such as a form tool reporting a new submission. The agent defines who does the job, and the workflow defines when it happens.

When the schedule fires, the run starts a fresh conversation with Kite. Kite reads your agents' descriptions, hands the job to the agent built for it, and receives the finished work back. Each run stands on its own, so the agent works from the knowledge base and the current month's material rather than from an old conversation.

You choose where results go when you set up the workflow. A team with Slack can have them posted in the thread where it asked, in a channel it names, or in its main channel. A team without Slack gets each result emailed from the team's Kite address to the person who set up the workflow, and replying to that email continues the conversation. Results can also stay in Kite, where each run appears as its own conversation in the Kite web app.

The agent's work follows the same approval rules as the rest of Kite. By default, Kite asks before it publishes a webpage, sends an email, or posts on social media, so a finished newsletter waits for you to approve it before it goes out. Approve it with the Approve button or a reply, in the Kite web app or in Slack.

Once you've seen enough issues to trust the agent, you can give Kite more autonomy, and the newsletter goes out on schedule without waiting for you. Work that stays inside your team, such as a weekly competitor report, needs no approval.

The agent improves through the same loop that improves the rest of Kite's work. When you correct or praise something it produced, Kite records that as a proposed rule. Once Kite's nightly review confirms the rule, it guides later work of the same kind, so a correction to one newsletter applies to the next one.

For a larger change, ask Kite to update the agent. It rewrites the instructions or changes the skill list on the same agent, so its name and its place in your schedules stay the same. When the job ends, ask Kite to remove the agent, and Kite can turn off or delete its workflow too.

When Kite finishes setting up, your team has an agent with a fixed role, a schedule that starts it, and a known place where its work arrives.