A Claude Code marketing system has two distinct parts: tools provide the connections, storage, computing, and creative models; agents use those tools to complete specific jobs. Keeping those roles separate makes it easier to see whether a failed post came from a connection, a schedule, or an agent’s decision. It also prevents a working automation from being mistaken for proof that a campaign will grow an audience or generate sales.
Start with the numbers the system must track
Before assigning work to agents, choose key performance indicators (KPIs): the measures that define progress for the business. A personal brand might track follower growth, total views, and average views per post. An app business might prioritize website visits, downloads, subscriber churn, and recurring revenue. Those choices determine which data the dashboard needs and which agents should collect it.
Set output targets separately from outcome targets. Output targets specify how many videos, posts, or outreach messages the operation intends to produce; outcome targets measure what happens afterward. In one basketball app launch plan, the first-month goals were 100 downloads and 1 million organic views. Those were targets for configuring the work, not achieved results or a guarantee that the approach would establish product-market fit.
Give tools stable infrastructure roles
The system can be divided into five functions. A tool does not need to be an agent to be useful: its job may simply be to store a record, run a scheduled process, or connect to an account.
| Function | Tool’s job | Example in the setup |
|---|---|---|
| Social publishing | Connect accounts, schedule posts, and retrieve results through application programming interfaces (APIs), which let software exchange data | Connected platform accounts and publishing queues |
| Storage | Keep assets, post records, metrics, and workflow states | InnsForge with PostgreSQL, a database for structured records |
| Cloud automation | Run scheduled jobs when a local computer is off | Railway |
| Creative generation | Provide image, video, and audio models for asset production | Replicate and Higgsfield |
| Model and agent layer | Interpret instructions and coordinate task steps | Claude and Claude Code |
Cursor provides a development workspace, while GitHub stores and tracks changes to code and configuration. GitNexus can turn a large codebase into a visual map of its routes, automations, database tables, and external services. That map helps an operator inspect how parts connect; it does not replace testing those connections.
▲ Infrastructure for connected marketing tasks
A shared BRAND.MD file gives generation agents a reference for logos, typography, exact colors, asset filenames, and voice or avatar configurations. The file belongs alongside working code, account connections, and database access. Complete those dependencies before asking agents to run the broader system. The setup stores sensitive API keys in cloud environment variables rather than in local files.
Assign agents bounded jobs
The clearest division is between a human-initiated creative task and a scheduled operational task. In the editing workflow, a person supplies raw talking-head footage and can add notes about style or pacing. An agent then transcribes the footage, selects clips, adds animated captions and a call to action, writes social copy, and prepares a cover image. Human input starts the job; the agent handles the production steps.
The posting workflow starts later, with a finished clip. Its agent analyzes the material, prepares a title and caption, attaches the relevant media or resource link, and places the post in the next available slot across connected accounts. A cloud process executes the scheduled action. That distinction matters: an editing agent should not also have to decide whether the cloud worker ran, and a scheduler should not have to invent the brand’s visual identity.
▲ Editing, publishing, and measurement flow
Other agents fill narrower roles. A research agent gathers potential content topics from product releases and news feeds. Data-collection agents record daily app downloads, social metrics, and clicks on trackable links. A monitoring agent checks schedules, execution logs, and errors. A KPI checker compares the collected results with the targets set at the outset. Together, these jobs connect research, production, publishing, and review without treating one agent as the owner of every decision.
Check the workflow before judging its impact
An initial check should confirm that the dashboard has the right KPI fields, account connections work, the database receives records, and the cloud host runs scheduled jobs. A low-stakes test post can verify the handoffs before a larger publishing schedule begins. Execution logs then show whether a missed result reflects a failed job or a post that published but performed poorly.
The case includes an edited clip that received more than 11,000 views and 240 likes within a day. It also describes a short-form edit taking about 30 minutes and an operator running 20 Claude Code instances on separate clips concurrently. These examples show what occurred or was proposed within that operation. They do not establish the results, turnaround time, or labor savings another team will achieve.
The same distinction applies to the dashboard. Views, follower counts, downloads, link clicks, and revenue are different measurements. A calendar full of scheduled posts confirms planned output; it does not confirm audience interest. Link tracking can help connect a post to subsequent clicks, while app and business metrics show what happened further along the path.
Build one verified loop first
Choose a business KPI and a realistic content-output target. Connect the data source, document the brand rules, and test one editing task and one scheduled post. Then check the job logs and the resulting metrics before adding more agents or accounts. The tools should keep the operation running, the agents should carry out defined tasks, and the dashboard should show whether those tasks worked—and whether they contributed to the outcome that matters.