ivanhapaz/marketing-context-brain

Context Brain is a Claude skill for building the background knowledge an AI marketing agent needs before it starts producing real work. The idea behind it is that before automating a marketing task, a person should first give the AI enough context to make good marketing judgment calls, and prove that judgment works on something low stakes, rather than plugging an under-informed AI straight into a live workflow. Instead of handing a user blank templates to fill in, the skill does research and drafting itself, pulls in whatever data it can reach, and then asks the user to weigh in on the parts that require real marketing taste. Running the skill produces eleven connected markdown documents that work together: a master spine document that resolves conflicts between the others, market and category research, a competitor reference, evidence from customers and past campaigns, an ideal customer profile, a voice and tone guide, a design system translated into rules an AI can follow, an index of brand assets, a description of the operating workflow and approval steps, a lean instructions document, and a decision log that records each hard rule along with the mistake that led to it. The author suggests that a minimal but useful version can be made from just three of these documents: the spine, the ideal customer profile, and the voice guide, which together stop the output from sounding generic, while a full build with all eleven takes several hours of back and forth with Claude. The process itself is meant to work as a back and forth: the AI investigates and drafts, the human decides. When Claude's research disagrees with the user's instinct, it states its case with evidence once and then follows the user's final call rather than arguing further. Judgments the user hasn't yet confirmed stay marked as uncertain instead of quietly being treated as settled fact. The skill can be installed on Claude.ai as a custom skill, in Claude Code by placing the folder in the skills directory, or simply pasted into a Claude Project since everything is plain markdown text. It includes worked examples and a set of acceptance tests to check whether the resulting documents actually produce good output on a new task. It is licensed under the MIT License.