context management is one of the core skills for a marketing engineer when you give an agent a campaign to work on, it needs to understand the business, use what you've learned from previous campaigns and follow your taste in the work it creates your second brain gives it the company knowledge, but you also need to connect it to performance data and a brand book you can work in together tldr > company brain = helps the agent understand your business and the work you're doing > warehouse = lets the agent measure performance and learn from results > brand book = guides the agent's design choices using your style and feedback keep references. md in the project so whichever harness you're using can find the relevant files, query the data and open the designs with your feedback 1. connect your context sources with references. md your harness is the software that runs the agent, like cursor, grok bot, claude code or hermes references. md holds links and instructions for finding context. the company files stay in the brain, performance data in the warehouse and designs in the brand book before making a decision, the agent should ask for your relevant context and experience, including what you've already tried, what you learned and anything that hasn't been written down yet for a landing page + social posts, the agent needs to: > work with you to define the campaign brief, starting from your idea, goal and experience > read company knowledge, previous campaigns, decisions and lessons from the brain > query previous campaign performance in the warehouse > open the brand book and approved designs, including your feedback > bring the evidence back to you and work through decisions together save the agreed brief and the reasoning behind your decisions in the campaign brain as the work progresses 2. keep company knowledge and decisions in the brain use markdown in git, with github as the shared place to keep it. the agent reads the files relevant to the job your company brain holds context that should remain useful across campaigns > what you sell, who buys and why > positioning, supported claims and exclusions > voice-dna. md, hooks. md and approved work > dated decisions and permissions to edit, spend or publish your campaign brain holds the context for the work in progress > this campaign's goal, audience, offer and constraints > decisions you've already made and why > approved versions, rejected drafts and your edit notes keep performance data in the warehouse. the visual/ folder holds links to design. md and the live brand file 3. keep performance in the warehouse use postgres or another sql database. define what each row represents, such as one asset in one campaign on one date keep the facts you use to make marketing decisions > spend, impressions, clicks, leads and conversions > content results, like saves, replies or watch time > lead source and stage in your customer relationship management system > experiments, dates and what you stopped running keep definitions beside the numbers. define what counts as a lead or conversion, which reporting window to use and how much the data can lag use shared identifiers like campaign_id and asset_id to connect records across sources recurring jobs pull the data on a schedule. the warehouse connects the records and calculates totals with sql, the language used to query the database your agent queries the warehouse for the results it needs, instead of pulling bulk history from each platform into its context for the landing page job, it can query the last similar campaign by asset over the last 28 days, using the same conversion definition save your interpretation in the campaign brain with a link to the query. keep the performance history in the warehouse 4. keep your design rules and visual feedback in the brand book this needs both written rules and a live design file you can work in together design. md gives the agent instructions it can read before creating anything > typefaces, sizes and when to use them > colors, contrast and logo placement > layouts, safe zones and formats for each channel > do / don't examples with the reason beside them > what it can create and what it should only adapt > links to the paper or figma file and the pages to open write your brand guidelines in design. md and link each guideline to an example in paper or figma. include your feedback on why a design works or what needs to change paper and figma mcp give the agent access to the live design file. mcp is the connection that lets it use the design tool keep your brand book, moodboard and approved work there > add a note explaining what you like about each reference > keep approved ads, landing pages and social designs > keep rejected versions with your comments on the frames > leave feedback next to the work you're discussing design. md explains the requirements. the live file lets the agent inspect examples, create designs and work alongside you as you review them 5. use all three context sources, and save what decisions and changes for the landing page + social posts, the agent reads the relevant company files and campaign brief, queries previous performance, then reads design. md and opens the brand page with approved examples work through the direction together before drafting copy and frames in paper or figma. send the work to your sign-off queue for review after approval, save the copy and the reason for your edits in a campaign folder. keep the approved frames and design feedback in the brand file after launch, results go to data warehouse and your interpretation goes in the company / campaign brain the models are interchangeable, but your company knowledge, performance history and design judgment carry into the next campaign you´ll be working on