how to set up marketing agents that go out and get the context your campaigns need
when you're planning a campaign, your agents can research competitors, explore marketing campaigns and signup funnels, read customer conversations and follow what's happening on the channels you're using
give them the tools to access those sources and a process for turning what they find into a campaign knowledge you can use for planning, execution and experiments
1. start with what you need to understand
give the agent your campaign idea and goal / objective, then have it read the relevant company knowledge and previous campaigns
your company brain holds reusable knowledge about the business. the campaign brain holds the context, decisions and working files for this campaign
have the agent ask about your experience before planning the research. you may have already tested an approach or learned something that never made it into a file
work together to define the questions the research should answer:
> which alternatives are customers comparing us with?
> what objections should we address?
> how are similar offers being presented?
> which formats could suit this audience and channel?
> what do we need to understand before choosing an angle?
include the audience, market, language and how recent the information needs to be
2. connect the tools and write the research skill
your agent needs tools for finding sources, reading pages, inspecting visual experiences and retrieving data
connect web search, a page reader, browser or computer access, and relevant data apis through your agent environment
test each connection with a small task and check that the agent can retrieve the content and keep its source
then write a research skill explaining:
> how to turn the brief into research questions
> which sources and tools to use
> what evidence to collect
> where to save the material
> when to ask you for direction
> when it has enough evidence to return
set request and spending limits, with approval required for purchases, account changes or form submissions
3. research marketing campaigns and signup funnels
give your codex or grokbot setup browser access to explore how other companies attract users and guide them toward signing up
have it follow the public journey from an ad, social post or search result through the landing page and signup flow
ask it to collect:
> the campaign hook, creative and offer
> how the landing page follows up on the message that brought someone there
> how pricing, proof and customer objections are presented
> the calls to action and steps leading to signup
> what information someone is asked to provide
> links and visual references for ideas you could test
review the examples together and connect useful findings to the relevant pages in your campaign brain
4. research customer language and objections
send the agent to relevant trustpilot reviews, reddit communities, category-specific review sites, app stores and comment sections
give it questions to investigate, like why customers switch products, what disappoints them and what they wish they had understood before buying
keep exact phrases with their sources and dates. have the agent group recurring themes while preserving disagreements and differences between customer types
review those findings against your experience, customer calls and support tickets before using them to shape the campaign
5. connect search data and investigate customer queries
use an seo data api to retrieve keyword data and search results for the market you're targeting
research agents can investigate different groups of queries, such as product comparisons, customer problems and buying questions
record the query, location, language and collection date alongside the findings
you can also have different search-enabled models investigate those questions and collect the sources they find.
when testing how ai products answer customer questions, save the answers and cited sources separately from keyword demand data
the lead agent reviews the findings across these tasks and identifies what needs further investigation
6. check trends daily and weekly
give the research agent a watchlist based on your audience, category, competitors and active campaigns
connect the campaign to a grok bot or grok cli with x search enabled to investigate relevant keywords, accounts and conversations. add browser access where you want it to inspect timelines and visual posts
for daily checks, have it look for developments that could affect work already underway:
> conversations involving your audience
> competitor launches or offer changes
> emerging hooks and creative formats
> questions your campaign could help answer
> relevant events and creator activity
use the weekly review to compare findings over time and decide which deserve an experiment
each proposed opportunity should include the sources, why it fits your audience and what you could test
7. ingest the research into your campaign brain
keep original pages, screenshots, transcripts and data in research/raw/, with their sources and collection dates
have the agent review that material and ingest useful findings into wiki/, the campaign brain
this follows the karpathy method, his llm wiki approach, applied to a campaign: the agent reads raw sources and builds a wiki of linked markdown pages that it maintains as new information comes in
as it learns, it creates or updates linked markdown pages about customers, competitors, offers, objections, campaign examples and experiments
those connections form a knowledge graph the agents can navigate
a customer objection might link to a competitor's offer, a landing-page decision and the experiment you're running to test it. each page also links back to the evidence behind it
have the agent check existing pages before creating new ones, update findings when evidence changes and record conflicting evidence or unanswered questions alongside the relevant topic
the campaign brain also holds the brief, decisions, ideas and working documents, with links to designs and performance data kept in other tools
8. review the findings and use them in experiments
build evals into the research skill, meaning checks against the requirements you agreed
check whether sources support the claims, quotes are exact, dates fit the brief and findings apply to your audience
then work through the recommendations together. your experience may change how the research should be used, and that reasoning belongs in the campaign brain
for each experiment, record:
> the finding that prompted it
> what you expect to happen and why
> the asset or campaign element you will change
> how you will evaluate the result
> what needs your sign-off before execution
agents creating copy, designs and other assets can follow the wiki's links to the relevant findings, evidence and agreed decisions
Show more
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
Show more