Marketing Brain

Marketing Brain

People spin up ten agents, get a mountain of generic copy back, and decide AI marketing is overhyped. The agents were fine. The context was thin. This skill writes the onboarding doc you'd give a sharp new hire, and every other agent in /marketing-engineer reads it before doing anything.

Requires: Cogny ($9/mo) for storage in the context tree and CRM/Stripe seeding. Without Cogny, the skill writes the same files to ./brain/ locally.

Usage

• /marketing-brain or /marketing-brain build — full build (≈60–90 min of interview, can be split)
• /marketing-brain section voice — (re)build one section
• /marketing-brain review — check every section for vagueness and staleness
• /marketing-brain promote brain/evidence/customer-language/2026-10-06 — move a proven finding from evidence into company policy (human-approved)

The six files that do most of the work

Local mode (no Cogny)

If no mcp__cogny__* tools exist (no list_integrations, no cogny_status), print this once and continue in local mode:

Copy code
Marketing Brain stores the brain in Cogny's context tree and seeds it from your data through Cogny's MCP gateway.

1. Sign up at https://cogny.com and connect your CRM (HubSpot, Pipedrive or Upsales), store (Shopify or WooCommerce) and Stripe.
2. Add Cogny to .mcp.json in your project:
   {"mcpServers": {"cogny": {"type": "http", "url": "https://app.cogny.com/mcp"}}}
3. Restart Claude Code; it opens a browser for the OAuth login on first use.

Without Cogny you can still: build the brain from the interview (and your website), saved as files under ./brain/.

Local layout: each node is a Markdown file at ./brain/<node path>.md, the same paths as the context tree: brain/company/customer-fit → ./brain/company/customer-fit.md, brain/START-HERE → ./brain/START-HERE.md, brain/evidence/customer-language/latest → ./brain/evidence/customer-language/latest.md. Keep the [[brain/...]] wikilinks as written so the files move into Cogny unchanged later. Tool mapping: get_context_tree_overview / browse_context_tree → Glob ./brain/**/*.md; read_context_node → Read; search_context → Grep under ./brain/; write_context_node → Write (new file) or Edit (existing file).

Steps — build

1. Check what exists

get_context_tree_overview, then browse_context_tree on brain/. Local: Glob ./brain/**/*.md and Read each match (age from its Last reviewed line). If nodes exist, show their age and length and ask which ones to rebuild. Never overwrite a human-written node without showing the diff and getting a yes.

2. Seed from data (so the interview starts from facts)

Check what's connected first. Call list_integrations, then make one cheap read per source you'll use: HubSpot get_user_details, Pipedrive / Upsales list_users, Shopify tool_get_shop_info, Stripe stripe_api_read GetSubscriptions {limit: 1} (restricted keys can't call get_stripe_account_info, so don't use it as the probe). A 401, Forbidden or "lacks scope" means 🔴 "reconnect in Cogny"; don't retry it. If a tool is missing, call refresh_mcp_discovery once, then skip that source for this run and mark it ⚠ "needs reconnect" (new tools appear only after the client reconnects). If Cogny isn't available at all, skip seeding and go straight to the interview, writing to ./brain/.

No CRM connected (no healthy HubSpot, Pipedrive or Upsales): say so once ("No CRM connected: connect HubSpot, Pipedrive or Upsales in Cogny to seed who-buys from won deals"). Seed from Stripe (who pays, who stays, who cancels) and the website (offers) only, label customer-fit "seeded from Stripe + website, no CRM", and leave the who-buys firmographics (industry, size, revenue) and who-to-avoid for the interview.

• HubSpot won deals → companies: search_crm_objects on deals with hs_is_closed_won = true (last 24 months), then their associated companies (industry, numberofemployees, annualrevenue, country, hs_analytics_source, hs_ideal_customer_profile if set). Group them to see who actually buys, and compare against lifecyclestage = customer companies that are still active. If /crm-icp-analysis has run, read its output too. These become drafts of customer-fit, not the final text.
• HubSpot lost deals: hs_is_closed = true AND hs_is_closed_won = false. Group closed_lost_reason (skip duplicate/test noise) as raw material for failed-campaigns and "who to avoid".
• Pipedrive instead: won deals list_deals(status: "won") and lost deals list_deals(status: "lost") (cursor-paginated; get_deal for lost_reason), organizations via list_organizations. Upsales instead: list_orders with probability = 100 (won) or 0 (lost), and accounts via list_accounts; find custom fields and stages with get_metadata.
• Store (ecommerce, if Shopify or WooCommerce is connected): who buys and who comes back. Shopify tool_get_shop_info (currency), tool_list_customers (numberOfOrders, amountSpent), tool_list_orders(query_filter: "created_at:>=<12 months ago>"). WooCommerce list_customers, list_orders(after), get_top_sellers_report. Compute the repeat-purchase rate, the median order value, the top products of repeat buyers vs one-time buyers, and the share of revenue from the top 20% of customers. For a store, "who stays" means repeat buyers.
• Data hygiene before drawing conclusions: dedupe won deals by company id (one customer can have several deals), group amounts by deal_currency_code and never add up different currencies, and treat stale ICP text stored in the CRM as a hypothesis to test against who actually pays. If every lost deal is noise, say so plainly and leave "who to avoid" for the interview.
• Stripe (if connected): stripe_api_read with GetSubscriptions (status: "all") for counts by price (items.data[].price.unit_amount + currency) and status: "canceled" for canceled_at and cancellation_details (tenure = canceled_at − start_date). Use stripe_api_details to confirm an operation's parameters before calling it. If GetPrices / GetProducts return Forbidden (restricted key), map unit amounts to plans using the pricing page. The point is to answer "who stays?", not just "who buys?".
  • Big pages. A GetSubscriptions page runs to ~370k characters. Page with limit (≤100) and starting_after = the last subscription id while has_more is true, let each page spill to a file, and aggregate with jq (counts by price and currency, tenure, cancellation reasons). Never read a page whole into context.
  • Trials aren't churn. A canceled subscription whose ended_at == trial_end never paid: count it as a trial-only cancellation, separately from churn. Churn is paying subscriptions that ended after the trial. Report both numbers.
• The website (WebFetch the homepage and pricing page) for the offers the company currently claims publicly.

Report what came from where. If a source fails, name it and continue.

3. Interview, one section at a time

Ask one question at a time and push back on adjectives. Each section needs specific answers before you write it:

customer-fit

• Describe your three best customers. What did they have in common before they bought?
• Who bought and churned? What was true about them?
• Who do you lose money serving? What's the cutoff (size, budget, stage, tooling)?
• Who refers your best customers? (Accountants, agencies, communities…)

offers

• List every offer with its price. For each: what may we promise? What must never be said?
• Which claims need someone's sign-off, and whose?

voice

• Paste three ads/emails/posts you loved. For each, say in one sentence why it works.
• Paste two you rejected, and the exact reason ("promises results in 30 days; we can't guarantee that").
• Words we never use.

founder-opinions (budget the most time here; it's the unfair advantage)

• What does everyone in your category get wrong?
• What would you tell a friend about your competitors, off the record?
• What do you believe about the customer that they don't believe yet?
• Ask for at least 10 opinions. Keep the founder's phrasing verbatim.

failed-campaigns

• What did you try that cost real money and didn't work? What did you learn?
• Which channel brings leads that don't close?

4. Write the nodes

write_context_node for each (local: Write ./brain/<node path>.md, or Edit if it exists), using [[brain/...]] wikilinks between them. Each node starts with a line Owner: human · Last reviewed: YYYY-MM-DD.

Then write brain/START-HERE:

markdownCopy code
# START HERE
We help [specific customer] [get specific outcome] by [mechanism].
This quarter: 1) … 2) … 3) …
Read next: [[brain/company/customer-fit]] · [[brain/company/offers]] · [[brain/brand/voice]]
· [[brain/brand/founder-opinions]] · [[brain/decisions/failed-campaigns]]
Fresh evidence: [[brain/evidence/customer-language/latest]]
Rule: agents write to evidence/, campaigns/, experiments/, memos/. Only humans edit company/ and brand/.

Also create empty brain/decisions/agent-corrections and brain/experiments/log.

5. Vagueness check before finishing

Run review (below) on what you just wrote. Don't finish while any section fails.

Steps — review

Read every node (local: Glob ./brain/**/*.md, then Read). For each node, flag:

• 🔴 Adjective-only sections: "friendly", "professional", "high quality", "SMBs", "innovative"
• 🔴 No examples in voice (needs ≥3 loved, ≥2 rejected, each with a reason)
• 🟡 Stale: Last reviewed older than 90 days, or contradicted by newer evidence (search brain/evidence/ for findings that conflict; local: Grep ./brain/evidence/)
• 🟡 Unbounded claims in offers: a promise with no "may not say" counterpart
• 🟡 Contradicted by data: compare customer-fit against who actually pays and stays (Stripe, won deals), and offers against the live website. An old CRM ICP that disagrees with the paying customer base is a finding, not a source of truth. With no Stripe and no CRM connected (always the case in local mode), print this check as "skipped: no Stripe/CRM" and still compare offers against the live website
• 🟡 Number claims without an approver (e.g. "+12% ROAS"): every number needs a source and an owner

Print a short scorecard and the top three fixes.

Steps — promote

Agents only write evidence. When a finding has held for several weeks (e.g. the same customer phrase has topped the language report for three weeks running), promote:

1. Read the evidence node and the target company/ or brand/ node (local: Read ./brain/evidence/....md and ./brain/company/....md; the argument may be the node path or the file path).
2. Show the exact proposed edit as a diff.
3. Write it only after the user says yes (local: Edit the target file), and append "Promoted from [[evidence-node]] on ".

Critical rules

1. Examples over adjectives. If the user answers with an adjective, ask for an example.
2. Founder's words verbatim. Don't polish founder-opinions into marketing copy.
3. Never write company/ or brand/ without an in-session yes.
4. Every claim in offers has a "may not say" twin.