fm.

How I run growth marketing on Claude Code and Codex

Two coding agents, my own tools, and a knowledge base they read before every task. What the setup looks like, what it has moved, and how to start.

Key takeaways

  • Two coding agents, Claude Code and Codex, run my reporting, research, ad operations, tracking fixes and builds. I make the calls and approve anything that spends money.
  • The part most people skip is memory. Every session is captured into a knowledge base the next session reads first.
  • Guardrails come before automation: approvals on spend, previews and rollbacks on writes.
  • It shows up in results: cost per SQO down 74% at Pipe17, a target list grown from 441 to 2,795 accounts, and about $24K a year in software replaced.

01The short version

Most marketers use AI to draft copy. I use two coding agents as the operating system for the whole job. They pull the reports, research accounts, mine search terms, fix tracking, build pages and write the first draft of almost everything. I make the calls that need judgment and approve anything that spends money.

The scale of it: 95.9 billion tokens through Codex alone, a 104-day daily streak, and one autonomous task that ran for 15 hours and 37 minutes.

02Two agents, one set of rules

Codex is my primary for long-running and scheduled work. Claude Code handles interactive deep work, browser tasks and research that fans out across many sources. Both read the same operating rules, a single AGENTS.md that says which tool to reach for, what needs my approval and how a session should close out. They also read the same map of my workspace, so they behave like one team instead of two contractors who never met.

03MCP first, my own tools second

Agents reach real systems through MCP connectors: GA4, Semrush, HubSpot, LinkedIn Ads, Google Drive, Docs and Slides, Figma, Chrome, my inboxes and my own knowledge-base server. Where a connector can't do the job, they fall back to command-line tools I wrote for exactly that:

  • gads for Google Ads reporting, search-term mining, negatives, ad rewrites and full audits.
  • hs for HubSpot contacts, deals, attribution and imports.
  • utm for one UTM standard, plus audits that catch a campaign splintering into five rows.
  • enrich for work-email enrichment as a single command, with a webhook receiver on my server.
  • recall for searching every past agent session by keyword.

Each tool is wrapped as a skill the agents can discover. Together they replaced about $24K a year in software.

04Parallel agents with written briefs

Research and production fan out to subagents, and each one gets a written brief: the goal, what's already ruled out, which files to read and exactly what to return. At Pipe17 a 12-agent pipeline researched accounts, scored fit and found the champion, and grew the target list from 441 to 2,795 accounts.

Most of the quality comes from the brief. A vague brief gets you twelve confident, wrong answers.

A brief looks like this:

Example brief

  • Goal: score 400 accounts for fit and find the likely champion at each one.
  • Already ruled out: companies under 20 people, agencies and current customers.
  • Read first: the ICP definition, the account list and the last two win and loss notes.
  • Return: one row per account with a fit score, champion name and title, the source URL and a confidence level.
  • Stop and report if: more than 10% of rows have no source.

05The knowledge base, the part most people skip

Agents forget. Claude Code deletes transcripts after 30 days by default. So every session reads from a knowledge base when it starts and writes to it when it ends.

LayerWhat it isScale
SourcesRaw agent transcripts, meeting notes, email and project files.7,926 transcripts, 329 meeting notes
CaptureEvery meaningful session ends with a structured note: summary, decisions, open loops and the next step.5,163 session notes
Canonical factsThe facts agents must never guess, such as account IDs, budgets and dates, ranked above everything else.184 facts
RetrievalHybrid keyword and embedding search in SQLite, exposed to both agents through my own MCP server.One index
InjectionHooks that load the right topic dossier and recent activity into a session automatically.14 dossiers
QualityEvals for retrieval and cited answers, health checks and secret redaction.55 scripts

The payoff: a session on a two-month-old project starts out already knowing the account IDs, the past experiments and what failed.

06Always on, with guardrails

A small server runs scheduled jobs, webhook receivers and alerts, so nothing depends on my laptop being open. 26 agent jobs run on a schedule, including ad quality monitors, a monthly LLM-referral report, weekly SEO and AEO runs and a weekly health check on the knowledge base. Alerts lead with the verdict and only fire when something needs a decision.

Honestly, the guardrails took more work than the automation. Spend and account changes need an explicit go. Writes go through previews and rollbacks. Credentials are redacted before they reach a model. Every sender has daily caps, bounce stops and a kill switch.

A normal week

WhenWhat runs
Every morningVerdict-first alerts on spend, pacing, conversion tracking and anything that broke overnight.
MondayA paid audit: search terms, negatives, ad strength and budget shifts, staged for approval.
Mid-weekThe AEO run: the prompt panel, citation changes and the next pages to ship.
FridayThe scorecard: spend against qualified outcomes, written up for leadership.
MonthlyThe AI referral report and a health check on the knowledge base.

What I learned the hard way

  • Agents will confidently guess an account ID if you let them. The canonical facts layer exists so they never have to.
  • Every write needs a preview and a rollback, because a fast mistake is still a mistake.
  • The narrower the brief, the better the work. Scope creep in a prompt shows up as noise in the output.

07What it has moved

WhereWhat the agents didResult
Pipe17A 12-agent account research pipeline.Target list from 441 to 2,795 accounts. Cost per SQO down 74%.
Pipe17Attribution rebuilt on what buyers said, plus monthly LLM-referral tracking.32% of new-business wins traced to AI assistants.
Pipe17Paid operations through my Google Ads and HubSpot tools.$66.5K in paid influenced $1.3M in pipeline in a quarter. About $24K a year in software replaced.
CommasAEO analysis of 10,560 ChatGPT and AI Overview answers.221 page opportunities and a 35-topic roadmap.
CommasThe rebrand site, funnels and an AI creative engine.100+ launch creatives behind about $1M a month in paid.

08How I'd start, if you run a marketing team

  1. Write the rules first: one AGENTS.md with tool choices, approval limits and how a session ends.
  2. Wrap one system you touch every day, your ad account or your CRM, in a small tool the agent can call.
  3. Capture every session into notes the next session reads.
  4. Put guardrails on anything that spends money before you automate it.
  5. Judge it like any channel. Did the cost of a qualified outcome go down?

FAQQuestions people ask

What does an AI-native growth marketer do differently?
The work itself runs through agents. Reporting, research, ad operations, tracking fixes and page builds are done by coding agents connected to real systems, and the marketer spends their time on decisions and approvals.
Do the agents change ad spend on their own?
No. Spend and account changes need my explicit approval. Agents stage the change, show a preview and wait for a go.
Which tools does Farhan Munshi use?
Claude Code and Codex, MCP connectors for GA4, HubSpot, Semrush, LinkedIn Ads, Google Drive, Figma and Chrome, his own command-line tools (gads, hs, utm, enrich, recall) and a knowledge base exposed through his own MCP server.