How We Run an Entire Agency Inside Claude Code

Most people meet Claude Code as a coding assistant — a thing developers use to write functions faster. That's the smallest version of what it is. For us, Claude Code is the operating layer the whole agency runs on. It's where ad creative gets generated, where the reporting dashboard gets built and deployed, where client automations get written, where research gets done, and where the day's work gets organized.

EBCD is three people running thirteen client accounts. That ratio doesn't work with a traditional toolchain. It works because a huge share of the execution lives inside an AI environment that has access to our files, our scripts, our APIs, and our context — and can actually do the work, not just describe it. Here's the honest, unvarnished tour of how that works in practice.

The shift: from "ask a chatbot" to "give an operator a workspace"

The mental model that unlocks everything: Claude Code isn't a chatbot you ask questions. It's an operator you hand a workspace. It can read the files in that workspace, run commands, call APIs, write and execute scripts, edit code, and deploy. The difference between "answer my question" and "go do this task in my actual environment" is the entire difference between a toy and a tool.

So the first thing we built wasn't a clever prompt. It was a structured workspace: a single file that explains who we are, what we're working on, where everything lives, and how we operate. Every session starts by loading that context. From then on, the AI isn't guessing about our business — it knows the client roster, the brand colors, the deploy commands, the API quirks, the team. Context is the multiplier. A generic AI gives generic output; an AI that knows your operation gives your output.

1. Ad creative — the production pipeline

The agency's creative promise is blunt: clients get the testing volume normally reserved for eight-figure e-commerce brands, and they don't write a single ad. That's only possible because creative production is automated end to end.

• Ad images. We generate performance ad creatives — multiple awareness levels, bilingual English and Spanish, on-brand — through an image pipeline driven from the workspace. What used to be a designer's week is a batch run that produces dozens of variants in an afternoon.
• UGC video ads. AI talking-head clips get converted into Meta-ready vertical ads with conversion captions, motion graphics at the key beats, and a CTA endcard — through a scripted edit pipeline, not a human editor dragging clips around for hours.
• B-roll and hooks. Hooks tested in both languages, b-roll generated on demand, the winners scaled. The operator directs; the pipeline produces.

The point isn't "AI makes ads." It's that the entire creative function — image, video, copy, in two languages — runs as a repeatable system one operator triggers, instead of a department you staff.

2. The reporting dashboard — built and run from the workspace

Every client gets clean weekly numbers: cost per lead, appointments booked, appointments shown, cost per sale. No fluff. The thing that produces those numbers is a custom dashboard that joins data across every client's ad accounts, CRM pipelines, and business profiles into one view.

We built it inside Claude Code — a single application, deployed to a host, serving both the data API and the views. When we need a new metric, a new alert, a new client added, that's a conversation in the workspace, not a ticket to a developer. The dashboard also drives the alerts: if a client's cost per lead drifts above target for a few days running, the team gets pinged before the client ever notices. By the time anyone asks "how are we doing," we already know, because the system has been watching.

3. Client automations — written, not waited on

Behind every account is a follow-up engine: lead qualification, calendar booking, SMS and email reminders, no-show recovery. Every lead that comes in gets nurtured automatically so reps walk into warm appointments instead of cold calls.

Those automations — the webhooks, the field mappings, the message sequences, the integration glue between the ad platform and the CRM — get built and debugged in the workspace. When a platform's API does something unexpected (and they all do), the fix is written and tested right there, with the quirk documented so we never re-learn it. The institutional knowledge compounds in files instead of evaporating when someone's out sick.

4. Research and content — at volume

Competitor ad research, market scraping, blog posts, landing-page copy, social content — all of it runs through the same environment. We can scrape what competitors are actually running, analyze it, and turn the findings into a content batch or an ad angle in the same sitting. The blog you're reading is produced this way: drafted, formatted into the site's template, wired into the index and sitemap, and deployed — as one continuous flow.

5. The operating rhythm — context that persists

The unsexy secret is the discipline around it. A few things make the difference between "AI helps sometimes" and "the agency runs on this":

• A living context file. One source of truth the AI loads every session — current work, priorities, where things live. We update it at the end of each session so the next one starts oriented instead of cold.
• Memory across sessions. The hard-won details — API gotchas, account IDs, the deploy command that only works from the right directory — get written down once and recalled forever. We stop paying the "re-explain it" tax.
• Commands for repeated work. The things we do every week — prime the session, end the session, run the research mine, generate a report — are saved workflows, not improvised each time.
• Honest scope. The AI does the production and the plumbing. The humans do the judgment: which client, which strategy, which creative bet, which number matters this week. The leverage comes from putting the machine on the volume and the person on the decision.

What this actually buys us

Concretely: three people comfortably run thirteen accounts with capacity for more, without a hiring cycle. Creative ships at a volume that would need a small department. Reporting that used to eat a half-day per client per month is real-time. New clients onboard in about a week because the system is templated instead of rebuilt each time. And the founder can step away for a stretch without the business stalling — because the operation lives in systems, not in one person's head.

None of that is magic, and I want to be careful not to oversell it. Claude Code doesn't run the agency by itself; it amplifies operators who know what they want. A bad brief still produces bad output, faster. The compounding only happens because we invested the first months building the context, the scripts, and the discipline around the tool. That upfront work is the part most people skip — and it's the entire reason the leverage shows up.

How to start, if you want to run your business this way

1. Write your context file first. Before any clever automation, write down who you are, what you do, where everything lives, and how you operate. This single document is what turns generic AI output into your output.
2. Automate one production function. Pick the thing you produce most — creative, reports, content — and move it into the workspace as a repeatable flow.
3. Build your dashboard or your follow-up engine next. Whichever is costing you more attention. Let the system watch the numbers so you don't have to.
4. Capture every gotcha in memory. The first time you solve a weird problem, write it down. The second time, you'll thank yourself.
5. Keep the human on judgment. Don't try to automate the decisions. Automate the production around them.

The agencies and businesses that win the next few years won't be the ones with the best prompts. They'll be the ones who treated AI as an operating layer and did the boring work of building the context and systems around it. The tools are here. The advantage goes to whoever actually rebuilds their operation on top of them.