How to Automate Your Marketing Agency With Claude: The Full System

Most agency owners who try AI stop after one good pitch email. It landed a call, they felt clever for a day, and by the following week they're back to writing client reports, churn-risk conversations, and case studies from scratch, one at a time, usually squeezed in after billable work. The prompt was never the problem. The problem is that nobody turned "a message that worked once" into a system that runs every week without being rebuilt from memory each time.

This guide is that system, built specifically for Claude, not for a generic AI chat window. It covers the ten pieces of writing a solo or small agency repeats every week, a business context file that makes every output sound like your agency instead of a generic template, a fixed weekly cadence for when each piece runs, a disclosure and honesty section most AI guides skip entirely, and what to do when Claude gets it wrong. By the end you'll have a working system, not just a list of prompts to lose in a chat history.

Everything here works in Claude.ai (a Project) or Claude Code (a CLAUDE.md file), and the underlying prompts are also in the Marketing Agency prompt pack as a standalone reference if you just want to copy one at a time.

What you're building

Two pieces, working together.

The first is a business context file: a short document that tells Claude who you are, so it stops writing like a generic marketing bot and starts writing like your agency. Without this, every pitch and every report sounds the same regardless of which agency sent it. With it, Claude knows your specialty, your tone, and the honesty line it should never cross when the numbers aren't as strong as you'd like.

The second is a weekly cadence: the same ten prompts from the prompt pack, placed on fixed days instead of used whenever you happen to remember them. A prompt you use once is a trick. A prompt on a fixed day is a system.

Quarterly business reviews and creative briefs aren't on the calendar. They run the week a review is actually due or a new campaign is actually starting, while the details are specific enough to be useful.

What you need

• A Claude account. Either works: Claude.ai with a Project (simpler to start), or Claude Code with a CLAUDE.md file in a folder (more setup, but it remembers everything across sessions without you re-pasting context, and can eventually reference your actual campaign data and client history directly).
• Ten to fifteen minutes to write your business context file once. This is the only setup step. Everything after this is copy, paste, review, send.
• Nothing else. No CRM integration, no API keys, no code. If you later want Claude to reference your reporting dashboards directly, Claude Code can do that too, but it's not required to get value from this system.

Step 1: Write your business context file

This is the step that separates a system from a party trick. Paste this into a Claude Project's custom instructions, or save it as agency-context.md if you're working in Claude Code.

I run a marketing agency. Here's my context:

Agency: [your agency name, or "independent" if solo]
Specialty: [what you actually do, e.g. "paid social for
  e-commerce" or "SEO and content for B2B SaaS" or "full-service,
  small local businesses"]
Client base: [the kind of client you mostly serve, e.g. "5-10
  person e-commerce brands" or "local service businesses under
  $2M revenue"]
Tone: [pick one and be specific: confident and direct / warm and
  consultative / sharp and a little blunt]
Things I never say: [e.g. "synergy", "10x your growth", any
  specific results number I can't back up with real client data]
Honesty rule: Never state a specific result (leads, revenue, ranking
  position, ROI) as guaranteed or typical unless I've given you a
  real, factual number from an actual completed campaign. When
  drafting a proposal or pitch, describe the approach and what we'll
  measure, not a promised outcome. Never fabricate a testimonial,
  a client name, or a case study number that wasn't given to you
  directly.

Fill in the brackets with real details, not placeholders. "Confident and direct" produces different output than "warm and consultative," and Claude will actually follow whichever one you pick, so it's worth being specific rather than generic. The honesty rule isn't optional. More on why in a few sections.

If you're using Claude Code, put this in a CLAUDE.md at the root of a folder you work from. Claude Code reads that file automatically at the start of every session in that folder, so you never re-paste it. If you're using Claude.ai, paste it once into a Project's custom instructions and every conversation inside that Project inherits it.

Step 2: Load the ten prompts

The prompt pack has all ten, each with a copy button. In Claude Code, the cleanest setup is a prompts.md file in the same folder as your CLAUDE.md, so both are always available together. In Claude.ai, add them as Project knowledge, or just keep the pack open in a tab and paste from there.

Either way, you don't need to memorize them. You need to know which one runs on which day, which is the next section.

Setting this up in Claude.ai versus Claude Code

Both work. The difference is where the context lives and how much it does on its own.

Claude.ai (a Project). Create a Project, name it something like "Agency Assistant." Paste your business context file from Step 1 into the Project's custom instructions. Add the ten prompts as Project knowledge files, or just keep the prompt pack page open in a browser tab. Every new conversation started inside that Project automatically has your context loaded, no re-pasting. This is the right starting point if you've never used Claude for work before, since there's nothing to install and no files to manage.

Claude Code. Create a folder on your computer, something like agency-ai/. Save your business context file as CLAUDE.md in that folder, Claude Code reads this automatically at the start of every session there. Save the ten prompts as prompts.md in the same folder. The real advantage over Claude.ai shows up once you want Claude working with actual files: drop your weekly reporting exports, a running client-status log, or past case studies into that same folder, and Claude Code can read and reference them directly instead of you copying numbers in by hand each time. It's more setup for less daily typing, which is worth it once this system is a genuine weekly habit rather than something you're still testing.

Start with Claude.ai if you're not sure yet. Move to Claude Code once you're running this every week and the copy-paste is the part that's starting to feel repetitive.

Step 3: Run the weekly cadence

Here's each day with a real worked example, not just the prompt in the abstract.

Monday: new business

Any qualified prospect worth pitching gets outreach first, before the week's client work eats the whole calendar. Writing this Monday morning, not "whenever there's a gap," is what actually keeps the pipeline moving instead of going quiet for a month at a time.

Worked example. A prospect you want to pitch:

Prospect: Coastal Supply Co, e-commerce outdoor gear brand. Noticed
their Instagram ads are running the same three creatives for at
least two months, engagement on the posts is visibly low relative to
their follower count. They just launched a new product line
(camping cookware) with no dedicated ad creative for it yet, based
on their feed.

The prompt (from the pack):

I want to pitch on . Here's what I know about their business and something specific I noticed: . Write a short outreach message (under 120 words) that references that specific observation, not a generic "we help businesses grow" opener. State the one gap plainly, and end with a low-pressure next step, not "let's hop on a call to discuss synergies."

A reasonable Claude output, with your business context file loaded:

Noticed Coastal Supply's Instagram ads have been running the same three creatives for a couple months now, engagement looks like it's cooled off relative to your follower count. Also saw the new camping cookware line launched with no dedicated ad creative yet, based on your feed. That's usually the easiest kind of win, a fresh product with an audience that already exists but no creative built specifically for it. If it's useful, I could put together two or three concept directions for the cookware line, no charge, just to show you what I'd actually do. No pressure either way.

Notice what it didn't do: no "synergies," no invented growth number, no vague "we help brands like yours scale." It used the two real, specific observations, stale creative and an unadvertised new product, and offered something concrete and low-risk instead of asking for a call up front.

Tuesday: proposals

Anyone who had a real discovery conversation gets an actual proposal, not a templated one-pager with the client's name swapped in. This is the highest-leverage prompt in the pack, because a proposal that reflects what the prospect actually said in the call closes at a different rate than one that reads like it was written before the call happened.

Worked example. Notes from a discovery call:

Prospect: Coastal Supply Co. Goal: sell through the new cookware
line before the fall camping season ends, roughly 10 weeks out.
Budget range given: $4k-6k/month ad spend, open to agency fee on
top. Explicitly said they don't want "influencer stuff," tried it
before and didn't like managing it. Currently running ads
themselves with no strategy, admitted it.

A reasonable Claude output:

Given the 10-week window before fall camping season ends, the recommended scope is a focused paid social push on the cookware line specifically, not a full-account overhaul, there isn't time for that and it's not what the goal calls for. Included: campaign strategy, creative direction (2-3 concepts to start), weekly optimization, and a monthly report. Not included: influencer management, per your note that it's not a fit right now, and organic content, which can be scoped separately later if useful. Recommended ad spend stays in your stated $4k-6k/month range, with the agency fee structured as . We'll track sell-through on the cookware line specifically as the core metric, alongside standard paid social benchmarks, rather than promising a specific revenue number this early, we'll have a much better read after the first two weeks of real data.

Notice it excluded influencer work explicitly, because the prospect said so directly, instead of pitching the agency's full standard package regardless of fit. It also didn't promise a revenue number, describing the approach and what gets measured instead, which matters enough that it gets its own section below.

Wednesday: client reporting

One report per client, sent the same day every cycle. The content changes monthly, the day doesn't. A report sent whenever you get around to it reads as an afterthought; the same report sent the first Wednesday of every month reads as a service the client can plan around.

Worked example. This month's raw data for a client:

Client: Coastal Supply Co, paid social, October.
Spend: $5,200
Results: 340 conversions (up from 265 in September), cost per
  conversion $15.30 (down from $18.10)
Note: cookware-line creative launched mid-month, most of the
  improvement traces to the second half of the month, not spread
  evenly

A reasonable Claude output:

October was a clear step up: cost per conversion dropped to $15.30 from $18.10, and total conversions grew to 340 from 265, on essentially flat spend. Worth knowing where that came from: the cookware-line creative launched mid-month, and most of the improvement traces to the second half of October specifically, not an even trend across the whole month. That's a good sign for the new creative, and also means November, a full month with that creative running, is the real test of whether this holds. Recommendation: keep spend steady through November and let the cookware creative run its full course before making any changes.

Notice it didn't just report the two headline numbers and stop. It explained where the improvement actually came from, so the client doesn't wrongly assume a steady month-long trend, and it ended with one specific recommendation instead of leaving the client to guess what happens next.

Thursday: harder conversations

Anything requiring a harder conversation, a churn-risk save, a scope-creep pushback, gets prepared with the relevant prompt before the call, not improvised during it. This is talking points, not a script to read verbatim. Claude is good at ordering the argument and keeping the tone collaborative instead of defensive; it has no independent read on whether your results genuinely underperformed or the client's expectations were simply unrealistic to begin with, so you still need to know which one it actually is before the conversation.

Worked example. A client showing churn signals:

Client: a different account, a local HVAC company. Two invoices paid
late in a row, first time that's happened. On last call, mentioned
"a friend's cousin does this stuff cheaper." Results this quarter
have genuinely been flat, no real growth in leads for two months
running, and we haven't changed strategy to address it yet.

A reasonable Claude output:

Lead with the results, not the invoices or the comment about a cheaper option, those are symptoms, the flat lead numbers for two months are the actual problem, and pretending otherwise in the conversation will read as dodging it. Acknowledge it directly: "Leads have been flat for two months, and that's on us to address, not something to wait out." Then bring a specific plan for what changes, not a general reassurance, since a vague "we'll keep working on it" is exactly what erodes trust further. If there's a real explanation (a seasonal dip, a platform change), name it honestly, but pair it with one concrete action either way. Do not address the "cheaper option" comment defensively or by badmouthing the competitor, a client who's already comparing prices is signaling value concern, not price concern, and the flat results are the actual value problem to fix.

Notice the shape: name the real issue plainly, don't let a smaller symptom (late payment, a stray comment) become the focus when a bigger one exists, and pair any honest explanation with a concrete next action rather than a vague reassurance. A scope-creep conversation the same day takes a gentler version of the same shape, acknowledge the ask, name the boundary plainly without it reading as a rejection, propose a specific path forward.

Prefer this built and wired up for your business instead of adapting it yourself? Apply for a strategy call.

Friday: proof and pipeline

Real, cleared results get a testimonial or referral ask while the client's satisfaction is still fresh. A recently completed campaign with strong numbers gets a case study draft, flagged clearly for the client's sign-off before it goes anywhere. And one social proof post goes out about the agency's own work, using only real, disclosed numbers. All three get skipped constantly because none feels urgent on a Friday, and all three compound if they aren't done, a testimonial asked for a month after the result loses most of the specific enthusiasm that made it worth asking for in the first place.

As they happen: quarterly reviews and creative briefs

The quarterly business review and the creative brief aren't scheduled because they can't be, they run the week a review is actually due or a new campaign is actually starting, while the details are specific enough to be useful.

Worked example. Prepping a quarterly review:

Client: Coastal Supply Co. Q3 summary: cookware line launch was the
big win, cost per conversion down 15% quarter over quarter. One
real miss: a planned email campaign never launched, kept getting
deprioritized behind the paid social work, client hasn't mentioned
it but it was in the original scope.

A reasonable Claude output:

Lead with the win: cost per conversion down 15% quarter over quarter, driven mainly by the cookware line launch, that's the headline. Then address the miss directly, don't wait for the client to bring it up: "One thing we didn't get to this quarter that was in our original plan, the email campaign. That's on us, and here's when it's actually happening next quarter." Naming it before they do reads as accountable; letting them notice it missing on their own reads as something being hidden. Close with next quarter's focus: continuing the cookware momentum, plus the email campaign as a committed, dated deliverable this time, not a rolled-over intention.

Notice it didn't bury or skip the miss just because the client hadn't raised it. Naming an unprompted shortfall plainly, with a specific fix attached, is what keeps a quarterly review from feeling like a highlight reel a client has learned not to fully trust.

The disclosure and honesty rules most guides skip

Agencies live and die on trust in a way that makes this section non-negotiable, not a nice-to-have. Two rules matter more than any other in this system, and an AI model has no innate sense of either one, it will happily produce whatever sounds most persuasive unless told otherwise.

Never let a draft promise a specific result you can't back up. "We'll 10x your leads," "guaranteed first-page ranking," these read as confident marketing copy to a model trained on marketing copy, not as a claim someone could reasonably hold you to. The honesty rule in your business context file exists specifically to stop this: describe the approach and what gets measured, never a promised number, unless that number is a real, factual result from a completed campaign you're explicitly referencing. Even then, a real past result from one client is evidence, not a guarantee for a different client with a different starting point, and a proposal should read that way.

Never let a draft fabricate a testimonial, a client name, or a case-study number. If you ask Claude to write a case study and give it thin or vague inputs, it will sometimes fill the gap with a plausible-sounding number instead of asking you for the real one. This isn't a small style issue, a fabricated testimonial or an invented result presented as real client feedback is the kind of claim that creates genuine legal exposure under advertising and endorsement rules that govern marketing claims and endorsements, and it's also simply dishonest to the prospect reading it. Every number and every quote in a case study or testimonial needs to trace back to something the client actually said or a result that actually happened, no exceptions, and the case study prompt in this system is written to flag every draft for the client's explicit sign-off before it goes anywhere, for exactly this reason.

A few concrete swaps worth having memorized, since these are the ones that slip through most often:

• Instead of "we'll guarantee you a 10x return," write "here's the approach, and here's what we'll track to know if it's working."
• Instead of inventing a round, impressive-sounding stat for a case study, use the real number the client gave you, even if it's a less dramatic one.
• Instead of writing a testimonial "in the client's voice" from a vague impression of how the engagement went, ask the client directly what they'd say, and use their actual words.
• Instead of naming a client or sharing their specific results without checking, confirm they're comfortable being named and having those numbers shared publicly, every time, not just for the first case study.

The pattern across all four: describe what's real and verifiable, never what sounds most impressive. This isn't legal advice, and if you have real questions about a specific claim or disclosure, particularly around sponsored content, paid partnerships, or client endorsements, an advertising attorney or a review of current FTC endorsement guidance is the right call, not this article. But honesty about results and sourcing is a foundational trust issue for any agency, worth checking on every single draft, not just the ones that feel obviously risky.

One more line worth being explicit about, separate from the above: never let Claude, or any AI tool, publish a testimonial, review, or case study result as coming from a real client unless that client actually said or achieved it. This system is built entirely around drafting communications, reports, pitches, proposals, drafts for review, not manufacturing social proof. Every claim about a real client's real experience needs to come from that client, informed by AI-assisted drafting of the words around it, never invented by it.

What Claude gets wrong, and how to fix it

It invents specificity when your notes are thin. If you give it three words about a prospect or a campaign, it will sometimes fill the gaps with generic marketing language instead of asking you for more. Fix: feed it more detail, not less. Five real observations produce a better pitch than fifteen generic value propositions.

It defaults to hype language unless told not to. "10x your growth," "game-changing results," these are the marketing-copy equivalent of filler words, and most models reach for them by default because they're common in training data. The "things I never say" line in your business context file is what stops this. If it still slips through occasionally, just tell it directly in that conversation: "drop the hype, be specific instead."

It can't tell a genuinely strong result from a cherry-picked one. When you ask it to draft a case study or a report, it will present whatever numbers you give it as favorably as the framing allows, but it has no independent judgment about whether those numbers are representative or the best three weeks out of a rough quarter. Use your own judgment on that before drafting. Claude organizes the argument; you're still the one who knows whether it's honest.

It doesn't know your industry's specific conventions. Every niche has norms Claude has no way of knowing: what a realistic cost-per-conversion actually looks like in your vertical, which platforms your specific clients care about, what "fast" turnaround actually means in your market. This is exactly what the business context file is for, and it's worth updating it the first few times you notice Claude missing something an experienced agency owner in your niche would automatically know.

Extending this to your own business

Every prompt in this system follows the same shape: give Claude real specifics, name what you don't want, ask for a length limit and a clear, honest ending. If you want to add an eleventh prompt for something this pack doesn't cover, an onboarding kickoff document, a win-back pitch to a former client, write it the same way. The business context file does most of the real work; the individual prompts just point it at a specific task.

If you're not in marketing, the shape of this system, ten pieces of recurring writing plus a business context file plus a fixed weekly cadence, is the same one used for the real estate, recruiting, and accounting prompt packs. The specific prompts change. The system underneath doesn't.

FAQ

Do I need Claude Code, or does Claude.ai work? Claude.ai with a Project works fine to start. Claude Code is worth it once you want Claude referencing your actual reporting exports or a running client-status log directly instead of you copying details in by hand every time.

Will this sound like every other agency using the same prompts? Only if you skip the business context file. The prompts are the same for everyone who copies them; your specialty, your tone, and your specific client and campaign details are not. Two agencies running this exact system on two different accounts will produce two different outputs, because the inputs are different.

How long does this take once it's set up? The business context file is a one-time, fifteen-minute setup. After that, each individual prompt takes under a minute to run, the time is in reviewing and editing the output, not generating it, which is the point: it should save you the blank-page time, not the strategic judgment time.

What if Claude writes something inaccurate about a campaign or a result? Treat every output as a first draft prepared in thirty seconds instead of ten minutes, not a finished claim. You still need to verify every factual claim, a metric, a client quote, a deadline, against your actual reporting data before it goes out. Claude doesn't know your client's real numbers unless you give it accurate details in that conversation.

Does this replace my reporting dashboard's automated summaries? No, and it isn't trying to. A dashboard's automated summary is the same generic template applied to every account regardless of what actually happened. This system produces one specific, individually reasoned report per client, which is a different kind of tool for a different part of the job. Many agencies run both: the dashboard handles the raw numbers on autopilot, this handles the writing that genuinely needs a human read of what those numbers mean.

What happens the first week I skip a day? Nothing breaks, and that's the point of building this as a cadence rather than a fragile all-or-nothing routine. If Monday's outreach doesn't happen, it's still worth doing Tuesday, a slightly late pitch beats no pitch. The value isn't in perfect adherence, it's in the writing getting done on most weeks instead of the rare week you happen to have spare time.

Can I use this for a team, not just a solo owner? Yes, with one change: give each account manager their own business context file rather than sharing one. The tone and client relationship should reflect the individual person writing the message, not a single agency-wide voice, otherwise every account manager's clients start receiving messages that sound identical to every other account manager's clients, which undercuts the personal relationship the whole system is meant to strengthen.

What this doesn't replace

None of this runs a campaign, builds creative assets, or decides strategy. Every output here needs a human read before it goes out, especially anything touching a promised result or a client's actual numbers, where the honest judgment call is not something a prompt should make for you.

Getting started today

Write your business context file first. It's the one step that makes every prompt after it better, and it takes fifteen minutes. Then pick one day, Monday is the easiest since it only requires one prospect worth pitching to exist, and run the outreach prompt once on a real prospect. If the output needs one edit, you're ready for the rest of the week. If it needs five, your notes going in were too thin, not the prompt failing, so add more specifics and try again.