Set Up Your AI Marketing Team in Claude (7 Roles, Zero Hires)

A real marketing team for a small business costs somewhere between $8,000 and $25,000 a month, depending on where you hire and how many people you need. A strategist, a copywriter, an editor, an SEO specialist, a designer, an analyst — that's five or six salaries before you've published a single post.

I don't have that team. I have Claude, a set of seven role prompts I've refined over months, and a weekly rhythm that takes about two hours of my time spread across five days. My content goes out on schedule, it sounds like me, and I can actually see what's working.

This post is the full setup. Not the "here's a cool prompt" version — the whole system: seven roles, the exact system prompt for each one, an example of what goes in and what comes out, how to save each role so you never rebuild it, and the weekly operating rhythm that ties them together.

One thing before we start: this is not magic, and I'll be honest at the end about what still needs a human (there's a real list). But the gap between "I should post more" and "I post consistently and it's good" is smaller than most people think. It's mostly a systems problem, and systems are exactly what AI is good at.

Why roles instead of one big prompt

The obvious approach is to write one giant prompt: "You are my marketing assistant. Help me plan, write, edit, and analyze my content." I tried that. It produces mush.

Here's why. When you ask one prompt to do everything, Claude averages across all the jobs. The strategist part wants to think in quarters; the copywriter part wants to think in hooks; the editor part wants to cut. Blend them and you get generic output that does none of the jobs well — a plan that's vague, copy that hedges, edits that don't bite.

Splitting the work into roles fixes this for three reasons:

1. Each role gets a narrow, specific job. Narrow prompts produce sharp output. "Write a LinkedIn post" is vague. "Write a LinkedIn post using this voice guide, this hook pattern, and this structure, under 250 words" is a spec.
2. Roles can check each other. My Editor role exists specifically to attack what my Copywriter role produced. You can't get that tension inside one prompt — a single prompt won't honestly critique its own output in the same breath.
3. Roles are reusable. A role saved as a skill or project instruction is an asset. You invest an hour tuning it once, then use it hundreds of times. That's the difference between prompting and building a workflow — and repeatable workflows are the whole game.

Think of it like hiring. You wouldn't hire one person and ask them to be strategist, writer, editor, designer, and analyst simultaneously. You'd hire specialists. Same logic, except these specialists cost you a Claude subscription.

How the pieces fit

Before the roles, a quick map of the system so the parts make sense:

• The Strategist runs once a month and produces the content plan. Everything downstream works from this plan.
• The Copywriter and SEO Writer produce drafts — short social posts and long articles respectively.
• The Repurposer multiplies: one long piece becomes a LinkedIn post, an X thread, a newsletter section, and a shorts script.
• The Editor/Humanizer is the quality gate. Nothing ships without passing through it.
• The Designer brief-writer turns finished posts into image briefs so visuals don't become the bottleneck.
• The Analyst closes the loop every Friday: what worked, what didn't, what to change.

Plan → produce → multiply → polish → visualize → measure. That's the pipeline. Now the roles.

Role 1: The Strategist

Job: Turn your business goals into a monthly content plan you can actually execute.

Most solo content fails at this layer, not at writing. People sit down on Tuesday with no idea what to post, write whatever comes to mind, and wonder why nothing compounds. The Strategist's job is to make sure every post exists for a reason.

Here's the system prompt. Paste it into a Claude Project's instructions, or save it as a skill (I'll cover both methods after the roles).

You are my content strategist. You work for a solo business owner, not an
agency, so every plan you produce must be executable by one person with
limited time.

MY BUSINESS CONTEXT (I will fill this in):
- What I sell: [product/service and price point]
- Who buys it: [audience, their situation, what they struggle with]
- This quarter's business goal: [e.g., 20 discovery calls, 500 email subs]
- Channels I actually use: [e.g., LinkedIn, newsletter, YouTube shorts]
- Publishing capacity: [e.g., 3 LinkedIn posts + 1 article + 1 newsletter/week]

WHEN I ASK FOR A MONTHLY PLAN, PRODUCE:
1. THEME OF THE MONTH — one sentence connecting the month's content to the
   quarterly business goal. If you can't connect them, say so and ask.
2. CONTENT PILLARS — 3 pillars max, with the % split between them
   (e.g., 60% teaching, 25% opinion/story, 15% offer-adjacent).
3. THE CALENDAR — a week-by-week table: date, channel, working title,
   pillar, format (post/article/thread/newsletter), and the ONE job of the
   piece (traffic, trust, or conversion). Never assign a piece two jobs.
4. THE ANCHOR PIECE — one long-form article per month that the Repurposer
   will multiply. Mark it clearly.
5. WHAT TO SKIP — explicitly list 2-3 tempting topics I should NOT cover
   this month and why (off-goal, wrong audience, or covered recently).

RULES:
- Match my stated capacity exactly. Never plan more than I said I can ship.
- Every piece must map to a pillar and a job. No orphan content.
- Prefer specific angles over broad topics. "Pricing your first AI service"
  beats "AI for business."
- Ask me at most 3 clarifying questions before planning, then commit.
- If last month's Analyst report is provided, adjust: double down on the
  top 20%, cut what flopped, and say what you changed and why.

Example input:

"Plan August. Goal this quarter: 30 booked calls for my done-for-you email marketing service ($1,500/mo). Audience: e-commerce founders doing $20–100k/month who write their own emails badly. Channels: LinkedIn and a weekly newsletter. Capacity: 3 LinkedIn posts and 1 newsletter per week, 1 long article this month. July's analyst report attached."

Example output (abridged):

Theme: "Your email list is an underpriced asset" — every piece shows an e-commerce founder the money they're leaving in their list, which is the exact pain my service solves.

Pillars: Teaching email tactics (60%), founder stories/opinions (25%), behind-the-scenes of client results (15%).

Week 1: Mon LinkedIn — "The 4-email sequence that recovered $11k in abandoned carts" (teaching, trust). Wed LinkedIn — "Why I tell founders NOT to send more emails" (opinion, trust). Fri LinkedIn — "What a $1,500/mo email retainer actually includes" (BTS, conversion). Thu newsletter — "Audit your welcome sequence in 20 minutes" (teaching, trust).

Anchor piece: "The E-commerce Email Playbook: 7 sequences every store needs" — publishes week 2, feeds the Repurposer for weeks 3–4.

Skip this month: SMS marketing (off-goal — you don't sell it), AI news commentary (wrong audience — founders don't care), open-rate benchmarks (covered in June, flopped per analyst report).

Notice what the rules force: the plan matches your real capacity, every piece has one job, and there's an explicit "don't cover this" list. That last one matters more than it looks — most content drift comes from chasing shiny topics that don't serve the goal.

Run it: once a month, first Monday, 30 minutes including your review.

Role 2: The Copywriter

Job: Write social posts that sound like you, not like a language model.

This is the role where most people give up on AI, because default Claude output has a recognizable flavor — smooth, balanced, a little breathless. The fix isn't a magic phrase. It's giving the role a real voice guide built from your actual writing, plus hard structural rules.

Before you use this prompt, collect 5–10 pieces of your own writing that sound like you on a good day. Posts, emails to friends, Slack rants — anything with your natural voice. You'll paste them in once, and the role keeps them forever.

You are my copywriter. Your only job is short-form social posts in MY
voice. You are not a general assistant; if I ask for strategy or long-form,
tell me to use the right role.

MY VOICE (learn from these samples, I will paste 5-10 below):
[PASTE WRITING SAMPLES HERE]

From the samples, follow my: sentence length habits, vocabulary level,
how I open posts, how much I hedge (probably very little), whether I use
questions, how I handle humor, and what I never say.

STRUCTURE OF EVERY POST:
- HOOK (line 1): a specific claim, number, tension, or moment. Never a
  greeting, never a question you immediately answer, never "I've been
  thinking about..."
- BODY: one idea only. Concrete beats abstract — numbers, names of tools,
  actual dialogue, real amounts of money and time.
- ENDING: a takeaway or a soft question. No "thoughts?" No engagement bait.
- LENGTH: 120-250 words for LinkedIn, under 280 characters per tweet.

HARD BANS (these instantly mark text as AI-written):
- "delve", "dive into", "game-changer", "landscape", "leverage" (as verb),
  "unlock", "elevate", "in today's fast-paced world", "it's not about X,
  it's about Y" more than once per week, emoji unless my samples use them
- Perfectly parallel three-item lists in every post
- Starting consecutive sentences with the same structure
- Summarizing the post at the end of the post

PROCESS:
1. I give you a working title + the ONE job of the post (from the
   Strategist's calendar) and any raw material (a story, a result, a take).
2. You ask for a real detail if I gave you nothing concrete. Do not invent
   client results, numbers, or stories. Ever. Placeholder them as [DETAIL].
3. You produce 2 versions: one safe, one with a riskier hook.
4. Keep drafts tight. I'd rather cut nothing than cut half.

Example input:

"Wednesday post: 'Why I tell founders NOT to send more emails.' Job: trust. Raw material: last month a client wanted to go from 2 to 5 emails a week, I pushed back, we went to 3 but segmented properly, revenue per email doubled."

Example output (version B, the riskier hook):

A client asked me to send his list 5 emails a week. I talked him down to 3 — and doubled his revenue per email.

Here's the thing nobody selling "email volume" tells you: your list doesn't have one audience. His had three — new subscribers, past buyers, and window shoppers who'd clicked but never bought. Blasting all of them 5 times a week trains two of those groups to ignore you.

So instead of more emails, we sent fewer, split three ways. New subscribers got the welcome story. Past buyers got the restock alert. Window shoppers got the objection-killer.

Same effort. Half the unsubscribes. 2x revenue per send.

More isn't a strategy. Matching the message to the reader is.

The two rules doing the heavy lifting here are the ban list and rule 2 of the process — never invent specifics. An AI copywriter that fabricates client results will eventually publish a lie under your name. Force placeholders and fill them yourself.

Run it: daily, in the 20-minute production loop (details in the rhythm section).

Role 3: The Repurposer

Job: Turn one long piece into a week of content across every channel.

This role has the best effort-to-output ratio of the seven. You write (or the SEO Writer drafts) one substantial article, and the Repurposer converts it into a LinkedIn post, an X thread, a newsletter section, and a script for a short video. One unit of thinking, four units of distribution.

The trap to avoid: naive repurposing just summarizes the article four times. Real repurposing understands that each channel wants a different slice — LinkedIn wants the story, X wants the list, the newsletter wants the personal angle, video wants one visual idea. The prompt encodes that.

You are my content repurposer. I give you ONE long-form piece (article,
newsletter, or video transcript). You extract channel-native derivatives.
You never summarize — you re-angle.

FOR EACH LONG PIECE, PRODUCE:

1. LINKEDIN POST (120-250 words)
   Take the single most surprising or contrarian point from the piece and
   build the post around ONLY that point. Do not mention "I wrote an
   article." The post must stand alone. End with a line that invites the
   curious to the full piece.

2. X/TWITTER THREAD (5-8 tweets)
   Take the most LIST-LIKE part of the piece (steps, mistakes, rules).
   Tweet 1 is the hook: promise + credibility + no hashtags. One idea per
   tweet, each under 280 characters, each readable on its own. Final tweet
   links the full piece.

3. NEWSLETTER SECTION (150-300 words)
   Take the most PERSONAL angle — why I wrote this, what I got wrong
   before I learned it, who specifically needs it. Written like an email
   to one person. Link the piece mid-text, not at the end.

4. SHORT VIDEO SCRIPT (45-60 seconds, ~130 words)
   Take the ONE point that can be shown or drawn. Format:
   - HOOK (first 2 seconds, on screen + spoken): the claim
   - BODY: 3 beats max, each one sentence + a visual note in [brackets]
   - CTA: one action, five words or fewer
   Write it for the mouth, not the page — contractions, short words.

RULES:
- Each derivative uses a DIFFERENT part of the source. If two outputs
  cover the same point, you've failed — redo one.
- Match my voice samples (inherited from the Copywriter role if saved in
  the same project).
- Flag anything in the source that is time-sensitive and will date badly.
- Output all four in one response, clearly separated, ready to paste.

Example input: the anchor article "The E-commerce Email Playbook: 7 sequences every store needs" (3,000 words).

Example output (abridged):

LinkedIn — built around the article's most contrarian point: "The highest-ROI email sequence isn't the abandoned cart. It's the post-purchase sequence — the one 80% of stores don't have..." (full 200-word post follows).

X thread — built from the list: "7 email sequences every e-commerce store needs (most have 2): ..." then one tweet per sequence with the single most useful line about each.

Newsletter — the personal angle: "I used to set up abandoned cart emails first for every client. I was wrong, and it cost one of them about $30k. Here's what I do now..."

Shorts script — the drawable point: HOOK: "Your best customer is the one who just bought." [on screen: a customer icon with a receipt] Beat 1... Beat 2... CTA: "Full playbook — link below."

Run it: the week after your anchor piece publishes. One run fills most of a week's calendar.

Role 4: The Editor / Humanizer

Job: Kill the AI flavor, tighten everything, and stop bad posts from shipping.

This is the quality gate, and it's deliberately adversarial. The Copywriter's job is to produce; the Editor's job is to attack what was produced. Keeping these as separate role invocations matters — when I ask the same conversation that wrote a draft to critique it, the critique is polite. A fresh Editor invocation with an adversarial prompt is not polite.

You are my editor. Your job is to make drafts sharper, shorter, and more
human. You are adversarial by design: assume every draft I give you has
problems, and find them.

PASS 1 — THE AI-TELL SCAN. Flag and fix:
- Hedging ("might", "could potentially", "it's worth noting")
- Empty intensifiers ("truly", "incredibly", "seamlessly")
- Symmetric structures repeated (three parallel clauses, every paragraph
  the same length, "It's not X. It's Y." patterns)
- Words real people don't say: delve, robust, landscape, tapestry,
  leverage, foster, elevate, unlock, harness, navigate (metaphorical)
- A conclusion that restates the whole piece
- Perfect grammar where a human would use a fragment. Like this.

PASS 2 — THE TIGHTEN. Cut 15-25% of the word count. Kill every sentence
that repeats an earlier sentence's job. If a paragraph works without its
first sentence, delete the first sentence.

PASS 3 — THE HONESTY CHECK. Flag any claim that is unverifiable, any
number without a source, and anything that overpromises. I will not
publish "guaranteed", "always works", or invented specificity.

OUTPUT FORMAT:
1. The edited draft, clean, ready to publish.
2. A change log: 3-6 bullets on what you cut/changed and why.
3. A verdict: SHIP / SHIP WITH FIXES / DON'T SHIP, with one sentence of
   reasoning. Use DON'T SHIP when the core idea is weak — do not polish
   a post that shouldn't exist.

RULES:
- Preserve my voice. You tighten it; you never replace it with yours.
- Never add new claims or examples. Editing only.
- If the hook is weak, propose 2 alternative hooks using only material
  already in the draft.

Example input: a Copywriter draft opening with "In today's competitive e-commerce landscape, email marketing remains one of the most powerful tools available to founders..."

Example output (abridged):

Edited draft: opens instead with "Most Shopify stores I audit have $10–40k a year sitting untouched in their email list." (rest of tightened draft follows)

Change log: Cut the throat-clearing first paragraph — the real hook was buried in sentence four. Removed "truly", "seamlessly", and two "it's worth noting"s. Cut 22% overall. Flagged "most powerful tool available" as an unverifiable superlative and replaced with the specific audit number.

Verdict: SHIP WITH FIXES — confirm the $10–40k figure is from your real audits before publishing.

The verdict system is the part I'd fight to keep. An editor that can only say "here's a better version" still lets weak ideas through in nicer clothes. An editor that can say DON'T SHIP saves you from publishing filler. Mine kills roughly one post in ten, and it's always right.

Run it: on every single piece, before it ships. No exceptions — this is the gate.

Role 5: The SEO Writer

Job: Draft long-form articles that answer search intent and feed the Repurposer.

Social posts build audience; search content builds an asset. This role drafts the long pieces — including your monthly anchor article — with structure that both readers and search engines can follow. Note the word drafts: SEO articles are where AI most needs your expertise layered on top, because generic long-form is exactly what search engines are learning to ignore.

You are my SEO writer. You draft long-form articles (1,500-3,000 words)
that satisfy a specific search intent and are worth reading by a human
who never thinks about SEO.

FOR EACH ARTICLE I REQUEST, I PROVIDE: target query, searcher's situation,
my unique angle or experience, and 2-4 internal pages to link.

BEFORE DRAFTING, OUTPUT A BRIEF FOR MY APPROVAL:
- Search intent in one sentence (what the searcher needs to walk away with)
- Working title (query near the front, but readable as a human headline)
- H2/H3 outline — the searcher's question answered in the first 150 words,
  then depth. No "What is X?" section unless the query is literally that.
- Where each internal link naturally fits

AFTER I APPROVE, DRAFT WITH THESE RULES:
- Answer the core question immediately. Never make the reader scroll past
  300 words of context to get the answer. Context comes after.
- Short paragraphs (1-4 sentences). Descriptive H2s a skimmer can navigate
  by. One table or list wherever it genuinely compresses information.
- MARK EVERY SPOT that needs my real experience with [YOUR EXPERIENCE:
  what happened when you...]. Do not fill these with plausible inventions.
  An article with zero experience markers is a failed draft.
- No keyword stuffing. Use the query and natural variants where a person
  would. Write the meta description (under 155 chars) at the end.
- Cite nothing you cannot verify; write "as of [date]" for anything that
  changes over time.

Example input:

"Target query: 'abandoned cart email sequence'. Searcher: e-commerce founder who just realized their one abandoned-cart email isn't enough. My angle: results from 12 client stores. Internal links: my email audit article and my services page."

Example output (abridged brief):

Intent: the searcher wants the actual sequence — how many emails, what timing, what goes in each — not a definition of cart abandonment.

Title: "Abandoned Cart Email Sequence: The 4-Email Setup I Use for Client Stores"

Outline: the sequence itself in the first 150 words (4 emails: 1hr / 24hr / 3 days / 7 days), then an H2 per email with copy guidance, then timing variations by price point, then [YOUR EXPERIENCE: recovery-rate ranges across your 12 stores], then mistakes, then FAQ from real query variants.

The [YOUR EXPERIENCE] markers are the difference between content that ranks and content that gets skipped. Search is saturated with competent generic articles. Your unfakeable client numbers, your failures, your specific process — that's the only moat, and the role is built to leave holes shaped exactly like it.

Run it: once or twice a month. Brief on day one, draft on day two, your experience pass on day three, Editor on day four.

Role 6: The Designer Brief-Writer

Job: Turn finished posts into image briefs so visuals never block publishing.

I can't design, and I'm guessing you don't want to. But posts with visuals consistently outperform text-only, and "I still need an image" is the number one reason a finished post sits unpublished for three days. This role doesn't make images — it makes briefs precise enough that an AI image tool, a Canva template, or a $30 freelancer can execute without follow-up questions.

You are my design brief-writer. I give you a finished post or article.
You produce a brief for its visual. You do not create images; you create
instructions precise enough that anyone (or any tool) can execute them
without asking me anything.

FIRST, PICK THE VISUAL TYPE (choose one, say why in one line):
- DIAGRAM: the post explains a process, comparison, or relationship
- BIG-TEXT CARD: the post's power is one quotable line
- DATA VISUAL: the post contains numbers worth seeing
- SCREENSHOT-STYLE: the post references something on a screen
- PHOTO-STYLE: the post is a personal story

THEN WRITE THE BRIEF:
1. ONE-LINE PURPOSE: what a viewer should understand in 2 seconds while
   scrolling, before reading anything.
2. EXACT TEXT ON IMAGE: every word, verbatim. Under 12 words total for
   cards; label text for diagrams. Typos here become typos in public.
3. LAYOUT: described spatially ("headline top-left, 3 boxes in a row
   below, arrow flowing left to right"). Never "clean and modern" — only
   things a person can draw.
4. STYLE CONSTANTS (from my brand, set once): colors with hex codes,
   font feel, light/dark, logo placement, aspect ratio per channel
   (LinkedIn 1200x1200, X 16:9, story 9:16).
5. AI-IMAGE PROMPT: if the visual type suits generation, write the
   ready-to-paste prompt for an image model, including negative
   instructions (no extra text, no watermarks, no hands if photo-style).

RULES:
- One visual concept per brief. Offer a second concept only if I ask.
- If the post doesn't need a visual, say NO VISUAL and why. A bad image
  is worse than none.
- Numbers and names on the image must match the post exactly.

Example input: the "3 emails beat 5" post from the Copywriter section.

Example output (abridged):

Type: DIAGRAM — the post's core is a comparison (blast vs. segmented), which is drawable in one glance.

Purpose: a scroller should instantly see "5 arrows to one blob = bad, 3 arrows to three groups = good."

Exact text: Left panel label: "5 emails/week → everyone". Right panel label: "3 emails/week → the right people". Bottom line: "2x revenue per send."

Layout: two panels side by side; left shows five identical arrows hitting one gray crowd; right shows three colored arrows each hitting a distinct smaller group; bottom banner with the result stat. 1200x1200.

AI-image prompt: included, with hex codes and "no additional text, no watermark" negatives.

Run it: in the daily loop, right after the Editor approves a post — brief today, execute or delegate tomorrow.

Role 7: The Analyst

Job: Review the week's numbers and turn them into next week's decisions.

Without this role, the other six run open-loop — you produce forever without learning. With it, the system compounds: every Friday the Analyst reads your numbers, finds the pattern, and hands specific instructions back to the Strategist and Copywriter.

One honest caveat: Claude can't see your analytics. You have to bring the numbers — a copy-paste from LinkedIn analytics, a CSV export, or even a typed list. That takes me about five minutes and it's the only data-entry chore in the whole system.

You are my content analyst. Every week I paste in my numbers (per post:
channel, date, topic/hook, impressions, engagements, comments, clicks,
plus any leads/replies/sales I can attribute). You find the signal.

PRODUCE THE WEEKLY REVIEW:
1. SCOREBOARD — a table of the week's posts ranked by engagement rate
   (engagements/impressions), with absolute numbers alongside. Note
   which had visuals.
2. PATTERN OF THE WEEK — the ONE most defensible pattern in the data
   (hook style, topic, format, day, visual type). State your confidence:
   with 3-5 posts a week, most patterns are weak signals. Never present
   noise as insight — say "not enough data" when it's true.
3. BUSINESS CHECK — engagement is not the goal. Which content produced
   profile visits, DMs, replies, subscribers, or sales conversations?
   If the best-engaging post produced zero business signal, say so.
4. THREE DECISIONS — specific and small:
   - One thing to do MORE of next week (with the evidence)
   - One thing to STOP or change (with the evidence)
   - One EXPERIMENT for next week (a single variable, testable)
5. NOTE TO THE STRATEGIST — 2-3 lines I will paste into next month's
   planning session. Cumulative memory is your most valuable output.

RULES:
- Compare against MY baseline (trailing 4-week averages), not industry
  benchmarks.
- Distinguish reach luck from content quality: one viral outlier does
  not make a pattern. Track medians, not means.
- Keep the whole review under 400 words. I read this in 5 minutes on
  Friday, then close the laptop.

Example input: pasted stats for four posts plus "2 DMs this week, both referencing the retainer-pricing post."

Example output (abridged):

Scoreboard: retainer-pricing post: 4.1% ER, 2 DMs. Segmentation post: 3.2% ER. Audit-checklist post: 2.9% ER. Industry-news comment: 1.1% ER.

Pattern: posts with a specific dollar amount in the hook took the top two slots — second week running, so confidence is moderate. The news commentary underperformed your 4-week median ER of 2.8% for the second time.

Business check: both DMs came from the pricing post, which had your second-best engagement. Transparency about pricing is out-earning pure teaching content even when it engages slightly less.

Decisions: MORE: one behind-the-scenes/pricing post weekly, not biweekly. STOP: news commentary — two strikes. EXPERIMENT: next week, run the same teaching topic with a dollar-amount hook vs. a question hook, same day-slot a week apart.

Note to Strategist: August plan should raise the BTS/offer pillar from 15% to 25%; audience responds to transparency. Kill the news-reaction slot.

Run it: Fridays, 15 minutes. The "note to the Strategist" lines are the system's memory — save them in a running doc and paste the doc into every monthly planning session.

Saving the roles so you never rebuild them

Seven roles are useless if you re-paste prompts every morning. Three ways to make them permanent, from simplest to most powerful:

Option 1: Claude Projects (easiest, claude.ai). Create one Project per role — "Marketing: Copywriter," "Marketing: Editor," and so on. Paste the role's system prompt into the Project's custom instructions, and upload standing materials as Project knowledge (voice samples for the Copywriter, brand colors for the Designer, the running Analyst-notes doc for the Strategist). Daily use becomes: open the right Project, type your input, done. Each Project keeps its own conversation history, which doubles as a work archive.

Option 2: Claude Code skills (most repeatable). If you're comfortable with Claude Code, save each role as a skill — a folder containing a SKILL.md file whose body is the role prompt, with a name and description in its frontmatter so Claude knows when to invoke it. Then /copywriter Wednesday post about segmentation runs the whole role from your terminal. The killer advantage: skills are files, so they're version-controlled. When the Analyst discovers dollar-amount hooks outperform, you edit the Copywriter skill's hook rules, commit, and the improvement is permanent. Your team literally gets better every week, and you can see the diff. My walkthrough of this setup is in the Claude workflow template post, and a packaged version of all seven roles is in the AI marketing team skill.

Option 3: One project, explicit role switching. Some people prefer everything in one place: paste all seven prompts into a single Project's instructions under clear headings, and start each message with the role name ("EDITOR: review this draft"). It works, but the roles bleed into each other more than I like — the Editor gets gentler when it shares context with the Copywriter that wrote the draft. I use this only for the Copywriter → Editor → Designer chain on days I'm moving fast, and keep the Strategist and Analyst strictly separate.

Whichever option you choose, the principle is the same: the prompt is an asset you improve, not a message you retype. That's what makes this a workflow instead of a habit of chatting with AI.

The weekly operating rhythm

Roles are the team; the rhythm is the job schedule. Without fixed times, you'll use the roles for two enthusiastic weeks and then drift. Here's the exact schedule I run. Total cost: about two focused hours a week, plus the monthly planning session.

Day	Session	Duration	Roles used	What happens
Monday (first of month only)	Monthly planning	30 min	Strategist (+ Analyst notes)	Paste quarterly goal, capacity, and the running Analyst-notes doc. Review the plan, push back once, lock the calendar.
Monday	Weekly kickoff	20 min	Strategist output, Repurposer	Read this week's row of the calendar. If an anchor piece exists, run the Repurposer on it — most of the week's drafts now exist.
Tuesday	Production loop	20 min	Copywriter → Editor → Designer	Draft tomorrow's post (2 versions), pick one, Editor pass, accept the verdict, generate the design brief. Schedule it.
Wednesday	Production loop	20 min	Copywriter → Editor → Designer	Same loop for the next slot. If it's an SEO week: review the SEO Writer's brief or fill in [YOUR EXPERIENCE] markers instead.
Thursday	Production loop	20 min	Copywriter → Editor → Designer	Same loop. Draft the newsletter from the Repurposer's newsletter section if there's an anchor piece this month.
Friday	Weekly review	15 min	Analyst	Paste the week's numbers. Read the review. Apply the three decisions. Append the "note to Strategist" to the running doc. Close laptop.

Three notes on making the rhythm stick:

Work one day ahead. Tuesday's session produces Wednesday's post. This buffer means a bad session (or a DON'T SHIP verdict) never creates a same-day scramble — you just fix it tomorrow.

The 20 minutes is a hard cap. If the Copywriter's drafts are weak, don't tinker for an hour — improve the role's prompt for two minutes and try once more. Slow output is almost always a prompt problem, and fixing the prompt pays forever; fixing today's draft pays once.

Friday's decisions must land somewhere. A decision like "more pricing posts" gets applied by editing the Strategist's calendar or the Copywriter's instructions that day, while it's fresh. An insight that doesn't change a prompt or a plan is entertainment, not analysis.

What still needs a human (the honest part)

I promised zero hype, so here is the unvarnished list of what this system cannot do. Skip this section and the system will quietly fail on you in about six weeks.

Your experience cannot be generated. The Copywriter needs your real client stories. The SEO Writer leaves [YOUR EXPERIENCE] holes on purpose. If you fill those holes with plausible-sounding inventions, you'll publish confident fabrications under your own name — and your audience, who lives in your niche, will catch it. The AI handles structure, drafts, and polish; the lived substance is your job, forever.

Taste is the moat. The Editor catches AI-isms, but it cannot tell you that a technically fine post is boring, off-brand, or beneath you. You are the final filter, which is why every session in the rhythm ends with you reading the output before it ships. The day you start auto-publishing without reading is the day quality starts decaying — slowly enough that you won't notice until your numbers do.

Relationships don't automate. Replies, DMs, comments on other people's posts — the actual social part of social media — is where trust converts to business, and it has to be you. My rule: for every 20 minutes producing content, spend 10 genuinely engaging. No role in this post writes your comments, on purpose.

The Analyst sees numbers, not causes. It can tell you the pricing post got 2 DMs. It can't know that one DM came from a founder who'd been following you silently for eight months. Attribution for a solo business is mostly qualitative — ask people who reach out how they found you, and feed that back into the notes doc yourself.

Strategy pivots are yours. The Strategist optimizes execution of the goal you give it. If the goal is wrong — wrong offer, wrong audience, wrong price — the system will efficiently produce content marketing for a business that shouldn't exist in its current form. AI amplifies direction; it doesn't provide it.

And the drafts are drafts. Even with a tuned voice guide and a merciless Editor, I change something in most pieces before publishing — a sharper detail, a cut line, a real number. The system's honest promise isn't "never touch content again." It's "never start from a blank page again, and never wonder what to post." For me that turned publishing from a 10-hour weekly grind I skipped half the time into a 2-hour rhythm I actually keep. That trade — not magic autopilot — is what's on offer, and it's plenty.

Start with two roles, not seven. The Copywriter and the Editor give you the fastest win: better posts, shipped consistently, this week. Add the Strategist at the start of next month, the Analyst the Friday after, and the rest as the rhythm settles. Six weeks from now you'll have a marketing team that never misses a Monday — and the only hire was a prompt library.

Resources & links

• /skills/ai-marketing-team — all seven role prompts from this post, packaged as a ready-to-install Claude skill set.
• /blog/humanize-ai-writing-prompt — the deep-dive on Role 4: the full humanizing pass and the complete AI-tell ban list.
• /blog/one-person-business-with-ai-30-days — the 30-day plan for running an entire solo business (not just marketing) on AI workflows.
• /blog/write-seo-articles-with-ai — the expanded version of Role 5: briefs, drafting, and the experience-marker method for articles that actually rank.
• /blog/claude-workflow-template — my template for turning any repeated Claude task into a saved, versioned workflow.
• /blog/twelve-ways-to-make-money-with-claude — twelve business models this marketing system can promote, from services to digital products.
• /skills/humanize-writing — the Editor/Humanizer as a standalone installable skill.
• /skills/seo-longform-writer — the SEO Writer role as a skill, with the brief-first process built in.
• /skills/context-engineering — how to structure voice samples, business context, and notes docs so your roles get smarter over time.
• /skills/productized-ai-service — turn workflows like this one into a service you sell to clients.
• /skills/superpowers — my collection of general-purpose Claude skills that pair well with the marketing team.

If this was useful, I send one practical AI workflow like this every week — no hype, just systems that work. Join the newsletter at alexandrrich.substack.com.