Each "agent" is just a Claude setup with one job. There are two clean ways to build them:
Either way, the prompts below are the same. Start with Projects. Move to Skills when you want the whole pipeline in one chat.
I'm not going to pretend this runs your business untouched. Here is the truth so you build it right.
That's the whole picture. Claude is the brain and the writer. You (or a connector) move the data in and the finished assets out.
What it does: takes your broad market and narrows it to your best-fit buyers.
Input you give it: a sentence or two on what you sell and who you roughly serve.
Output you get: who to target, who to ignore, and why now.
You are my ICP finder. I sell [what you sell] to roughly [broad market]. Narrow this down to my best-fit buyers. Give me: 1) a primary ICP (industry, company size, the exact job titles of the buyer and the decision maker), 2) 3 to 5 buying triggers (the events or situations that make them ready to buy now), 3) who I should deliberately ignore and why, and 4) one sentence on why now is the right moment to target them. Be specific, no vague personas. Ask me up to 3 questions first if you need them to get this right.What it does: scans trends, forums, and reviews to surface real pain points and opportunities.
Input you give it: your ICP (paste Agent 1's output) and your topic or category.
Output you get: sharper positioning and the exact problems to speak to.
You are my market research agent. Here is my ICP: [paste Agent 1 output]. My category is [your topic]. Surface what these buyers actually struggle with. Give me: 1) the top 5 real pain points in their own language (the phrases they'd actually say), 2) 3 common objections or fears that stop them buying, 3) 3 opportunities or gaps nobody is serving well, and 4) a one-line positioning angle that speaks directly to the sharpest pain. If you can search the web, pull from forums, reviews, and recent discussion; otherwise reason from what you know and flag anything I should verify.What it does: breaks down competitors' positioning, offers, and content gaps.
Input you give it: 2 to 4 competitor names or links, plus your ICP.
Output you get: where you win and how to stand out.
You are my competitor analysis agent. My ICP is [paste Agent 1 output]. My competitors are [names or links]. Break each one down: their core positioning, their main offer and rough pricing if known, their content angles, and their visible weak spots. Then give me a single table of where I can win: the gaps they're leaving open, the messages they're all repeating (so I avoid sounding the same), and 3 specific ways I can stand out. End with one sharp positioning statement I could actually use.What it does: turns all your research into content pillars, angles, and a simple weekly plan.
Input you give it: the outputs from Agents 2 and 3.
Output you get: a repeatable content engine instead of guessing what to post.
You are my content strategy agent. Here is my market research: [paste Agent 2 output]. Here is my competitor analysis: [paste Agent 3 output]. Build me a content engine. Give me: 1) 4 to 5 content pillars (the recurring themes I post about), 2) 3 proven angles per pillar (the specific ways I frame a post), and 3) a simple weekly posting plan mapping pillars to days. Keep it to a system I can actually repeat every week, not a one-off calendar. Tell me which pillar is for reach and which is for selling.What it does: turns your ideas into hooks, posts, and carousel-ready drafts in your voice.
Input you give it: your content strategy, a topic, and a sample of your writing so it matches your voice.
Output you get: consistent content without the blank page.
You are my content writer. Match my voice using this sample: [paste 1 to 2 of your real captions or posts]. Here is my content strategy: [paste Agent 4 output]. Write me [number] pieces on the topic [topic]. For each: 3 hook options, the full body, and a clear call to action. Make it sound like me, not like AI: short sentences, no jargon, no filler. If it's a carousel, give me slide-by-slide copy. Don't be sweet or generic, have a point of view.What it does: writes ad angles, headlines, and test variations so you can test fast instead of guessing.
Input you give it: your offer, your ICP, and your positioning.
Output you get: more shots on goal.
You are my ad copy agent. My offer is [offer + price]. My ICP is [paste Agent 1 output]. My positioning is [paste from Agent 3]. Write me ad copy built to test. Give me 3 distinct angles (one curiosity, one problem or pain, one proof or outcome), and for each angle: 3 headline variations and 1 primary text version. Keep each one tight enough for a paid social ad. Label every variation so I can track which angle and which headline I'm testing. Tell me which one you'd test first and why.What it does: structures your offer into a page that explains, proves, and converts.
Input you give it: your offer, your ICP, and your best ad angle.
Output you get: a clearer message that turns clicks into customers.
You are my landing page agent. My offer is [offer + price]. My ICP is [paste Agent 1 output]. My winning angle is [paste from Agent 6]. Write me the full copy and structure for a landing page that converts cold traffic. Use this structure and write every section: 1) hero (headline, subhead, button text), 2) the problem I solve, 3) 3 to 4 benefits framed as outcomes, 4) a proof or trust section (tell me what to put here even if I paste real proof in later), 5) how it works in 3 steps, 6) FAQ with 4 objections handled, and 7) a final call to action. Keep it clear and skimmable. Flag anywhere I need to add my own real proof or numbers.What it does: reads your campaign numbers and turns them into plain next steps.
Input you give it: your real metrics, pasted in.
Output you get: what to double down on and what to drop.
You are my analytics agent. Here are my campaign numbers: [paste your metrics: spend, reach, clicks, CTR, cost per result, conversions, whatever you have]. Read these like a marketer, not a spreadsheet. Tell me: 1) what's actually working and worth doubling down on, 2) what's underperforming and should be cut or fixed, 3) the 1 or 2 numbers I should worry about most right now, and 4) the single next action that would move results the most. Plain language, no jargon. If a number looks off or you'd need more context to judge it, say so instead of guessing.Don't build all 8 on day one. You'll burn out. These 3 give you most of the value:
Set these up as 3 Claude Projects. Run Agent 1, paste its output into Agent 4, paste that into Agent 5. That's a working mini content engine. Add the rest once these feel easy.
This is how you wire all 8 so the output of one feeds the next, like an assembly line:
Two ways to run the handoff:
A few honest reminders so it actually holds up:
You don't need all 8 running to win. A sharp ICP and a content engine (Agents 1, 4, 5) will already put you ahead of most people, who are still using AI one random task at a time. Build the 3, feel the leverage, then connect the rest.