Most 'agent' setups are over-engineered. Claude Projects do 80% of the work. Here are the five that actually earn the setup time.
Every "marketing agents" post I've read this year starts with a flowchart involving n8n, Zapier, MCP servers, a vector database, and a Python script. The author has spent two weeks setting it up. The agent does roughly what a Claude Project with a good system prompt does in ten minutes of setup.
This is the case for the simpler stack. Five agents. Each one runs in Claude Projects with a pinned system prompt and a Google Sheet for I/O. No code, no orchestration layer, no $200/month vendor stack. Cost: $20/month for Claude Pro, plus whatever you already pay for sheets.
Spinning up n8n, Zapier, MCP servers, and a custom RAG layer to do what a Claude Project plus a Google Sheet would handle is a vanity setup. You're spending engineering time, not winning back operator time. The orchestration is the impressive part, not the useful part.
Start with the simplest stack that works. Upgrade only when you've hit a wall the simpler stack can't get past, not when you anticipate one.
Job: Watches 3-5 competitors weekly and flags real moves, pricing changes, launches, positioning shifts.
Stack: Claude Project + a Google Sheet with each competitor's homepage and pricing URL. Manual trigger every Monday.
How: Paste the current state of each competitor page into the Project. Ask: "What changed vs. last week? Flag anything that affects positioning, pricing, or ICP."
Output: A 3-bullet Slack message to your team with real changes only. No summaries of unchanged pages.
What to watch: Don't ask Claude to "browse" the sites, that's unreliable. Grab the text yourself, then hand it to Claude. The parsing is the leverage, not the scraping.
Job: Every pillar article published, six+ channel-native assets out, each one matching your voice.
Stack: Claude Project with your extracted voice guide pinned + the Atomic Content Multiplier prompt.
How: Drop the new pillar article into the Project. Run the prompt. Edit lightly. Ship.
Why this matters: The voice guide is what makes the output not look AI-written. Skip the voice extraction step and the repurposer produces generic LinkedIn-bot posts.
Job: Categorises inbound (press, partnerships, community, support), drafts responses, flags what needs a human.
Stack: Claude Project + Gmail labels. Optional: Zapier/n8n flow for forwarding emails to Claude via webhook. Manual review is the start.
The rule: Never give the agent send permission. Drafts only. A well-meaning AI reply to press goes viral in the worst way. You stay in the loop. Always.
Job: Reads your weekly numbers and writes the standup narrative.
Stack: Looker Studio or a Google Sheet with the consolidated metrics + a Claude Project with the Performance Readout prompt pinned.
How: Monday morning, export the week as CSV. Paste into the Project alongside goals and last week's readout. Get headline, three findings, two follow-up questions, one recommendation per finding.
What to watch: Claude will try to see patterns in noise. The Performance Readout prompt has "refuse to declare a trend from fewer than two data points" baked in, keep that guardrail in the system prompt.
Job: Interviews the team lead, fills in a campaign brief, routes it for approval.
Stack: Claude Project where the system prompt IS the brief template (the Campaign Brief Generator prompt). Optional: Slackbot for intake. Start without Slack, just paste into Claude.
How: Team lead chats with the Project, answers Claude's questions, Claude assembles the brief, lead reviews, ships to writer.
The rule: Claude's job is to refuse to output the brief until every required section has real content. No placeholder briefs. No "we'll figure out the audience later."
If you build all five agents, they should all read from the same brand context. One voice guide, one positioning statement, one ICP doc. Pinned in every Project.
If you fragment the context (each agent has its own slightly different brand description) the agents will drift apart. The Content Repurposer will produce stuff that doesn't match what the Campaign Brief Copilot is briefing. Pick the source of truth (probably the output of the Voice Extractor) and have every agent reference it.
Claude Projects cover 80% of the agent use case. The remaining 20% is when you need:
For everything else, start with Claude Projects. Add complexity only when the simpler version has hit a real ceiling, not when it might hit one in theory.
One agent per month. Five months in, you've got a marketing operation that runs on autopilot for the parts that should run on autopilot, and reserves your time for the parts that genuinely need judgment.
The full Agent Pack is in the paid library, five complete blueprints with the system prompts, the stack recommendations, and the guardrail patterns. $159. If the calendar above looks right, the pack is the thirty hours of setup you skip.
Each agent's full setup: system prompt, stack recommendations, guardrail patterns, and the prompts the agent depends on. $159, lifetime access, 30-day refund.
See the Agent Pack →