Cindy Zhu.
Free guides
Building & apps Claude

Build a 37-agent company operating map with Claude

Hey, it's Cindy 🌱 This guide gives you the complete structure behind a six-department, 37-role map, plus the prompts to design it without creating 37 disconnected chatbots.
btw the waitlist for AI System Club is now open! If you're keen on using AI to automate your content, video editing, job search, business or build AI agents, feel free to sign up (only if you want to hehe). Ps. the waitlist gets first dibs:)
Start with three working agents, not 37. Prove the shared knowledge and routing system first, then expand only when a repeated workflow has a clear owner and measurable output 🌱
the architecture

What makes an agent team real

One company brain A versioned knowledge base holds policies, offers, customers, brand rules and definitions. Agents retrieve from it instead of inventing their own truth.
One role per agent Each agent owns a narrow outcome, such as qualifying a lead or preparing a weekly cash summary. A role is not “marketing assistant.”
Tools with boundaries Every role lists what it may read, draft or change. High-risk actions such as sending, paying, deleting or publishing require human approval.
A routing layer A coordinator receives the request, chooses the right role, passes the minimum context and records the result. It does not do every job itself.
the 37-role blueprint

Six departments and the jobs inside them

  • Communications, 6: inbox triage, meeting preparation, meeting recap, internal updates, customer-response drafting and calendar coordination.
  • Sales, 7: account research, lead qualification, call preparation, proposal drafting, follow-up drafting, pipeline review and win-loss analysis.
  • Finance, 5: invoice checking, expense categorisation, cash summary, variance explanation and renewal monitoring.
  • Clients, 6: onboarding, requirement capture, project-status reporting, feedback synthesis, risk escalation and renewal preparation.
  • Marketing, 7: audience research, content briefing, repurposing, campaign planning, performance analysis, SEO research and creative QA.
  • Technology, 6: bug triage, documentation, release notes, test planning, access review and incident-summary drafting.
These are role templates, not a claim that every business needs every agent. Remove anything without a recurring input, a repeatable decision and a useful output.
build sequence

Build the system in four passes

  1. Map repeated workflows List work that happens weekly or monthly. Record the trigger, inputs, decision rules, output, human owner and consequence of an error.
  2. Design the shared knowledge base Create folders for company, offers, customers, operations and brand. Give every document an owner and review date. Agents should cite the exact source they used.
  3. Build three low-risk pilots Start with research, summarisation and drafting. Keep sending, publishing, deletion, payments and account changes behind explicit approval.
  4. Add routing and observability Log the request, selected role, sources, tool calls, output, reviewer and correction. Expand only when the pilots are reliable.
prompt one

Turn your company into an agent org chart

Act as an AI operating-system architect. Interview me about my company, team, offers, customers and recurring workflows. For every possible agent role, define its trigger, approved inputs, decision rules, output, tools, human owner, approval gates and failure risks. Group the final roles into departments, remove duplicates and rank them by value, frequency and risk. Recommend only three low-risk agents for the first pilot and explain why those three should come first.
prompt two

Write one agent’s role card

Create a production-ready role card for the agent named [ROLE]. Its outcome is [OUTCOME], its human owner is [OWNER], and it may use only [TOOLS]. Include the trigger, required inputs, step-by-step SOP, source-of-truth documents, output schema, approval gates, prohibited actions, escalation conditions and a five-case test set. The agent must cite its sources, state uncertainty and stop before any irreversible action.
prompt three

Design the router without creating chaos

Design a coordinator that routes incoming work to the smallest suitable specialist agent. It must classify the request, check that required inputs exist, pass only relevant context, reject requests outside the approved catalogue and log every handoff. Provide the routing table, conflict rules, fallback path, human-approval conditions and the exact record saved after each run. Include examples for ambiguous requests and multi-department work.
quality gate

Test before you add another agent

  • Run five normal cases, two missing-input cases and three adversarial cases.
  • Confirm every factual output cites the shared knowledge source it used.
  • Check that the agent stops at approval gates instead of guessing permission.
  • Record corrections and update the SOP, not only the prompt.
  • Retire agents whose jobs overlap or whose output is not used.
Do not give 37 agents broad access on day one. More roles multiply permission risk, stale context and debugging work. Scale the catalogue only after the routing, logs and human review loop are working.
your first week

A practical rollout plan

Days 1 and 2 Map ten repeated workflows and choose the three lowest-risk, highest-frequency candidates.
Days 3 and 4 Build the knowledge folders, role cards, output schemas and approval gates.
Days 5 and 6 Run the test cases, log failures and tighten the SOPs.
Day 7 Put one pilot into real supervised use and measure time saved, correction rate and whether the output was actually used.
the waitlist is now open

AI System Club is launching soon

Build AI systems with me through recorded lessons you can follow at your own pace. Join the waitlist for first access and the founding price.

btw this list gets first dibs
the moment doors open :)

Privacy · Terms