# AI Teachers — Build Your Own AI Agent Team Updated 2026-10-25 · Free written training workbook Use invented or permitted data. Your AI tool may require a separate subscription. Download this workbook and save answers locally; this course site saves only completion checkmarks. ## The outcome Coordinate a practical team of specialized AI roles around one small project, with evidence, handoffs, explicit authority and a tested review loop. ## What you will learn - Choose useful agent roles - Design reusable briefs and context - Create reliable cross-agent handoffs - Evaluate and handle failure cases - Complete one human-approved team project ## Lesson 1: Build the team around an outcome (30 min) A team is an operating model, not a collection of impressive character names. ### Follow these steps 1. Choose a fictional business project: a one-page landing page for a new service. 2. List the decisions and deliverables needed from discovery to release. 3. Assign roles: chief of staff, researcher, marketer, designer, developer and reviewer. 4. Keep a single human product owner responsible for quality and external actions. ### Copyable prompt Design a small AI team for a fictional service landing page. Define a mission, five specialist roles, their inputs and outputs, decision rights, approval gates and a RACI-style owner table. Remove any role without a clear unique contribution. One human approves publication and spending. ### Deliverable Team charter, responsibility chart and deliverable plan. ### Quality gate No task lacks an accountable owner; no agent can publish autonomously. ### Watch out Multiple roles can add overhead; start with the smallest team that improves quality. My result: What I verified: What I will change: --- ## Lesson 2: Create role briefs and shared context (40 min) Specialists need the same facts and different responsibilities. ### Follow these steps 1. Make a short shared project brief containing audience, promise, constraints and evidence. 2. Write individual role cards with purpose, required input, output and limits. 3. Give the researcher source verification, marketer messaging, designer UX and developer implementation tasks. 4. Write a reviewer role explicitly empowered to reject bad handoffs. ### Copyable prompt Create reusable role briefs for Researcher, Marketer, Designer, Developer and QA Reviewer. Every card needs mission, permitted sources, input/output schema, success criteria, no-go actions, and when to ask a human. Include one shared context document that distinguishes known facts from hypotheses. ### Deliverable Five role cards and one source-of-truth brief. ### Quality gate Each role can work independently without inventing project data. ### Watch out An AI persona or system prompt alone does not create a separate autonomous tool. My result: What I verified: What I will change: --- ## Lesson 3: Orchestrate safe handoffs (45 min) A handoff should preserve evidence, owners and unresolved questions. ### Follow these steps 1. Define the order: researcher → marketer → designer → developer → reviewer. 2. Create a structured handoff with objective, sources, decisions, deliverable and unknowns. 3. Run one manual handoff using fictional customer interviews; don't assume automatic multiagent coordination. 4. Stop the next agent if the incoming package has no evidence or acceptance criteria. ### Copyable prompt Design a handoff contract between the research and marketing agents. Fields: objective, evidence with citations, audience hypotheses, confirmed claims, uncertainties, recommended message, acceptance criteria, reviewer, and approval status. Show a complete fictional example and a rejected incomplete example. ### Deliverable Tested handoff schema and one rejected weak handoff. ### Quality gate No downstream role silently promotes speculation to fact. ### Watch out Automating handoffs without observability can amplify errors. My result: What I verified: What I will change: --- ## Lesson 4: Test the team like a product (45 min) A good team handles missing and conflicting data without producing confident nonsense. ### Follow these steps 1. Test normal input, missing data, contradictory notes and a malicious instruction embedded in a document. 2. Measure source accuracy, task completion, human review time and cost. 3. Check that permissions, approval and stop rules are enforced in the selected platform. 4. Revise one role brief based on an observed failure. ### Copyable prompt Develop a six-case evaluation matrix for the team: normal, missing evidence, contradiction, hostile source instructions, connector failure and duplicate request. For each case define expected behavior, unacceptable result, human escalation and an observable pass/fail condition. ### Deliverable Evaluation matrix, run log and improved role brief. ### Quality gate Failures are reported to an owner instead of covered up. ### Watch out Prompt instructions are guidance; platform controls, permissions and approvals enforce actual boundaries. My result: What I verified: What I will change: --- ## Lesson 5: Ship a small human-approved result (45 min) The capstone is one finished, checked business artifact—not five agent conversations. ### Follow these steps 1. Use your team to prepare a fictional service landing-page brief. 2. Hand off a sourced positioning statement, wireframe outline and build specification. 3. Run the reviewer and correct at least one finding. 4. Present the package for human go/no-go; record the next experiment. ### Copyable prompt Coordinate a complete, nonpublished landing-page package for the fictional service using the role briefs and handoff schema. Deliver a sourced audience summary, positioning statement, page outline, copy draft, test plan, accessibility checklist, risk log and final human approval checklist. Label hypotheses as hypotheses. ### Deliverable A coherent, reviewed multi-role project package. ### Quality gate Each handoff has evidence; final approval is clearly human-owned. ### Watch out Do not make real claims about customers, conversion or revenue from fictional exercises. My result: What I verified: What I will change: --- ## Final review - [ ] Every material fact has a permitted source. - [ ] Missing data is visible. - [ ] External actions still require authorized approvals. - [ ] I have tested at least one failure case. - [ ] I know who owns the result and how to stop the workflow. Official reference links: - AI Teachers Agent to Outcome: https://aiteachers.cc/agent-to-outcome/ - Claude plugins: https://support.claude.com/en/articles/13837440-use-plugins-in-claude - ChatGPT Work and Codex: https://help.openai.com/en/articles/20001275-chatgpt-work-and-codex AI Teachers · https://aiteachers.cc/academy/build-your-ai-agent-team/