# AI Teachers — Claude for Work: From Assistant to Agent 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 Turn scattered working notes into a reviewed weekly research or status brief; understand when to use chat, Cowork and more advanced agent features. ## What you will learn - Choose the right Claude mode - Create reusable project instructions - Evaluate skills, connectors and plugins - Delegate a multi-step deliverable - Test quality and approval boundaries ## Lesson 1: Choose chat, project or Cowork (25 min) The task, not the feature list, should determine your starting mode. ### Follow these steps 1. Open Claude on an account that supports your intended feature; verify availability in your own interface. 2. For a single question use chat; for reusable context consider a Project; for a multi-step deliverable explore Cowork. 3. Choose the weekly briefing example and make a three-column list of available sources, required output and approvals. 4. Create an empty sandbox/project workspace or use fictional notes, not real client files. ### Copyable prompt I need a recurring Friday review of three fictional project updates. Recommend whether to use Claude chat, a Project, or Cowork, based on required file handling, recurring context, review, permissions, and complexity. Provide the simplest working setup. ### Deliverable A mode decision and safe practice workspace. ### Quality gate You can explain why the chosen mode fits the task; unsupported options are marked optional. ### Watch out Cowork, connectors and local file access vary by account, device and administrator settings. My result: What I verified: What I will change: --- ## Lesson 2: Write project instructions that actually help (25 min) Reliable results start with a stable brief and clear boundaries. ### Follow these steps 1. Define the audience, scope, voice, source hierarchy, report format and success standard. 2. Ask Claude to first produce a plan and list what it cannot verify. 3. Use project instructions or a reusable note; keep date-specific material out of permanent rules. 4. Test the same brief against two different fictional updates. ### Copyable prompt You are my weekly project briefing assistant. Audience: a busy project sponsor. Use only the provided fictional notes. Sections: summary, completed work, blocked work, owner questions, next-week risks. Cite each note used. If a date, owner or outcome is missing, say UNCONFIRMED. Never send anything or modify source files. Ask for my approval before any external action. ### Deliverable Versioned instructions, a short sample brief and gap log. ### Quality gate The report distinguishes evidence from suggestion and passes a second sample. ### Watch out Do not hide missing context or silently infer ownership. My result: What I verified: What I will change: --- ## Lesson 3: Add skills, plugins and connectors deliberately (30 min) A packaged workflow is useful only when tools and permissions match a real need. ### Follow these steps 1. List data the workflow truly needs and which approved source owns it. 2. Inspect supported Skills/plugins and connector permissions in your Claude account; features vary. 3. Draft one reusable 'project update analyst' skill with input schema, output format and checks. 4. Run it first without external write access; compare before and after for improvement. ### Copyable prompt Draft a reusable project-update skill specification. Inputs: weekly notes and a project glossary. Outputs: status summary, risk list, source list, missing-information questions. Include minimum access scope, approval rules, three evaluation cases, and a version/date line. Make no claims that reading a prompt enforces security. ### Deliverable Skill specification and least-privilege access checklist. ### Quality gate The skill can be run manually without installing additional software. ### Watch out Plugins may bundle subagents and connectors; install only trusted components you understand. My result: What I verified: What I will change: --- ## Lesson 4: Delegate a complete reviewed report (30 min) A complete outcome is more valuable than an impressive-looking prompt. ### Follow these steps 1. Supply the approved fictional source pack and your reusable brief. 2. Ask Claude to plan the report, draft a one-page narrative and a structured action table. 3. Inspect the document or exported result; compare numbers, owners and dates with sources. 4. Request fixes and save the revised file under a clear version name. ### Copyable prompt Using the provided fictional sources, prepare a one-page executive report plus a table of action items. In the report distinguish confirmed facts, interpretation, and open questions. Before finalizing, run a second pass specifically looking for unsupported claims, mismatched dates, or duplicate tasks. Never send the report. ### Deliverable One checked brief, action tracker and a list of corrections. ### Quality gate No unverified fact is presented as a confirmed result. ### Watch out AI-created documents and tables still require actual human review. My result: What I verified: What I will change: --- ## Lesson 5: Make the next run reliable (30 min) Repeatability requires an error plan, not just saved instructions. ### Follow these steps 1. Test a complete update, one missing status and a contradictory due date. 2. Log output quality and review time for each case. 3. Add a stop rule, approval step and escalation message. 4. Only after testing, explore scheduled tasks or supported automated handoffs. ### Copyable prompt Create an operating runbook for this weekly reporting workflow. Include purpose, trigger, permitted sources, step-by-step execution, quality rubric, approval owner, failure messages, duplicate prevention, log fields, cost considerations, and a 7-day improvement plan. ### Deliverable A repeatable, human-reviewed runbook. ### Quality gate The workflow remains safe if a source is unavailable or a connector fails. ### Watch out A scheduled instruction is not a guarantee that a service action ran successfully. 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: - Claude Cowork help: https://support.claude.com/en/articles/15520349-use-claude-cowork-on-web-desktop-and-mobile - Claude plugins help: https://support.claude.com/en/articles/13837440-use-plugins-in-claude AI Teachers · https://aiteachers.cc/academy/claude-for-work/