Build Scheduled AI Agents with Claude Co-Work

Prioritize Co-Work Over Chat or Code for Production Workflows

Claude offers three interfaces powered by the same models: Chat for one-off Q&A and brainstorming, Code for terminal-based power users handling software builds and MCP integrations, and Co-Work for visual, accessible automation of multi-step tasks. Co-Work stands out because it opens local files, connects to apps without command-line setup, plans complex processes, runs on schedules, and delivers outputs to folders—ideal for non-technical users scaling business operations. The speaker admits initially overlooking it but now runs client businesses on it, emphasizing its edge for recurring deliverables over Chat's passivity or Code's technical barrier.

Principle: Match interface to use case—start with Co-Work for 80% of automation needs, graduate to Code later. Download Claude Desktop (free at claude.ai/download) on Pro, Max, Team, or Enterprise plans. It's labeled 'research preview,' so features evolve, but core reliability supports production use.

Common mistake: Sticking to Chat for anything beyond questions, missing Co-Work's ability to execute autonomously. Trade-off: Less flexible than Code but zero-config for most users.

"Claude co-work, this is the one that actually opens up your files. It pulls your data from your applications. It runs on a schedule, and it drops finished deliverables into a folder on your computer."

Leverage Connectors, Skills, and Plugins as Agent Building Blocks

Access extensibility via the Customize sidebar: Connectors act as the agent's 'hands' for app integrations (e.g., Gmail, Google Calendar, Outlook, Drive, Slack, Notion, Apollo, DocuSign, Fireflies, Ticket Taylor). One-click OAuth or API key setup with granular permissions (read-only vs. read/write). Anthropic adds connectors regularly; extend via MCP (Model Context Protocol) for custom apps.

Skills are reusable prompt recipes: Build a task once, save it (e.g., 'morning brief'), invoke by name like "run morning brief." Accumulate a library mirroring your business processes. Use built-in '/skill creator' slash command—it analyzes working tasks, drafts YAML-wrapped skills, generates test prompts (e.g., "Generate my morning brief," "Refresh daily brief"), self-heals errors, and tests scenarios.

Plugins bundle skills + connectors into playbooks for roles like sales or content (e.g., Anthropic's Clockwork for sales/finance). Non-technical: Hand a human a playbook; Co-Work executes it.

Principle: Layer from connectors (access data) → skills (reusable logic) → plugins (packaged workflows). Avoid context bloat—keep instructions concise to preserve token limits.

Quality criteria: Skills must handle trigger variations, edge cases (e.g., no meetings), and iterate via conversation. Test with real data.

"Connectors are effectively how co-work touches any of your applications... you can think of them as the agent's hands."

"Skills, these are going to be the reusable recipes... you can just say like, okay, run the morning brief skill."

"A plugin is that playbook, except co-work executes it instead of a person."

Construct and Schedule Agents via Projects and Prompts

Start with a Project (workspace tied to a local folder for read/write isolation—security model limits visibility). Steps for morning briefing agent:

1. Create project 'daily briefs' in a folder (e.g., AI/co-work-demo). Add optional instructions.md for context (company name, team info—keep <400 words).
2. Enable connectors: + icon → Google Calendar (OAuth login, toggle read access).
3. Prompt precisely: "Read Google Calendar today (now-11:59pm local). For external meetings, web search attendee + company for 2-3 facts (funding, news, LinkedIn, company info)—skip internals. Top 3 AI news (last 24h, prioritize Anthropic/OpenAI/DeepMind/launches, no rumors). Output 'today-brief.md' in project folder: structured Markdown (## Meetings, ## AI News), <400 words. Open in Markdown viewer."
4. Monitor: Progress shows plan/steps; Context logs sources (e.g., 20 web results, calendar pull). Review output (e.g., Sunday: no meetings, curated news like 'Google Gemma launch'). Iterate: Tweak formatting via chat.
5. Save as skill: '/skill creator' → "Turn morning briefing into 'morning-brief' skill." It drafts, tests (with/without meetings), adds features (e.g., "Add unread emails from yesterday—triage important ones").
6. Schedule: Scheduled sidebar → Set daily 6:45am. Test run confirms delivery (Slack/email optional via connectors).

Model choice: Opus for complex tasks (best reasoning, higher cost/time); Sonnet/Haiku for simple. Invoke skills casually ("What's on my plate today?") for natural triggers.

Extend pattern: Apply to KPIs, QBRs, pipeline checks. Dispatch from phone (QR scan app). Ideas sidebar sparks automations (sales playbooks, data analysis).

Mistakes to avoid: Over-prompting (causes bloat/delays); skipping tests (misses edges); broad permissions (security risk). Before: Manual email/calendar checks. After: Autonomous brief ready pre-wakeup.

Prerequisites: Claude Pro+ subscription, basic prompting. Fits early in workflow: Prototype in Chat → Build in Co-Work → Scale with Code.

"First, read my Google Calendar for today only... And then search the web for the top three AI news stories from last 24 hours... write a single markdown file called today brief."

Key Takeaways

• Download Claude Desktop and switch to Co-Work tab for visual automation—skip Chat for anything executable.
• Connect apps via one-click connectors with read-only perms; start with Calendar/Gmail/Slack.
• Build once, reuse forever: Prompt → iterate → '/skill creator' for YAML skills with test cases.
• Structure projects as isolated folders; keep instructions.md lean to avoid token waste.
• Schedule via sidebar for daily/weekly runs—test immediately to verify outputs.
• Bundle into plugins for team handoffs; extend with MCP for custom tools.
• Use Opus for reasoning-heavy tasks, but optimize costs with lighter models.
• Pattern-match: Calendar + research + news → brief; adapt for emails, KPIs, triage.
• Monitor progress/context logs; converse to refine before skill-ifying.
• Phone dispatch + ideas sidebar accelerate on-the-go starts.