Claude Masterclass: Prompts to AI Operating System

Master Claude's Model Hierarchy for Task Efficiency

Claude offers three models—Opus, Sonnet, Haiku—each optimized for specific workloads. Opus handles deep reasoning like coding or complex planning but consumes more tokens and runs slower. Sonnet serves as the daily driver for 90% of tasks, balancing speed and intelligence. Haiku excels at quick, bulk operations where depth isn't needed. Default to Sonnet; escalate to Opus for shallow responses and drop to Haiku for speed. Toggle 'extended thinking' for step-by-step reasoning on tough problems, but avoid on free/Pro plans due to cost and time.

Practical rule: Match model to job to avoid waste. In Alex's P&L analysis, Sonnet suffices for data synthesis and charts; Opus only if reasoning falters. This prevents overkill—Sonnet processed credit card statements, ad receipts, and revenue data into pie charts showing Meta ads dominating expenses (27k spend) versus YouTube's efficiency (1.35k spend yielding 9.6k leads).

Voice input accelerates prompting: Use Claude's built-in voice mode or tools like Whisper Flow (hold key to dictate anywhere—docs, terminals, chats). Typing slows thinking; voice captures fluid ideas, cutting prompt creation time.

Plans enforce paced usage: Free for basics, Pro ($20/mo) unlocks Co-work/artifacts, Max ($100-200/mo) for heavy lifting (equivalent to $3-5k API spend). Monitor via claude.ai/upgrade or platform.anthropic.com/usage; space queries to dodge timeouts.

Build Persistent Context with Projects and System Prompts

Projects centralize work: Create via sidebar > New Project, name it (e.g., 'PNL for Boss'), add custom instructions (system prompt), and upload files. System prompts prepend every chat message, embedding role, tone, and rules—write once for consistent outputs.

Alex's prompt: "I'm the marketing manager at a B2B SaaS company. Lead with numbers, then reasoning. Bullet points only. Recommend boldly, no hedging. Visualize data. Match brand voice: direct, confident, no fluff."

Upload scattered data (credit cards, ad platforms, CRM exports)—Claude ingests PDFs/CSVs instantly. Enable Memory (Settings > Capabilities) for cross-chat recall of preferences/projects; toggle Artifacts for side-panel outputs (charts, decks, tools) over inline text.

This setup transforms ad-hoc chats into role-aware workspaces. Alex prompts: "Break down P&L from uploaded data—pie charts for revenue/expenses." Claude delivers interactive visuals: revenue from monthly subs dominant, expenses Meta-heavy. Follow-up: "Rank channels by leads per spend." Reveals YouTube/Instagram organic outperform paid—Instagram/blog/YouTube for doubling down, cut Meta.

Key principle: Context-first prompting scales analysis. Without projects, repeat instructions; with them, Claude knows your SaaS context, brand (pull from bookend.ai), and style automatically.

Generate and Share Production-Ready Artifacts

Artifacts turn insights into polished deliverables: Prompt for decks/tools; Claude builds in side-panel (React-based interactivity). Alex: "Build presentation on P&L, findings, recommendations per brand guidelines." Outputs branded Google Slides-ready deck: agenda, snapshots, story flow (e.g., 'Cut Meta, boost organic').

Elevate to interactive: "Build budget reallocation tool—sliders for Q1 spend vs. projected leads/conversions/CPA, CEO-playable, branded." With extended thinking on, Sonnet codes sliders projecting real-time impacts (e.g., shift from paid to organic drops CPA).

Share via Publish > Web link—embeddable widget, no code needed. Claude Club example: Interactive guides as artifacts for community step-by-steps.

Common pitfalls: Stuck on one model mid-chat (chat locks it)—switch via new chats or Co-work (Level 2). Free plan limits artifacts/Co-work. Update OS for desktop app (80% course here)—browser for quickies only.

Before: Manual data hunt (20-30min/week), static Excel. After: One project prompt yields charts/decks/tools, export to Slides/Docs. Criteria for good artifacts: Interactive, branded, actionable (numbers lead, visuals explain), shareable.

Automate Repetition with Co-work Transition

Weekly reports expose chat limits: Regather data, repeat prompts. Solution: Level 2 Co-work (desktop app, Pro+ required)—persistent workspaces for automation.

Alex's weekly ask: CEO wants Monday P&L decks. Co-work fixes data ingestion/prompt repetition, evolving to agents (later levels) for full AI ops.

Build alongside: Download desktop (bottom-left icon, drag to apps), log in, switch via top-left (Claude/Chat > Co-work/Code). Settings: Usage bars, Memory/Artifacts on.

Broader workflow: Level 1 proves one-offs (chat/projects/artifacts); Level 2+ scales to ops (Co-work agents replace hated tasks). Prerequisites: Free account, recent OS. Practice: Replicate Alex's P&L project with your data.

"Most people use only 10% of Claude—typing a message and closing. We're building AI that runs operations."

Notable quotes:

• "Use Sonnet, escalate to Opus when shallow, Haiku for bulk." (Model selection rule, early setup.)
• "System prompt goes first—sets tone/rules before your message." (Projects explanation, persistent context.)
• "Artifacts: Documents, decks, diagrams on side-panel, not dumped text." (Settings toggle value.)
• "Pace usage—$200 Max = $3-5k API value." (Plans ROI, upgrade guidance.)
• "Voice is faster than typing—think clearer." (Whisper Flow recommendation, productivity hack.)

Key Takeaways

• Download Claude desktop app immediately—unlocks Co-work, Code, power tools; keep OS updated.
• Default Sonnet model; toggle extended thinking sparingly for complex builds.
• Every project needs a system prompt: Define role, style, rules once for all chats.
• Upload all data to projects—prompt for visuals/insights/decks to skip manual analysis.
• Build/share artifacts for presentations/tools: Interactive sliders > static slides.
• Enable Memory/Artifacts in settings; upgrade to Pro for scaling beyond one-offs.
• Use voice (Whisper Flow) for faster, clearer prompts.
• Practice Alex's flow: P&L project > charts > deck > interactive tool.
• Pace queries to avoid timeouts; monitor usage.
• Build alongside course—10 levels compound to AI workforce replacing busywork.