20 AI Tricks
You Can Use Today

Instead of reading every email yourself, tell Claude to scan your entire inbox, identify which emails actually need your reply, draft responses for each one, and generate a dashboard showing everything at a glance. What used to take 30–45 minutes becomes a single command.

Walking into a meeting unprepared? Claude can pull your calendar, check your email history with each attendee, find relevant documents, and produce a one-page briefing with background on who you're meeting, what you discussed last time, your open action items, and suggested talking points.

Drop a folder of scanned PDF receipts onto your desktop. From your phone, tell Claude to turn them into a categorized expense report. It reads every receipt, extracts amounts and vendors, categorizes spending, and generates an interactive HTML report with charts and breakdowns.

Create an empty folder. Tell Claude to set itself up as your personal knowledge assistant with an orchestrator agent, a local SQLite database for structured data, and a simple HTML interface to browse it. You now own 100% of your data, it works offline, and you can swap AI models anytime.

Claude Dispatch lets you fire off multiple independent tasks from your phone. Each runs as a separate parallel agent on your desktop. While you're at the gym, one agent sweeps your inbox, another researches a topic, a third preps your meeting, and a fourth generates a presentation.

Claude's native connectors cover about 38 apps. But using Zapier's MCP server, you can extend that to 8,000+ apps with 30,000+ actions. Connect to School, Airtable, HubSpot, Stripe, or any app Zapier supports — then trigger actions from Claude or even from your phone via Dispatch.

Dreaming is the biggest shift since agents themselves. Between sessions, an idle agent reviews everything it did in its last jobs, pulls out the patterns — what worked, what kept failing — and writes new memory entries and playbook files the next session can use. Anthropic compares it to how your brain consolidates memory during sleep. Legal AI company Harvey reported task completion rates up roughly 6x after enabling it. Your agents stop making the same mistake twice — without you writing a single instruction.

Stop writing longer prompts and start writing rubrics. With Outcomes, you define what a successful result looks like, and a separate grader — running in its own context window, uninfluenced by the agent's reasoning — evaluates the output against your criteria. If something falls short, the grader pinpoints what, and the agent takes another pass. Wisedocs cut document review time by 50% with this. It's the single best answer to "the output is almost right but not quite."

Agents no longer need you to press the button. Routines let you schedule recurring agent work: a 7am daily briefing that's waiting when you wake up, a Friday afternoon weekly report, a Monday morning pipeline review. Combined with Dreaming, a routine agent gets better at its recurring job every single week — it's the compounding loop I keep writing about, running on autopilot.

This wasn't possible in March. Claude Code can now write its own orchestration script — breaking a big job into subtasks, fanning out tens to hundreds of parallel subagents, and adversarially verifying results before anything reaches you. The proof point: Bun's creator used dynamic workflows to port the entire runtime from Zig to Rust — roughly 750,000 lines, 99.8% of the test suite passing, eleven days from first commit to merge. For your world: full-codebase audits, mass content reviews, bulk data migrations.

Claude Code agents can now spawn child agents up to three levels deep: a lead agent delegates to per-area leads, which delegate to per-file workers. Each specialist gets its own context, prompts, and tools; results flow back up the tree. The practical difference from flat parallelism: the middle layer can coordinate, dedupe, and quality-check its workers before anything hits the lead — like a real org chart, minus the meetings.

Claude Code sessions can now publish their output as a live Artifact page on claude.ai — one that updates in place as the session keeps working. Kick off a long-running analysis, share the link with your team, and they watch the report fill in as the agents progress. No more "I'll send you the results when it's done" — the results are a URL from minute one.

Claude accumulates memory of your conversations over time. Ask it to create tone-of-voice guidelines based on what it's learned about how you communicate. This becomes a reusable markdown file that every AI skill references, ensuring all content sounds like you.

Claude Design now sticks to your design system across projects: feed it your colors, type, and components once, and every asset it generates stays on-brand. You can edit directly on the canvas instead of re-prompting, and it works fluidly with Claude Code — design a component visually, then have Code wire it into your actual site. For the 80% of design work that doesn't need a professional, this replaces the entire brief-wait-revise cycle.

When AI extracts data from documents, it guesses instead of admitting it doesn't know. Fix this by explicitly giving AI permission to leave fields blank and requiring an explanation for each blank. You only review blanks instead of everything.

AI defaults to guessing because it treats a wrong answer the same as a blank answer. Add a single line that changes the incentive: tell AI that a wrong answer costs 3x more than saying "I don't know." This dramatically reduces hallucinations. Like telling a new hire: "If you give me wrong info it costs the company 3x more than just saying you'll check."

That's it. One line. Add it to any extraction, analysis, or review prompt. Combine with Trick #15 for maximum effect.

Even after you tell AI to only extract from the document, it will start inferring on complex tasks. This safety net catches it. Require a "Source" column on every field with two possible values: Extracted (word-for-word from the document) or Inferred (derived from context). For inferred values, require a one-sentence evidence explanation. Now you only need to review the inferred fields.

The complete anti-hallucination stack: Combine Tricks #15 + #16 + #17 in every data extraction prompt. You'll go from checking everything to only checking blanks and inferred fields.

Before Claude does anything complex, switch to Plan Mode. Instead of executing immediately, it asks clarifying questions, shows you a step-by-step plan, and waits for your approval. This prevents expensive mistakes, runaway agents, and "overkill" implementations. In Claude Code, press Shift+Tab to toggle. In Cowork, add "plan first, don't execute yet" to your prompt.

Two model moves worth acting on. First: Sonnet 5 launched in June as the most agentic model yet, with introductory pricing ($2/$10 per million tokens) through August 31 — go re-run the automations that failed for you in February; several will just work now. Second: Fable 5, the new tier above Opus, went globally available July 1. The trick is routing, not loyalty: cheap fast models for mechanical work, Sonnet 5 as your default, and Fable 5 reserved for the genuinely hard reasoning — the gnarly contract, the architecture decision, the analysis you'd otherwise hire for.

Anthropic shipped Claude Finance — a package of ten predefined agents for financial workflows: a pitch builder, meeting preparer, market researcher, evaluation reviewer, and month-end closer, among others. Built for financial services, but here's the SMB trick: the month-end closer and market researcher work just as well for a normal company's books and competitive landscape. You're getting workflows that a consultant would charge five figures to design, pre-built.