20 AI Tricks You Can Use Today

20 AI Tricks
You Can Use Today

Refreshed for July 2026. The evergreen tactics that still pull their weight, plus nine new tricks that simply weren't possible in March — dreaming agents, outcome rubrics, dynamic workflows, and more.

Each trick includes the exact prompt to use and how to set it up. The "New in July" tag marks what shipped since the March edition. Organized into seven categories: Productivity, Mobile & Remote, Self-Improving Agents, Orchestration at Scale, Content & Design, Prompting & Quality, and Models & Business.

Productivity & Everyday Automation

Inbox Sweep: Triage 100+ Emails in One Command

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.

Prompt

How to set this up

Connect Gmail to Claude via Connectors (Settings → Connectors → Gmail)
Open Cowork and select your working folder
Paste the prompt above — or just say "inbox sweep" if you have the skill installed
Review the HTML dashboard that appears in your folder

Meeting Prep Briefing in 60 Seconds

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.

Prompt

How to set this up

Connect Google Calendar and Gmail via Connectors
Optionally connect Google Drive for document access
Run the prompt before your next meeting (or trigger via Dispatch from your phone on the way there)

Receipt Scanner to Instant Expense Report

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.

Prompt

How to set this up

Create a receipts folder in your Cowork project
Scan or save your receipts as PDFs into that folder
Run the prompt from Cowork or Dispatch

Replace Notion with a Folder + Claude

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.

Prompt

How to set this up

Create a new empty folder on your desktop (e.g., "PKA")
Open it in Claude Code (cd ~/Desktop/PKA && claude) or select it in Cowork
Paste the prompt and let Claude set up the entire structure
Drop files into the Team Inbox — Claude auto-organizes them

Mobile & Remote AI

Run 5 Tasks in Parallel from Your Phone

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.

Prompt

How to set this up

Set up Dispatch (Cowork → Dispatch → scan QR code with phone)
Keep your computer awake (System Settings → Energy → Prevent sleeping)
Send the prompt from the Claude mobile app
Each task runs independently — results appear on both phone and desktop

Connect Claude to 8,000+ Apps via Zapier MCP

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.

How to set this up

Go to Zapier → MCP → Start Building → New MCP Server
Select "Claude" as your client
Search and add tools for any apps you want (e.g., School, HubSpot, Airtable)
Copy the generated URL
In Claude: Connectors → Zapier → paste the URL
Select "Always allow" for the tools you added

Self-Improving Agents — New for 2026

Let Your Agents Dream — and Write Their Own Playbooks

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.

Prompt

How to set this up

On Claude Managed Agents: enable Dreaming in your agent settings — it runs on a schedule while agents are idle
Review the generated playbook .md files weekly and prune anything stale
Solo users: use the DIY prompt above — same principle, manual trigger

Grade the Work, Not the Prompt: Outcomes

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."

Prompt

How to set this up

On Managed Agents: define an Outcome rubric in the agent config — the independent grader loop is built in
Everywhere else: use the DIY prompt — self-grading against an explicit rubric catches most of the gap
Keep rubrics to 3–5 criteria. More than that and the grader's signal gets mushy

Put Your Agents on a Schedule with Routines

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.

Prompt

How to set this up

In Cowork: set up a Routine with a schedule and a standing prompt
In Claude Code: cron-style scheduled sessions do the same job
Start with one routine you'd otherwise do manually every day — the morning briefing is the classic

Orchestration at Scale — New for 2026

Dynamic Workflows: Hundreds of Agents, One Command

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.

Prompt

How to set this up

Update Claude Code and just describe the fan-out you want — "use a workflow" opts you in
Best fits: jobs that decompose into many similar pieces (per-file, per-module, per-record)
Always ask for a verification stage — parallel agents are fast, adversarial checking is what makes them trustworthy

Hierarchical Agent Teams, Three Levels Deep

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.

Prompt

How to set this up

Describe the org structure in your prompt — lead, coordinators, workers
Give the middle layer an explicit review job, not just routing
Watch the audit trail to see what each sub-agent did, in what order

Turn a Working Session into a Live, Shareable Dashboard

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.

Prompt

How to set this up

Currently in beta on Team and Enterprise plans
Ask the session to create an artifact and keep updating the same page
Artifacts are private by default — you choose when to share the link

Content, Design & Brand

Clone Your Writing Voice from AI Memory

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.

Prompt

How to set this up

Have several conversations with Claude first so it learns your style
Run the prompt above to generate the guidelines
Review and refine the output
Save it as a skill markdown file so all future content references it

On-Brand Design Without a Designer: Claude Design

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.

Prompt

How to set this up

Find Claude Design in the sidebar of the Claude desktop app
Upload your brand kit or design tokens once — it persists across projects
Edit results directly on the canvas instead of re-prompting from scratch
Hand finished designs to Claude Code to implement on your site

Prompting & AI Quality Control

Force Blank Answers to Kill Hallucinations

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.

Prompt

The "3x Penalty" Trick: One Line That Changes 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."

Prompt

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

Source Tagging: Catch AI When It Infers Instead of Extracts

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.

Prompt

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.

Plan Mode: Make AI Explain Before It Acts

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.

Prompt

Models & Business

Right-Size Your Model: Sonnet 5 by Default, Fable 5 for the Hard Stuff

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.

How to set this up

Keep a "failed prompts" folder — re-run it on every major model release
Default to Sonnet 5 for daily agent work (it's the Free/Pro default now anyway)
Escalate to Fable 5 only when the task is worth the premium — hard reasoning, high stakes
Review your routing quarterly: yesterday's premium task is today's default-tier task

Claude Finance: 10 Pre-Built Agents for Your Back Office

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.

Prompt

How to set this up

Available as part of Claude's managed agent offerings — start with one agent, not all ten
The meeting preparer and market researcher generalize far beyond finance
Feed each agent one past example of the output you want (last month's close, last quarter's research)