Cheat sheet

A Day with Claude, on one page.

Everything we covered, distilled to what you'll actually reach for. Print it, pin it, or keep the tab open while you build.

The four ways to work with Claude

1 API

Claude wired straight into your own software, no person typing in between.

Built for: developers & automations

2 Chat

Claude.ai in your browser or phone. Everyday thinking, writing, analysis.

You give it: a Master Prompt

3 Cowork

Works inside a folder on your computer. Turns a pile of files into output.

You give it: folder instructions

4 Code

Your whole machine and the web. Big builds it can research and ship.

You give it: a CLAUDE.md

Talking to Claude well

›Give it context, not clever words. The shift from prompt engineering to context engineering: hand over the brief, the files, the goal, and examples of good and bad. The answer is only as sharp as what you give it.
›Talk, don't type. Dictating is about 3× faster and gives the full thought, not a rationed keyword. (Wispr Flow: press a key, speak, clean text lands anywhere.)
›Treat it like a colleague. Co-define the task before you delegate, stay in the messy middle while it's fuzzy, then say go.

Know its memory cutoff. Every model is trained up to a date and only learned what was common. For anything recent, niche, or obscure, make it search instead of recalling.

"Search the web for this and cite your sources. Don't answer from memory."

Your Master Prompt: tell Claude who you are, once

1Your name and what you do

2Years in your field, where you've worked

3The two or three projects that define you

4What you're working on right now

5The biggest problem you're solving today

6Your strongest belief about your work

7The tone you want, and the one you don't

8The role Claude should play: assistant, partner, coach

9The habits that work against you, to flag

10Who you work with and for

11The language you work in

Answer the eleven out loud (or hand Claude your pre-work voice note), then have it write your Master Prompt and paste it into your profile. A living document: refresh after any big change.

Answer 11, by voice→ Claude writes it→ Paste into profile

Build once, then reuse

Project

Give Claude one job plus its resources. You've built a specialist.

Skill

A folder of instructions Claude loads when relevant. Portable, invoke it anywhere.

Chain skills

One skill's instructions can call another, so small pieces become big jobs.

Plugin

Pack many skills (plus hooks, connectors, commands) into one installable, shareable, versioned unit.

Routines (Claude Code): save a task once and let it run by itself in the cloud, on a schedule. Great for a morning briefing or a weekly report. A "done" status means it ran, not that it nailed it: spot-check.

Other models worth knowing

Gemini

Grounded in Google Search and YouTube, pulls many sources. Best when the answer changes weekly: deep research.

ChatGPT

Strong all-rounder, still a go-to for image generation.

NotebookLM

Answers only from sources you upload, with citations. Audio Overview turns docs into a podcast. Free with Google.

Two researchers beat one (triangulation): run the same deep-research prompt on Claude and Gemini, then ask Claude to reconcile the two: where they agree (trust it), where only one found something (a real lead), and where they clash (dig in by hand, errors hide there).

Trust and verify: minimizing hallucinations

›Make it search, not recall. For anything factual, recent, or niche, ground it on live web sources, not memory.
›Cross-reference. Run the same task on another model and compare.
›Spot-check the data. Trace 5 to 10 generated cells back to the source.
›Ask for sources. "Where did you get this?" and make it cite the passage.
›Make it check its own work. List each claim, verify one by one, flag the unsure ones.
›Keep a human in the loop on anything high-stakes.