50 Prompt tricks are the small part

I use Claude every day and still felt like a tourist

July 17, 2026

Sponsored by

Type a prompt, get an answer, move on. That has been my loop for months. It works, which is exactly the problem. It works well enough that I never went looking for the rest of it, even though I could always hear a bigger machine humming somewhere behind the chat box.

Then I came across a creator who had crammed 50 Claude tips into one enormous infographic, and it landed harder than it should have. The thing separating casual users from power users is not talent. It is a list of features nobody ever bothered to teach as one system.

How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

*Ad

Their framing is what stuck with me:

They're not using the wrong AI. They're using it like a search engine.

That is the whole thing in two sentences. Extended thinking, MCP connectors, memory, Projects, Cowork, Claude Code. None of it is hidden. It is all sitting right there in the interface. It just gets explained one blog post at a time, so you never see the shape of it.

So I pulled the twelve tips the author flagged as the ones that changed their work most, and I sorted them by where the gain actually comes from. Three layers. The smallest one is the one everybody talks about.

Layer one: the prompt tricks, and there are only three

This is the layer the entire internet is obsessed with. It is also the shortest section in this email, and that is the point.

Give Claude a role, a goal, a format, and a constraint. Four ingredients, every time. "You're a technical editor. Rewrite this changelog for non-engineers. Return it as bullet points. Keep it under 100 words." Output quality jumps on the first try. If you do one thing from this whole email, do this one.

Add an anti-guessing line when accuracy matters. Paste this in: "If you are unsure, say so explicitly. Do not guess." It gives Claude permission to admit uncertainty instead of filling the gap with something that sounds confident and is wrong. Use it for anything with numbers, dates, or facts you cannot check yourself.

Use XML tags to separate the parts of a long prompt. Wrap your source material in one set of tags and your instructions in another. Claude never mixes them up. This is the trick that makes a prompt with a pasted document actually work instead of turning into soup.

That is it. Three tips. Learn them in about four minutes and you are done with this layer forever.

Stop typing what you could say in 10 seconds.

Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.

*Ad

Layer two: the setup nobody screenshots

Here is where the compounding lives. Every item below is something you configure once and then stop thinking about, which is precisely why nobody makes content about it.

Match the model to the task instead of your habit. Haiku for speed, Sonnet for daily work, Opus when accuracy is not negotiable. Most people pick one and never switch, which is like owning three cars and only ever driving the one in the middle. Quick reformatting or classification? Haiku flies. Reviewing a contract? Opus.

Create a Project for every recurring workflow. This one is quietly enormous. A Project holds your context, your tone, your style guide, your uploads, so you are not re-explaining yourself at the top of every session. Weekly report, support queue, codebase you maintain: those are Projects. Set it up once, benefit forever.

Connect your actual stack through the Connectors tab. This is the moment Claude stops being a chat window and starts being an operator. Through Zapier MCP alone that reaches past 8,000 apps.

Write a CLAUDE.md in your project root. Standing instructions for Claude Code so it never writes a line before it understands your standards. Coding conventions, testing rules, the files it must never touch. Write it once and every future session opens already knowing your rules.

Notice what none of these are. None of them are clever. They are all boring, one-time, and they pay every single day afterward.

Layer three: the runtime you have never opened

This is the layer I was most guilty of ignoring. I knew these existed. I had read about them. I had never actually used them.

Extended thinking, for anything where the first answer is usually wrong. That is the author's rule of thumb and it is a good one. Complex code, multi-step math, strategy calls. It gives Claude room to reason before committing. Quick email? Skip it. Debugging a race condition? Turn it on.

Cowork, for autonomous multi-step work on your own machine. When you need Claude to read, edit, and create files on your computer as part of a bigger job. Think "reorganize these 40 research notes into a structured folder," not "answer my question."

Artifacts, for live interactive output. HTML pages, React components, data visualizations, working tools, generated right in the chat. My favorite of the twelve, because it turns "describe the dashboard" into "here is the dashboard, click it."

Claude Code, for autonomous coding. It reads your codebase, plans the architecture, runs the tests, and fixes the errors in a continuous loop. That last part is the whole pitch. It is a loop, not a one-shot suggestion.

Feed it big things, because the context window can take it. With a 1M-token window, Claude handles entire codebases and giant documents in one pass. The chunking habit is a leftover from 2023. Paste the whole thing.

The best voice models, now fully orchestrated across all channels

ElevenAgents puts full orchestration on the voice models the market builds around. Voice, text chat, transcription, and reasoning in one integrated stack, <400 milliseconds, and human-sounding. Plug in any LLM, integrate tools, A/B test, and deploy across channels. More human-like conversations, lower latency, flat $0.08 per minute.

*Ad

Why everyone chases the smallest layer

Look back at the count. Three prompt tips. Nine setup and runtime tips. And yet almost every piece of AI advice you have ever scrolled past lives in that first bucket.

The reason is dumb and human. Prompts are visible. A prompt fits in a screenshot, gets a hook, gets shared. "I made a CLAUDE.md file" is not a viral post. Nobody screenshots their Connectors tab. So the compounding gains sit quietly in the configuration layer where no content ever points, and we all keep trading tricks for the layer that matters least.

I think there is a bigger version of this too. We collectively decided AI tools are easy, so nobody reads the manual, and then we blame the model when the output is average. Every serious tool has a skill curve. Nobody opens Photoshop expecting to be a designer by lunch. Somehow with AI we expect mastery on day one, get mediocre results, and conclude the thing is overhyped.

The order I would actually run this

Do not try to adopt twelve things at once. You will do zero of them. Effort versus payoff, here is the sequence I would follow:

  1. Today, five minutes. Put role, goal, format, and constraint into your next prompt. You will feel the difference on the first response.

  2. This week, twenty minutes. Build one Project for whatever workflow you repeat most. Upload the style guide, the context, the examples.

  3. This week, ten minutes. Start switching models on purpose. Notice when Haiku was plenty and when you actually needed Opus.

  4. Next week. If you touch code at all, write a CLAUDE.md. A rough one beats none by a mile.

  5. When you are ready. Open Connectors and Cowork. That is where it stops being a chatbot and starts being infrastructure.

The feature I had read about and never opened

The original poster closed with a question I have not been able to shake: what is the one Claude feature you have not tried yet?

For me it was Cowork. I had read about it. I had nodded along in three different conversations about it. I had never once opened it. And I suspect that is most of us. We know the feature exists, we just never cross the gap between knowing and doing, and then we call the gap a talent problem.

Once you see the full list together, the 50 tips stop feeling like tricks and start feeling like features you would have found eventually anyway. Just years later. Because Claude is not a chat box with a model behind it. It is a workspace with a model, a file system, a memory, a tool layer, and an agent runtime.

Pick the one feature you have been nodding along about for months. Open it tonight. Ten minutes, no goal, just poke at it.