The Claude Features Almost Nobody Turns On, Part 1: Memory, Real Builds, Better Thinking, and Honest Pushback
Part 1 of a practical guide to getting far more out of Claude: how to give it memory, make it build working things instead of describing them, make it reason on the decisions that matter, and turn it from a flatterer into something that tells you the truth.
I want to start with the moment most people quietly give up on Claude. They open it, type a question the way they would type it into a search bar, get a competent paragraph back, and conclude it is a nicer Google. Useful, a little overhyped, nothing that changes their week.
Then they watch someone hand the same model a half-formed goal and get back a finished research brief, a working prototype, and a blunt list of why their plan will fail. Same model. Same subscription. The difference is not a secret prompt. The difference is that the second person stopped asking Claude questions and started configuring it.
“A search box forgets you the moment you close the tab. A configured assistant remembers who you are, builds real things, reasons before it speaks, and is willing to tell you that you are wrong.
”
That sentence is also the structure of this piece. This is Part 1 of two, and instead of marching through a feature menu, I have organized it around the four things you actually want Claude to do inside a conversation: know you, build for you, think properly, and disagree with you. Part 2 leaves the chat window for the power tools that act on your computer and run while you sleep. Let us take the four in the order I would set them up.
First, make it know you before you type a word
Every other technique in this series gets better the moment Claude stops starting from zero, so this is where I always begin.
A Project is a persistent workspace. You drop in your reference documents and write standing instructions once, and Claude carries all of it across every chat inside that project. The reference material matters, but the instructions field is where the real leverage hides, and it is the field almost everyone leaves blank. Tell it who your audience is, what you never want explained to you, and the exact phrasings that make you wince.
Project knowledge
- 📄 audience-profile.md
- 📄 past-issues.md
- 📄 brand-voice.md
Custom instructions
”My readers are technical. Never explain what an LLM is. Find the surprising angle. No filler, no corporate language. Short paragraphs, prose not bullets.”
A starting point worth stealing and editing
You are my content research partner. I publish for a technical audience that already knows the fundamentals, so never explain the basics.
When I share a topic or a link, do three things:
- Name the three most counterintuitive angles in it.
- Connect it to something recent I probably missed.
- Show me how to frame it as a story, not a summary.
Style: direct, specific, no filler. Short paragraphs. No bullet points unless I ask. Never open with “Great question”. No em dashes.
A Project remembers your context inside one workspace. Memory, where your plan offers it, does the same thing across all of your chats: switch it on, tell it once who you are and how you like to be helped, and a brand-new conversation already knows your role and your current focus. Projects are the right tool when you want a walled-off context for one effort. Memory is the right tool when you want Claude to simply know you everywhere. Most serious users end up running both.
Then, make it disagree with you
Here is the counterintuitive one, and the one I get the most mileage from. By default Claude is agreeable. It builds on your idea, finds the upside, and reassures you. That feels nice and is frequently worthless, because the thing you actually need before a big decision is someone who will tell you it is a bad idea.
You fix that by assigning a role. Claude commits to a stance when you give it one, and it changes what it questions and what it refuses to let slide. These are the five I keep on hand, in the order I reach for them.
The devil's advocate (the one I use most)
When you have already made up your mind is precisely when this earns its keep. Ask for the full case against your decision: not polite reservations, but the three most realistic ways it fails and the assumptions you are quietly betting on. Five minutes here has talked me out of more bad calls than any amount of optimistic planning ever talked me into good ones.
The brutally honest mentor
Slightly broader than a single decision. Tell it to be an operator who has built things, failed at some, and watched people make the same confident mistakes. Its job is not to encourage you, it is to find the blind spot you are avoiding because you want the plan to work, and to end with the one thing you should think hardest about before moving.
The conversation you are dreading
Most people rehearse what they will say and never rehearse what they will hear back. Have Claude play the other person, in character, answering the way they actually would and pushing on your weak points. Step out between exchanges for a quick note on what landed. By the third run the real conversation has lost most of its teeth.
The questioning therapist
For the decision you keep circling rather than resolving. Instead of handing you advice, this role asks questions, listens for distortions like catastrophizing or all-or-nothing thinking, and names them so you reach your own conclusion. It is the opposite of using Claude as a validation machine, which is what most people do without noticing.
The specific trainer
A reminder that roles are not only for thinking. Hand it your real constraints, age, goal, equipment, available days, any injuries, and ask for a structured program with the reasoning explained, then report back weekly so it adapts. The point is not fitness advice, it is that a role plus your real numbers beats any generic template, in any domain.
The mechanism under all five is the same: stop letting Claude be a neutral assistant and assign it the stance that is useful right now. Here is the prompt I run before committing to anything that matters.
Run it before you commit, never after
I have decided to do the following, and I want the strongest possible case against it before I commit.
The decision: [exactly what you plan to do] My reasoning: [why you think it is right] Already considered: [objections you have handled]
Do not balance this with positives. I already believe in it.
- Attack the assumptions I am making that could be wrong.
- Describe the three most realistic ways this fails.
- Tell me what I am underestimating, and what I would have to believe for this to be a clear mistake.
Be ruthless. If it is a mistake, I need to know now.
Make it build, not describe
This is the shortest section and the biggest mindset shift. Most people think Claude can only return text, so they ask it how to build something and then go build it themselves. Artifacts are working things rendered right beside the chat: a calculator, a dashboard with live charts, an interactive diagram, a small web app you can actually click. Inline SVG, charts, and Mermaid diagrams all work, and it is on the free plan.
”Build me a habit tracker. Five habits, daily check-offs, a 7-day streak that resets if I miss a day. Dark, minimal, satisfying to click.”
Live artifact
The shift is one word. Stop asking “how would I build” and start saying “build me.” You get a working version you refine by talking to it, which is a very different relationship to the work than copying a code sample into an editor.
Make it think slowly, but only when it counts
The last in-chat lever is Extended Thinking, which tells Claude to reason step by step before answering, with the working shown. Almost nobody turns it on. The trick is knowing when: it is wasted on a quick fact and genuinely different on a hard call, a strategy question, or a tangled tradeoff. Switch it on, ask something you have asked before, and read the two answers next to each other. The careful one is usually a different quality of thought, not just a longer version of the fast one.
Now stack them
Here is the part the feature lists never show you, and where the real compounding happens. These are not four separate tricks, they are one workflow. The next time you face a genuine decision, run it as a chain inside a Project that already knows your context: turn on Extended Thinking and have it reason through the options, then in the same thread switch it into the devil’s advocate and make it attack its own recommendation, then ask for a small Artifact that models the numbers so you are deciding on something you can poke rather than a vibe.
“Any one of these is a feature. Chained together, with a Claude that already knows you, they are a thinking partner that most people do not realize they are allowed to have.
”
That is the in-chat half: a Claude that knows you, builds for you, reasons on the calls that matter, and is willing to disagree. In Part 2 we walk out of the chat window entirely, to the tools that let it read your screen, edit your files, write and fix your code, and run on a schedule while you are asleep.
