Volume 40: You Do Not Need to Code, You Need to Decide What to Build

I spent one afternoon building a working AI assistant for my website, and I never wrote a line of the code. I expected the technical difficulty to be the holdup. It almost never was. What slowed me down was a choice I had not yet made, sitting on my side of the screen.

🧭 Founder's Corner: Why the barrier to building with AI was never technical skill, and the part of the work that stays yours no matter how capable the tools become.

🧠 AI Education: How an AI assistant actually builds context over time, the difference between the memory you already have for free and the features you pay for, and the setting most people never check.

✅ 10-Minute Win: Set up a workspace that remembers your project so you stop re-explaining the same background at the start of every session.

Let's dive in.

1) How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery

Summary: Immunologist Derya Unutmaz used GPT‑5 Pro to revisit a three-year-old puzzle centered on a special type of immune cell that helps the human body fight cancer and other illnesses. When asked to simulate one of his unpublished experiments, GPT-5 Pro correctly predicted the boost in CD8+ cells' ability to kill lymphoma cells, results the model could not have gleaned from the internet.

Why it matters: Frontier AI is moving from "helps you write" to "helps you reason about biology you have spent a career studying." For anyone in clinical research, drug discovery, or specialty pharma, this is the new baseline. The labs that pair domain experts with reasoning models on legacy datasets are going to compress years of analysis into weeks.

2) Trump admin allows Anthropic to release Mythos AI model to some companies, government agencies

Summary: The US government on Friday granted Anthropic permission to release its Mythos 5 model to a group of roughly 100 companies and federal agencies, two weeks after disabling it under an export control directive. The letter, addressed to Anthropic co-founder Tom Brown, did not grant Anthropic approval to restore access to Fable 5.

Why it matters: This is the first time the federal government has designed an allow-list for who can run a specific commercial AI model. For any organization buying or building on frontier models, the message is that access is now a policy decision, not just a procurement decision. Your AI roadmap needs a Plan B for the day your preferred model is restricted.

3) Rhode Island passes ambient AI scribe opt-out law

Summary: Rhode Island Governor Dan McKee signed a 12-piece healthcare legislation package, which includes safety guidelines for AI chatbots and the use of AI in mental healthcare, requiring AI scribe disclosure to patients and giving them the right to opt out. Chatbot operators face up to $15,000 per day in fines for failing to route users expressing self-harm to crisis services, and unlicensed AI companions marketed for mental health or emotional support are now prohibited.

Why it matters: Ambient AI scribes have rolled out faster than the rules around them. Rhode Island just made disclosure and opt-out a default, and other states are reading the same script. If your health system or vendor is deploying ambient scribes, patient consent workflows need to be in place now, not when your state catches up.

4) NVIDIA powers over 400 of the world's 500 fastest supercomputers

Summary: According to the latest TOP500 rankings released at the ISC High Performance conference in Hamburg, NVIDIA technologies now power more than 400 of the world's 500 fastest supercomputers, with nearly nine of every 10 systems new to the ranking built on NVIDIA technologies. NVIDIA systems across the TOP500 now deliver more than 2x the AI training and nearly 3x the AI inference throughput of every other platform combined.

Why it matters: When one vendor sits underneath 81% of the world's heaviest compute, the strategic question shifts from "which model do we use" to "who controls the floor those models run on." For procurement, governance, and resilience planning, the practical move is to know which workloads in your stack ultimately depend on a single supplier, and what your contingency looks like if that changes.

5) AI Is Pushing Healthcare Toward a Breaking Point, ZS Research Shows

Summary: Based on responses from more than 10,000 healthcare consumers and providers across the United States, Germany and China, the ZS Impact Institute 2026 Future of Health Report found approximately 90% of people who use AI and digital tools for health information trust it nearly as much as their doctor, and 42% of US consumers research symptoms online before deciding whether to see a doctor.

Why it matters: Patients are no longer arriving at the visit cold. They are arriving with an AI-generated working hypothesis, a question list, and a trust level rivaling their doctor's. For health systems, the front door to care is shifting from the call center to the AI assistant. The redesign question is whether your intake and education systems are ready to meet patients where they already are.

An assistant named Neura now lives on my website. Open her from any page and ask where to start with AI agents, or which piece explained retrieval. She answers in a few plain sentences and points you to the right posts. She has read everything I have published, all one hundred and fifteen of them, and she keeps the whole archive in mind when she replies.

She came together over about three hours of focused work across a single day, with the World Cup and the US Open on in the background. I am not a developer, and I have never claimed to be. Not one line of the code is mine. I directed Claude Code from my terminal, made every real decision myself, and watched a retrieval-powered product come to life in production on my own site.

My reasons for this build went past content for the newsletter. I wanted a living tool inside my own site, not a static archive. I wanted to build it myself, and to apply knowledge I had been developing over more than eighteen months. Retrieval techniques and vector databases have been recurring topics in Neural Gains Weekly, yet I had never wired them together with my own hands. Explaining a concept and building a system around it are two different skills, and I wanted to close the gap. Neura was the chance to do exactly that, with my own archive as the proving ground. That is the part that pulled me in.

Every Stall Was a Decision I Had Not Made Yet

Before any building, Claude Code got one instruction from me, about how we would work together. I told it plainly, "I need to move slow as this world is newer to me." As new ideas entered the stack, a vector store here, an embedding step there, Claude Code paused to explain each one in plain language before we built on it, teaching me the concepts as we went along. Unfamiliar terms became clear within minutes, and the aha moments came faster than they would have alone.

Directing the build looked less like coding and more like a run of decisions. That shift reaches well beyond me. As of January 2026, roughly nine in ten professional developers reported regularly using at least one AI tool at work, in a JetBrains survey of more than ten thousand of them. Even the people who can write code by hand now lean on AI as a matter of course. The same help has simply reached someone who could not write a line of it. I described what I wanted in plain words, and Claude Code proposed a plan. The first real decision was how much to trust it. Rather than let it build and ship on its own, I connected it to my GitHub repository and had it open a pull request for every change, a proposed set of edits waiting for me to review before anything went live. That was the governance I built in on purpose. Claude Code wrote the code and opened the request. My part was to confirm the details I had asked for were actually there, not to hunt through the code for bugs, and to approve it only once they were.

The bigger decisions came before any code was written, and there was no engineer down the hall to settle them. Some I could not have answered without assistance. When it came time to pick where the whole thing would run, I did not know enough to choose between Cloudflare and Google Cloud on my own, so I asked. Claude Code laid out the tradeoffs, one vendor with everything I needed in a single free tier against a more powerful platform built for enterprise scale I did not have. The facts made the call obvious, and the call was still mine to make. The smaller, cheaper model was smart enough for the job, so the expensive option came off the table. Full retrieval beat a quick shortcut, so the archive could grow without boxing her in later. Claude Code wrote the code behind every one of those decisions, but the deciding was the real work, and it was mine.

While building the vector database in Cloudflare, I loaded all one hundred and fifteen posts so Neura could search them, and the terminal work began. This was a new rhythm of commands, so I ran them one at a time and watched what came back. On one of them, nothing happened. Nothing happened, just the command sitting there waiting on me. The cause turned out to be a single missing quotation mark. I worked out why, fixed it, and the posts flowed in. I checked my footing as I went, asking Claude "Did the terminal work?" and "We good?" until the output told me we were. Each command that landed made the next one feel less foreign.

The deployment had noticeable bugs that were only apparent after Neura went live on the site. The chat button hid behind another button where no visitor would find it. Dark text rendered on a dark background, making the header unreadable. And the early answers ran long, stuffed with citation numbers that pointed nowhere. I caught each one the way a reader would have, live on the page, and fixed them in minutes.

A year ago, stalls like these would have overwhelmed me. This time they did not. I stayed calm, worked each one methodically, and knew the next step to take, even without being able to read a line of the code. That steadiness was the surprise. I kept expecting the technical difficulty to be the holdup. It almost never was. Each stall traced back to a decision about what I wanted that I had not yet made. The moment I got clear on the goal, the work moved again quickly. The thing slowing me down had been sitting on my side of the screen the whole time. By the end, I honestly felt comfortable. That comfort was earned, the sum of every rep that came before, going back to my first AI coding project. What I did not expect was how much the job itself had changed underneath me.

A Senior Developer Who Cannot Read Your Mind

In an earlier issue, I treated an AI coding tool like a senior engineer who already knew everything, and I broke my own homepage in the process. The lesson was simple and humbling. I could not hand over the responsibility and step away. What I had then was a junior developer, capable but green, and nothing it produced could go unchecked.

Building Neura, that same kind of tool no longer behaved like a junior. Claude Code, now running on Opus 4.8, reasons through tradeoffs, raises problems before they surface, and explains its choices as it makes them. It feels like a senior dev. The constant supervision the homepage demanded was mostly gone.

It would be easy to read that as the human mattering less in the build. That is not how it played out. As the tool took on more, the weight of the work shifted onto me. A senior developer who cannot read your mind is powerful and a little dangerous at the same time. Hand a vague request to a capable engineer and you get exactly what you asked for, done well and done fast, even when what you asked for was wrong. The sharper the tool gets at execution, the more your half of the job narrows to deciding precisely what to execute. Speed only makes a bad decision arrive sooner.

None of this belongs to one project or one tool. The more capable the machine becomes, the more the work depends on a clear head and a well-chosen problem, which is to say it depends on you. That is the shift worth carrying into whatever you decide to build next.

The Part No Tool Will Ever Do For You

So what is left for you, the capable professional who has been watching all of this from a distance? The most important part, and it is not close. You decide which problem is worth solving and which tradeoffs you will accept. You define what good looks like, and you choose when something is finished. Taste and judgment are yours to supply. They always have been.

The growing ease of these tools does not worry me, because the skill that carried me through this build was a different kind entirely, learning to translate between the problem in my head and the language a machine needs to act on it. That translation holds up far beyond a terminal, in any room where someone has to turn a fuzzy goal into a clear instruction.

So here is the one move worth making this week. Take a single problem that nags at you, the kind of small, recurring friction you have learned to tolerate, and describe it in plain English, the way you would explain it to a colleague. Then open Claude Code, or another tool like it, and start. There is no product to ship here, only your first real rep at directing a machine to build something, which is quickly becoming one of the most useful things a person can know how to do.

Open Neura if you want to see where a few spare hours can land, though the tool itself is beside the point. What matters is this: the starting line is not a coding course, and it never was. You do not need to learn to code. You need to decide what to build, and then decide to begin. That decision was always the hard part. It was just hiding behind the code.

What Is Actually Going On Here

A professional opens Claude on a Monday, asks for help with a board memo, and Claude already seems to know the house style, the audience, and the fact that the quarterly numbers are due Thursday. Nobody pasted that context in. The conversation where it came up happened two weeks ago. To the user it feels like the tool simply remembers them, the way a colleague would.

What is actually happening is more deliberate, and more controllable, than that impression suggests. Claude is not silently watching everything. This continuity only appears when a particular setting is doing its job, and understanding the mechanism is the difference between a user who fights the tool and one who works with it. Two distinct systems sit behind that Monday-morning moment, they do different jobs, and they are not handed out on the same terms.

Two Systems That Get Blurred Into One Word

"Memory" gets used as a catch-all for two features that are actually separate.

The first is memory itself. When it is active, Claude reads back over a user's past conversations, writes a running summary of the durable facts, the kind of work they do, how they like to communicate, what projects are in flight, and carries that summary into every new chat. The user does not re-explain themselves each morning. That summary refreshes on a daily cycle rather than the instant a chat ends.

The second is to search across past chats. This is not the summary. It is the ability to reach into one specific old conversation and pull an exact detail back out on demand. A user asks what they decided about the vendor contract in March, and Claude goes and finds that conversation rather than leaning on the general summary.

Here is the part that surprises people, and it is the most useful thing to take from this section: these two are gated differently. Memory, the running summary, is now rolling out across every plan, including the free one. Search across past chats is a paid feature. So a free user already has a colleague who remembers the gist of who they are. The paid plan adds the colleague who can also go pull the exact thing you said six weeks ago. Most people never realize these are separate capabilities, which means free users often do not know they already have the more powerful of the two.

The Setting Worth Knowing About

Memory lives in settings, under Capabilities, and it is something a user can see and control rather than a black box running in the background. That control is the whole point, and it is where the people who get real value diverge from the people who do not.

The users who benefit most are the ones who know the feature is there and treat it as something they manage. They check what Claude has captured about them, correct anything off, and tell Claude directly what to keep. The users who get less are not doing anything wrong. They simply never look at the setting, so they never shape what it holds or confirm it is working the way they want. Same product, two very different experiences, and the gap is entirely about whether someone knows the control exists.

Where Projects Come In

Projects are the second half of the picture, and they solve a problem the running summary does not.

A project is a dedicated workspace inside Claude. A user creates one, gives it a set of standing instructions, and uploads the documents that matter for that body of work. Every conversation started inside that project automatically inherits those instructions and has those documents on hand. A board-reporting project holds the board's format and the prior decks. A client project holds that client's history. None of it gets re-pasted at the start of each chat.

Underneath this is a mechanism that has come up before in this newsletter. A single conversation can only hold so much in view at once, the context-window idea from Vol 8. On a paid plan, when a project's uploaded material grows past that limit, the project switches to retrieval, the RAG pattern from Vol 19 through 22, fetching the relevant passages at the moment of the response instead of trying to hold every document in view. That is what lets a project draw on far more reference material than any single chat could.

The detail that catches people off guard is that each project keeps its own separate memory. What Claude learns inside the board-reporting project stays there. It does not leak into the client work, and it does not fold into the general day-to-day chats outside any project. For a professional running several distinct streams of work, that separation is the point. The contexts stay clean.

What Persists and What Does Not

Honest boundaries matter as much as capabilities, so here they are plainly.

The running memory summary covers conversations outside of projects. Project conversations are deliberately walled off from it, and each project remembers only itself. Incognito chats, the temporary mode marked by a ghost icon, remember nothing at all, and they are available on every plan for exactly the moments a user does not want a conversation folded into the record.

One more boundary worth naming: the summary is a synthesis, not a transcript. It captures the durable shape of who a user is and what they are working on, not every sentence they ever typed. Memory holds the standing gist. Reaching back for an exact quote from a specific past chat is the paid search feature doing a different job.

What This Means for How You Work

A few concrete moves follow from understanding the mechanism.

First, find the memory setting and check what it has captured. The summary is viewable and editable under Capabilities. A user can read exactly what Claude believes about them, fix anything wrong, and state directly what to remember. The people who get the most from the tool treat that summary as something they curate, not something that merely happens to them.

Second, use projects to separate streams of work that should not bleed together. One project per distinct body of work, each with its own instructions and its own walled-off memory. That structure is what keeps a busy professional's contexts from contaminating each other.

Third, reach for incognito when a conversation should leave no trace. A sensitive draft, a one-off question, anything that should not shape the running summary. It is the clean-slate option, and it costs nothing on any plan.

How This Connects

This is the second part of the Claude Deep Dive. Vol 39 oriented around first contact and the free plan's genuine core. This volume went under the hood on the features that let Claude accumulate context: memory that summarizes who you are and rides along into every new chat, search that retrieves what you specifically said, and projects that wall off each stream of work with its own instructions and its own memory. The retrieval machinery inside projects is the same RAG pattern from Vol 19 through 22, and the reason it is needed traces back to the context-window limit from Vol 8. Next week, Vol 41 turns to Artifacts and long-form work, where Claude does its most distinctive professional lifting on documents, reports, and presentations. The thread running through this series is what it actually feels like to work inside the tool, not what the marketing says it does.

Part 2 of 4 in the Claude Deep Dive series.

Stop Re-Explaining Your Project Every Time

You are three weeks into something real: a launch, a research effort, a planning sprint. The work lives across dozens of AI chats. Every time you open a fresh one, you paste the same background again: the goal, the constraints, the documents that matter. You spend the first five minutes of each session rebuilding context the AI forgot when you closed the last tab. Across a multi-week project, that is hours lost re-explaining yourself.

A project workspace fixes this. In Claude, the feature is called Projects. You configure the context once, custom instructions and reference files, and every chat inside that Project starts knowing your work.

Two notes before we start. First, this is a paid feature, not a free one, and that is the point of today: knowing when a paid AI feature earns its cost is a fluency skill of its own. Second, it pairs with this week's AI Education section, Projects and Memory, which explains how Claude builds context over time. That section is the theory. This is the practice.

Why this matters: a chat session means starting over; a workspace means picking up where you left off. Once you feel that difference, you stop treating AI as one-off conversations and start treating it as a place your work lives.

The Workflow

1. Measure the re-explaining tax (2 minutes). Open your usual AI tool and run this on your current project:

Copy/Paste Prompt: "I am working on a multi-week project. List everything you would need to know about it to give me genuinely useful help, assuming you know nothing. Be specific."

Look at that list. That is the context you retype, in some form, every single session. Estimate how many minutes per week you spend rebuilding it.

2. Run the cost rubric (2 minutes). Claude Projects requires Claude Pro, which is $17 a month billed annually or $20 month to month. Ask the one question that decides any tool purchase: does it earn its cost? If you lose 30 minutes a week re-explaining context, the feature buys that time back for under a dollar a day. If you run one project a month, maybe not yet. Decide honestly.

3. Create your Project and write its instructions (3 minutes). If you upgrade or already have Pro, click "Projects," create one, and name it for your actual project. In the custom instructions box, paste a version of this:

Copy/Paste Prompt: "This project is about [PROJECT GOAL]. Key constraints: [CONSTRAINTS]. Stakeholders: [WHO]. When I ask for help, assume this context. Ask me what is unclear rather than guessing."

4. Add your reference files (2 minutes). Upload the two or three documents the project keeps coming back to: the brief, the plan, the key data. Every chat in the Project can now see them without re-pasting.

5. Validate the value (1 minute). Start a fresh chat inside the Project and ask a real question with no background. If it answers with your project's context already loaded, the workspace is working, and you just felt what your money bought.

The Payoff

If you stayed free, you walk away with a clear-eyed decision and the rubric to make it. If you upgraded, you have a live workspace that ends the re-explaining tax for this project. Either way, you can now see the line between a chat and a workspace, which is the real fluency gain.

The AI Concept You Just Used

Persistent context. A normal chat forgets everything when it closes. A workspace holds your instructions and files so every conversation starts informed. It is the same idea that powers custom assistants and enterprise AI: configure context once, reuse it everywhere.

Transparency & Notes

• Claude Projects is a Claude Pro feature, not free. Pricing was verified on Anthropic's official pricing page, but prices change, so confirm before you subscribe.
• Not ready to pay? Gemini Gems offers a free take on persistent context, custom instructions plus reference files, though it organizes around a reusable assistant rather than a full project history.
• Do not upload PHI, confidential metrics, or NDA-protected material to any project workspace unless your plan and your employer's policy allow it.
• For very large document sets, the workspace pulls the most relevant parts per question rather than reading everything at once.