Hi friend,
Last issue I asked which gap in your three numbers was widest, traffic to list or list to conversations. That question is still open, so send yours if you have not.
This week I wrote a table that tells a nurse which AI tool is allowed near a patient chart. It took most of a day, and most of that day went to three separate products that share one word. The post is here, with a ten-card summary you can download from inside it.
Somewhere in the middle of it I noticed I was doing the exercise I keep recommending to founders and keep skipping myself. So this issue runs that exercise three times: once for a hospital, once for your build stack, and once for the thing I underrate most, which is a design system.
If you have ever opened four tabs to compare AI tools and closed all four without deciding anything, this one is for you. Hit reply. I read all of them.
🧠 Big Idea
You are not choosing a tool. You are choosing whose rules get enforced, and if you have not written the rule down, the vendor already wrote it for you.
Start with how bad the naming is. Three different things are called Copilot inside a hospital. Microsoft 365 Copilot lives in Word and Teams and Outlook, outside the chart entirely. Dragon Copilot is ambient clinical documentation and writes into the chart, which is what it was built and cleared to do. Copilot Health is a consumer preview with no place in patient care at all. One word, three answers, and a clinician who has only heard the word in a hallway cannot tell which one anybody means.
The scoreboard does not settle it. Epic reports that more than 85% of its customer organizations actively use its AI tools, with clinicians at John Muir Health saving 34 minutes a day on notes and Summit Health taking 42% off prior-authorization time. NHS England is rolling Copilot out to more than 505,000 staff after a trial found 43 minutes a day. Every one of those figures is published by the vendor whose product it flatters, which does not make them false and does make them a sales sheet rather than a decision. The research settles it no better: a 2026 JAMIA study found ambient scribes improved how clinicians engaged with patients and created new reconciliation work at the same time. All of it is true. None of it tells a nurse what to open on Monday morning.
What settles it is one sentence. Does this task touch a patient chart or billing? Yes means Epic AI or Dragon Copilot. No means Microsoft 365 Copilot. That is the entire framework. It fits on a card, it can be applied in about four seconds, and it needs none of the numbers above in order to work.
Now move it onto your own desk. Cursor, Claude Code, Lovable, whatever is sitting in your tabs. You have been reading comparisons hoping one hands you a winner, and none of them can, because the answer is not a property of the tools. Write down what each one is allowed to touch. Production data. The deploy. Your customers' inboxes. The ranking falls out of the table, and it will not match anybody's list.
A design system is the same artifact pointed at pixels. Hand a generative tool your tokens, your component inventory, your spacing scale, and it builds your product. Hand it nothing and it builds a demo you rewrite inside a month. Same tool, same prompt, two outcomes, and the tool was never the variable. The variable was whether the rule sat somewhere the tool could read it.
Key takeaways:
The tool question resolves to a table, not a preference. Write the table once and stop reopening it every time something launches.
A design system is a decision table for pixels. It is why the same tool builds a real product for one person and a throwaway for another.
If you have not written the rule down, the vendor's default is your rule, and you agreed to it by writing nothing.
⚡ Quick Wins
1. Write your one-sentence test before you open another comparison post. One question, two branches, a named tool on each side. Mine is the chart question. Yours might be whether the task touches production, or whether a customer ever sees the output. The test is good when someone who does not work with you can apply it correctly without asking a follow-up question.
2. Write the never list before you write the tool list. Money leaving an account, a message sent to a real person, anything published, anything deleted, anything that grants access to something else. Those five cost you a relationship rather than an afternoon. Every decision underneath the never list gets faster, because the expensive branch is already closed.
3. Put your design system where the tool reloads it, not where your team remembers it. Tokens, component inventory, spacing scale, in the file your tool reads at the start of every session. A design system in someone's head is a preference. A design system in a file the model opens every time is a constraint, and only one of those changes what gets generated.
🛠 Tool of the Week
One this week, and it is the least exciting tool I have recommended here, which is the point.
Evidently
What it is: an open-source Python library for evaluating, testing, and monitoring machine learning and LLM systems. More than a hundred built-in metrics, from data-drift detection to LLM-as-judge scoring. Apache-2.0, genuinely free, and you can self-host the monitoring interface or use their hosted version. github.com/evidentlyai/evidently
How to use it: start with one invariant, not a dashboard. Name the thing that must never be true in your system, write it as a condition, and run it in CI so it fails a build rather than sending an email you will archive. One version note that cost me an hour: on 0.7.x the tests are conditions living inside a Report, not a separate TestSuite class the way older tutorials show it.
Why it matters: on the build in the personal note below, one invariant caught a real gap on the first day, an episode closing without an audited transcript attached. I did not catch that. I had written the rule and then written code that quietly broke it, and the only reason I know is that the rule sat somewhere a machine could check. A decision table you write and never enforce is a diary entry.
🧭 Product Hunt Radar
Three launches from today, read by someone who was on the other side of this page nine days ago and has not fully recovered.
Blume An AI-ready, Markdown-first documentation framework. Docs written so that a model can consume them as readily as a person can.
Why it matters: the design system argument above has a prerequisite nobody says out loud. The rule has to be machine-readable before it constrains anything. A beautiful Figma file the agent cannot open is decoration.
HarnessRouter Community Edition An open-source unified interface for agent harnesses, so one project can address several of them without rewriting the plumbing each time.
Why it matters: a router is what you build once you have accepted that no single harness wins. That is the mature version of the tool question, and it arrives about a year after the argument everyone is currently having.
Vidaya A healthspan score assembled from your wearables, your labs, and your DNA.
Why it matters: the consumer end of the argument at the top of this issue. All of the data reach, none of the chart's constraints, no clinician in the loop. Whether that is useful or reckless depends on rules nobody has written yet.
💻 Prompt for Productivity
Prompt: Help me build a decision table for the AI tools I already pay for. Do not recommend new tools and do not rank anything yet.
Ask me one question at a time until you can fill in all four of these, and do not answer any of them for me:
1. Which tools do I actually use in a normal week, and what does each one touch that is hard to undo?
2. For each one, what is the worst thing it could do if I approved a bad output without reading it carefully?
3. Which of those consequences costs me money, which costs me a day, and which costs me a relationship?
4. What would have to be true for me to let one of these run without me watching it?
Then give me a table, one row per tool, three columns: what it may touch, what it must never touch, and what has to be true before it runs unattended. Add a final row called Never, listing the actions no tool on the list may take under any circumstance. Where two tools overlap, tell me which one wins and the one sentence that decides it.
WHEN TO USE IT: the next time you catch yourself comparison-shopping AI tools instead of shipping, which for me is roughly every second Tuesday.
WHAT IT DOES: it converts a preference argument into an artifact. Most tool debates never resolve because the two people arguing are optimizing for different consequences and neither has said which one. Naming the blast radius first makes the ranking fall out on its own.
TIP: write the Never row first and out loud. If you cannot say what a tool is never allowed to do, you have not decided anything, you have picked a favorite.
🌎 In the News
Cognition is reportedly in talks to raise at a $40 billion valuation my POV: reported, not announced, so hold it loosely. Three months ago Cognition raised $1 billion at $26 billion on a $492 million run rate, and the number now being discussed sits on a run rate of about $1 billion. The line that stayed with me came from the CEO: enterprises grew their usage of Devin by 50% month over month, for six months running. The money is going to the agent that lives inside the repository, not the one that answers questions about it. The repository is the constraint, and the constraint is the moat.
LinkedIn published how its feed actually ranks now my POV: this went up in March and I only read it properly this week. They replaced a stack of separate retrieval systems with one LLM-based retrieval layer, then rank with a model that reads more than a thousand of your past interactions as a sequence rather than as separate events, weighted against profile signals like industry, skills, and geography. The funny part is that the platform published its decision table on its own engineering blog, and the guides most of us read instead confidently describe parameter counts and reach penalties that appear nowhere in it. The rule was written down. We went and asked somebody else.
Hackensack Meridian is the first health system certified for responsible AI use my POV: the Joint Commission program launched June 1 and assesses six domains: governance, privacy, risk and bias, monitoring, transparency, and staff training. Read the scope carefully, because it is the most important sentence in the story. It certifies the organization's process, not the AI products moving through it. That is a table about tables. Worth having, and not the same thing as knowing whether the tool works.
🔧🎓 Workshops, Events & Happenings
Two of these are mine, and one of them has a question attached that I would genuinely like answered.
Claude Loves Lovable. Five-part class, starting this fall, with dates landing in September. Five live 90-minute working sessions, one a week, in the order a real product comes together: brand, then pitch, then MVP, then website, then code. Every session pairs Claude for the thinking and the copy with Lovable for the building, because the handoff between the two is the part nobody teaches and everyone gets wrong. You build alongside the demo rather than watching one, and you leave with a working Claude skill you keep. No coding experience needed. Reply and I will make sure you hear the dates first.
The clinician course catalog is live. Two of the six speak to everything above. Chart Smarter is AI for clinical teams, CE-eligible pending accreditation, covering the zero-PHI rule, safe prompt structure, spotting hallucinations and drift, and where human vigilance is not optional. Guardrails is AI governance and standards for teams, including healthcare AI procurement and vendor evaluation, which is this issue's table at department scale. Both are built as something a department brings in, so here is my question: would you want Chart Smarter as an open cohort you could join on your own, without waiting for your employer? Reply yes or no and I will build it or shelve it accordingly.
AI Builders Hackathon. August 21 to September 15, 2026. Online. Twenty-five days, solo or team, and the eligibility page explicitly names startup founders and entrepreneurs rather than only students, which is rarer than it should be. No prior hackathon experience required. The prize pool is still to be announced, so enter for the deadline rather than the money. Details here
RevenueCat Shipaton 2026. August 1 to September 30, 2026. Online. More than $685,000 in cash prizes with a $100,000 grand prize. The constraint is the interesting part: the app has to be released for the first time inside the window, and it has to charge money through the RevenueCat SDK. A two-month hackathon about shipping something that actually takes payment is a better forcing function than anything most of us apply voluntarily. Six weeks left. Enter here
✍️ Personal Note
The other thing I built this month is an outbound follow-up engine for crisis call centers, working the 24 to 72 hour post-crisis queue that guidance requires and most centers ration by staffing. It is not finished, and no live call has ever been placed.
The design decision arrived early and shaped everything after it. The model interprets, a deterministic layer decides, and a human owns anything that cannot be taken back. The language model is only ever allowed to fill in a closed schema, and every decision after that is a plain function of what is in it. When anything is unclear the default is not to retry, it is to hand the whole thing to a person. A handoff is not complete when the system decides it went fine. It is complete when a named human accepts it.
Here is the developer tip, and it is the one I would give anyone building agents this month. Refusal has to be a different result type from failure. If your agent saying "I will not do this" looks identical to "this errored," your retry logic will quietly undo your safety rule at three in the morning and nobody will notice for a week. Refused is not failed. Give it its own type.
📬 Closing CTA
If you are staring at a tool decision and cannot tell whether it is a real fork or just noise, that is the conversation I like having most. Book a Vibe Check. There is nothing to buy on it, and I will tell you plainly if I am not the right person.
💌 One Ask
What is the one thing your tools are never allowed to touch? Reply with just that, one line. I am collecting them, and I suspect the list is going to be shorter and stranger than anyone expects. 🌿