The Edge
Issue #005 · May 13, 2026
◆ This week
01 Working Theory: companies are moving faster than anyone predicted. The race is real. Nobody knows what they're running toward. 02 Harvey raised $200M on top of Anthropic's Claude. Last week, Anthropic launched Claude for Legal. The company that powers Harvey just became their competitor. 03 Isomorphic Labs just raised $2.1 billion to design drugs with AI. They're based in London, not Silicon Valley. That's worth noting. 04 This week's Workshop: don't figure out what to ask. Tell Claude what you're trying to do and ask it to build the prompt. Then refine it. Then a clean new chat.

Working Theory

A recurring section where I share what I'm currently seeing in the AI landscape and what I think it means. These aren't predictions or hot takes. They're working theories, meaning they're my best read of the moment, written with the awareness that the moment keeps moving. I'll update them when the evidence changes.

A Good Decision From a Fast One

Over the course of my career, I've watched a lot of technology land in large organizations. The pattern is usually the same. The technology arrives, consultants explain it, leadership schedules a strategy session, a working group forms, a pilot gets approved, 18+ months later we are live in production.

Today feels different. What I'm watching right now is more about the race. Oddly, nobody can quite name what they're racing toward and nobody wants to find out what happens if they finish last. You can sense the urgency. Not in press releases but in how decisions are getting made.

A year ago, I'd have said the kind of adoption I'm seeing today was still three years away. The technology continues to move fast. That part I expected. What I didn't expect was that change management at scale would speed up to match it. Large organizations are being tested to absorb change at this pace. Keeping up with what these tools can actually do, week to week, is genuinely hard. The decisions getting made aren't necessarily wrong. They just feel like decisions made to avoid being last more than decisions made to do this the right way, because there is certainly real value here.

Companies appear to be responding in two ways. Some are building internally. That tends to be the slower path. It requires hiring, training, governance, infrastructure, and a tolerance for moving at the pace your own organization can actually move. Some don't have all of that, so they turn to third parties to move faster. Harvey for legal. Glean for enterprise search. Hebbia for research workflows. There are hundreds of these now and the good ones are genuinely good. They're specialized in ways the general-purpose tools aren't yet. They've spent years building workflows, integrations, and guardrails for a specific job. For a lot of teams, that depth is the reason to buy.

It's also worth knowing what you're actually paying for. Most of these products are wrappers. Built on top of an AI provider (OpenAI, Anthropic, Google), packaging the underlying capability into a workflow and reselling access as a service. You pay the vendor. The vendor pays for Tokens. The buyer rarely touches the technology directly, which means the buyer also doesn't develop much understanding of what's actually doing the work.

That tradeoff used to be straightforward. You gave up understanding in exchange for a more polished product. The AI providers are now starting to change the second half of that equation. Claude tools and add-ons are showing up directly in design work, legal work, code, and desktop task automation. These companies that used to sit one layer back are now building the products on top too. And living inside one provider's ecosystem is starting to look meaningfully different from living across multiple wrappers. This is also part of what makes building internally worth the patience it requires. A team that builds directly on a provider's platform inherits that same interconnection, plus the understanding that comes from doing the work themselves.

None of this means the specialized tools are going away. Harvey isn't getting displaced by Claude in legal work next quarter. But the calculus is shifting, which could be problematic if your company is already halfway through the race.

I don't think every decision made while running this race was a bad one, but I do think they were familiar. Some companies know how to buy software and benefit from working with vendors. The difficulty is understanding a good decision from a fast one.

AI News

The Company That Powered Harvey Just Became Its Competitor

Dark moody courtroom interior

Harvey raised $200M last year. They built their legal AI product on top of Claude, Anthropic's model. A few weeks ago, Anthropic launched Claude for Legal, their own product aimed directly at the same market.

Anthropic has begun developing tools that directly compete with companies building on top of their models. Harvey's bet was that their product layer would be defensible. The workflows, the trust, the enterprise contracts. That bet just got harder.

Read full story →

A $2.1 Billion Bet That AI Can Design Better Drugs Than We Can

Dark moody molecular science laboratory

Isomorphic Labs just closed a $2.1 billion Series B. They're using AI to design drug molecules from scratch, not to accelerate existing research methods but to generate candidates that human scientists wouldn't have known to look for.

They're based in London. The AlphaFold research that underpins all of this came out of DeepMind, also London. This isn't just a funding story. The talent doing the most serious AI research in biomedicine isn't concentrated in one zip code in California. That window is open in a lot of places right now.

Read full story →

The Workshop

Give Me a Prompt
A short, practical experiment. Try it now.

Most people sit down with a task and try to figure out what to ask AI. That's the hard part. Skip it.

Tell Claude what you're trying to do and ask it to build the prompt for you.

You:

I need to write a performance review for someone who did solid work this year but missed a few deadlines. Give me a prompt I can use to write this well.

Claude:

Here's one:

"Write a performance review for an employee who delivered strong work this year and demonstrated [specific strengths]. They missed [X] deadlines related to [context]. Acknowledge the good work specifically, address the deadline pattern without over-penalizing it, and frame the path forward as motivating rather than punitive. Tone: direct and fair. Length: 300-400 words."

Now refine it. You don't have to take the first version.
You:

Make it more specific to someone in a client-facing role.

Claude:

"Write a performance review for an employee in a client-facing role who delivered strong work this year and demonstrated [specific strengths, e.g., client relationship management, responsiveness under pressure]. They missed [X] deadlines on [types of deliverables, e.g., client reports, proposals]. Acknowledge the good work with specific examples where possible, address the deadline pattern directly without framing it as a character issue, and close with a concrete path forward in a tone that is motivating and specific to their growth in client work. Tone: direct and fair. Length: 300-400 words."

When it feels right, stop. Open a new chat. Paste the prompt. Get your output. The reason a new chat matters: the back-and-forth you used to build the prompt is now clutter. A clean context gets you a cleaner result.
Try it this week on something that's been sitting on your list. DM me on LinkedIn with what you built it for.

Resources

Claude for Legal
claude.ai →
Anthropic's direct legal AI offering. If your firm or business touches legal work, worth knowing it exists.
Isomorphic Labs
isomorphiclabs.com →
The DeepMind spinout using AI to design drugs. Worth following if you want to track where AI is being aimed at hardest problems.
Harvey
harvey.ai →
The legal AI company built on Claude. Still one of the stronger vertical AI tools. Useful context for understanding the ecosystem shift.