LLM OS.

Increasingly, AI platforms like Claude and ChatGPT are the place teams do real work from — not a side panel they tab to occasionally, but the actual surface they spend their day in. LLMs have firmly moved into a phase where you can accomplish meaningful work in your systems directly from the chat if you set it up the right way. But that's a big if!

Context, context, context. LLMs without context are useless. LLMs with a little context are okay. LLMs with as much of your firm's real context as you can possibly give them are invaluable. Everything else in this module (provider choice, plan, skills, instructions, connectors, memory) is in service of that one principle: get more of your firm's thinking, data, and judgment in front of the model.

None of what follows works on individual logins. You need every person on a shared team or enterprise plan from the same provider — so skills, connectors, instructions, and shared workspaces actually propagate across the firm rather than living in five disconnected accounts.

A team plan is usually enough. It doesn't preclude you from any of the major features. Enterprise can be a touch better for firms with strict privacy or security postures (SSO, audit logs, stronger data handling guarantees) but the step that actually matters is going from individual to team.

We recommend Claude. Anthropic is, right now, the most innovative company in the world moving at an almost unimaginable pace. The frontier moves week to week, and they have been at or near the front of it consistently.

There is no world in which a firm should try to build its own AI platform. None. Anything you stand up will be obsolete in months, possibly sooner. This is a 99-out-of-100 buy decision, not a build. Pick a provider and ride their roadmap.

ChatGPT is a perfectly fine choice and many firms will be happy with it. But for more ambitious teams — firms that want to build real workflows, custom automations, and connect their own infrastructure — Claude is the better bet. It is a more open ecosystem. ChatGPT is tightly confined by comparison; good luck wiring a custom API into it without a fight. Claude's surface (skills, MCP, connectors, computer use) is built to be extended.

Honestly, you probably do not need an agent building platform like Writer, Lindy, or Relay. You can do almost everything you'd want to do through your LLM provider directly. The models themselves are becoming their own agents — the line between “chat model” and “agent” has completely blurred.

Claude with skills, connectors, and computer use can already execute multi step work across your systems. ChatGPT's new agents are the same idea. A separate agent layer adds another vendor, another set of credentials, another place context lives, and another thing to maintain — for capability you already have. Save the spend and the surface area.

Skills are the most important piece of this stack to get right today, and luckily they are also the easiest to build. A skill is a small, packaged instruction set the model loads on demand when the work calls for it — a memo template, a diligence pattern, a brand guide, a specific way you want a deliverable formatted.

They turn ad-hoc prompting into reusable firm IP. The same skill runs the same way for every analyst, every time, with the same quality floor. This is how you scale taste and judgment across a team without scaling headcount. Build them once, share them across the firm, version them as you learn.

The Skills Library in this site is the working set we use and recommend. Start there, fork what fits, and write your own as your workflows reveal themselves.

Instructions (the system prompt) and memory get passed to the model on every single message anyone on the team ever sends. Configuring sensible firm wide defaults here is enormously high leverage. A single well written system prompt raises the floor of every interaction the firm has with the model, forever.

This is where you set the firm's posture: how to disagree, how to cite, how to handle uncertainty, when to push back, what voice to write in. The LLM General Instructions in this library are our recommended starting point — you can drop them in as the team system prompt and tune from there.

Memory is the same idea on a longer timescale: facts about the firm, the people, the portfolio, and the recurring projects that the model should always know without having to be told again. Treat memory deliberately — what you put in is what shapes every answer that comes out. This isn't a one time job. Memory can degrade over time and become polluted with irrelevant information that crowds the context window without adding value. We've noticed that one time prompts that do not contain reusable useful info can end up here and have harmful effects on model responses.

Connectors are how the chat surface reaches into your actual systems of record — CRM, drive, calendar, Slack, the wiki, the data vendors. Without them, the model is reasoning in a vacuum. With them, you can ask a question in plain English and get an answer grounded in your firm's live data.

This is where Claude's open ecosystem really pays off. MCP (Model Context Protocol) has become the standard, and almost every tool worth using either supports it natively or has a clean wrapper. Wire up the systems the team actually uses, scope permissions properly, and let the model do the rest.