What “Persistent” Actually Means Here
The foundation is a file called CLAUDE.md. It’s a plain markdown file that Claude Code reads at the start of every conversation, so your project conventions, the commands, the folder structure, the things you’d otherwise repeat, are loaded before you type anything. You can even nest them, dropping a separate CLAUDE.md inside a tests folder or a database directory for rules that only apply there.
Skills go a step further. Instead of instructions you write once, a skill is a packaged capability, often a folder with its own docs and scripts, that Claude pulls in on its own when a task matches. You don’t call it manually. It gets discovered.
Subagents are the third piece, and they’re the one that makes the whole thing scale. A subagent is a separate Claude instance with its own context window. When Claude needs to read a pile of documentation or scan a big repo, it hands that off to a subagent, which does the messy work in isolation and returns just the answer. Your main conversation stays clean and doesn’t blow past its limits.
The Patterns Developers Keep Repeating
Read enough of these guides and the same advice comes up. Keep the context lean, because a bloated conversation makes Claude drift and forget its own instructions. Turn anything you do more than once into a skill instead of re-explaining it. And let Claude ask you clarifying questions before it starts, since a thirty-second back-and-forth up front beats debugging a confident wrong answer later.
There’s also a structure people keep recommending for the actual prompt: spell out the context, the role, the goal, the instructions, examples, constraints, the output you want, and a review step. It reads like overkill until you’ve watched a vague request produce vague code.
On cost, the official tip is prompt caching. If you’re sending the same large chunk of stable context over and over, a big CLAUDE.md or a set of reference docs, Anthropic’s Claude Platform guidance is to cache it so repeated calls run cheaper and faster instead of paying to reprocess the same tokens every time.
Anthropic Is Leaning In
The clearest signal that this isn’t just community enthusiasm is who Anthropic is hiring for it. Addy Osmani announced he’s joined the company to work on Claude Code and, in his words, make it better for the developers who use it. He spent 14 years at Google, most recently as a director in Google Cloud AI leading Gemini’s developer experience, and he’s the name behind tools a lot of web developers use daily, including Chrome DevTools, Lighthouse, and Core Web Vitals.
Pulling someone with that background off a direct competitor’s flagship AI product says something about where Anthropic thinks the fight is. Not just building a smarter model, but making the thing developers actually sit in all day worth staying in.