Making It Yours: Skills, Memory, and the Long Run
Everything so far has been one-off — a good prompt, a good session. The last step is making the good stuff reusable and durable, so you stop re-explaining yourself.
What changed›
- New chapter — added in the June 2026 restructure.
You've been teaching Claude the same things over and over — your conventions, your taste, how you like the output. The jump from using it to building with it is making that knowledge stay.
Every session so far started cold. You re-explained your house style, re-pasted the same context, re-stated what “good” means to you. That repetition is the signal that something should become permanent. There are three places to put it, and choosing the right one is most of the craft.
Skills: package the expertise once
A Skill is an organized folder — instructions, and any scripts or files they need — that Claude discovers and loads on its own when a task matches. At its heart is a small SKILL.md with a name and a one-line description. The clever part is progressive disclosure: only that name and description sit in context normally; the full instructions load when the task actually calls for them, and bundled files only when they're needed. So your packaged expertise is effectively unbounded, but costs almost nothing until it's used.
The placement rule is simple. Context that's broadly true for a project goes in CLAUDE.md. Know-how that's only sometimes relevant goes in a Skill, loaded on demand. And something that must happen every single time, with no room for judgement, goes in a hook — deterministic, not left to the model to remember.
Memory: surviving the long run
A big task outlives a single context window — the desk fills, and early details slide off the edge. The move that holds a long run together is external memory: state written down where a fresh window can pick it up — a progress.md, the git history, a tests file. Models are good at discovering state from the filesystem, so when a session gets long and foggy, you're often better off starting a clean window that reads the notes than nursing a bloated one. Write down what “done” means and where you are; let the next window stand on it.
Pick the right engine
Finally, the model is a choice, not a fixed thing. Opus for the hardest, longest-horizon reasoning — and it's the one with an effort dial you can turn up when the problem deserves more thinking. Sonnet for everyday agent and coding work, the workhorse. Haiku when speed and cost matter more than the last few percent of depth. Match the engine to the job; you don't send the heaviest model to do the lightest task.
Where it breaks
A long autonomous run with nothing written down. Three hours in, the thread has forgotten what “done” meant, and it drifts — confidently — off the goal. External memory isn't bureaucracy; it's the thing that keeps a long run pointed where you aimed it.
Try it yourself
Take the thing you explain to Claude most often — your format, your tone, your one non-negotiable rule — and write it once as a short SKILL.md: a name, a description, the instructions. Next time the work matches, watch it apply without being asked. That's the moment Claude stops being a tool you operate and becomes one you've shaped.
Grounded in Anthropic's writing on Agent Skills (progressive disclosure), effective context engineering (external memory), and the current Opus / Sonnet / Haiku lineup with the effort control.