Interfaces, features and setup steps for Claude Code, Cursor and other tools are likely to have changed. Focus on the concepts, and check the provider’s current documentation when following along.
Watch, pause, and follow along in your own workspace.
Claude's working memory. Think of it like milk: best when it's fresh and condensed. Overloading it causes drift, hallucinations, and lower quality output.
Two rules to follow:
Never let the context fill up. Run /compact when the ring hits 50%.
Keep separate tasks in separate conversations. Don't mix unrelated work in one chat.
Commands you'll use constantly
/clear — wipes the context completely. Use when starting a new task.
/compact — summarizes and condenses context to free up space. You can specify what to preserve before running it.
/context — shows what's loaded in your session and how full the context is.
/init — lets Claude read your project folder and generate a CLAUDE.md for it.
What loads into context at session start
Your CLAUDE.md files (global + project)
Skills
MCP connections (tools)
Auto-memory notes Claude has saved
Keep this layer lean. Everything that loads at start costs tokens — and shapes every response Claude gives you.
CLAUDE.md: two levels
Global — applies to all projects on all machines. Lives at ~/.claude/CLAUDE.md.
Project — specific to one project folder. Lives at the root of that folder.
Rule of thumb: only put something in CLAUDE.md if it needs to be in every conversation. If it's situational, load it on demand.
Skills are how you turn Claude from a generalist into a specialist. Each skill is a folder containing a SKILL.md file. You can teach a skill what good output looks like, give it templates, and refine it over time with feedback.
Project-level skills — only available in that project
User-level skills — available across all projects
One way to build a strong skill: gather real expert opinions and embed them as reference material. Superme.ai is useful for this — it surfaces expert perspectives you can use to ground your skill's logic and outputs.