Context: the one concept that matters

The one constraint behind everything

Two abilities come out of this lesson: sessions that stay sharp for hours, and the means to diagnose any teammate's 'Claude got dumb' complaint with one command. The mechanism behind both: Claude's working memory, the context window. Everything it knows during a session lives there - your messages, the files it has read, the output of commands it ran, its own replies. The window is large but finite, and quality drops as it fills with junk.

Anthropic's own best-practices guidance now centers on this single constraint. Once you get it, every confusing behavior makes sense: why long rambling sessions get dumb, why a fresh session nails what the old one fumbled, why specific file references beat vague descriptions.

What fills the window

• Your conversation - every message both ways.
• Files Claude reads - the full contents, not a summary.
• Command output - whatever its tools print back.
• Standing instructions - CLAUDE.md files and rules (next lesson).
• Skill descriptions and connected tools - small per item, real in aggregate.
• Auto memory - notes Claude keeps about your project (below).

And what Claude can NOT see: anything outside the folder you opened (unless you grant it), your screen, other apps, browser tabs, and anything from a previous session you cleared. If you find yourself thinking "but it should know this" - it almost certainly is not in the window. Put it there: @-mention the file or paste the content.

Run /context inside a session to see exactly what is loaded and how much room is left. Do it once today just to make the invisible visible.

The four context habits

Four commands manage the window. Together they are the difference between sessions that stay sharp for hours and sessions that slowly rot.

1. /clear between unrelated tasks. Finished the report and starting on the proposal? /clear. This is the default, not the exception.
2. /compact when a long session is still on-topic but heavy. It summarizes history to free space. You can steer it: "/compact keep the client requirements and the final file list".
3. /btw for side questions. It answers without the exchange ever entering your session's context. Curiosity stops costing you quality.
4. Double-press Esc to open /rewind. Checkpoints let you restore the code, the conversation, or both to an earlier point - they persist across sessions, and you can summarize up to a point instead of keeping everything.

Models and effort: the other two dials

Context is one dial on session quality. Which model is thinking, and how hard, are the other two. The next lesson is the full routing table with prices; here is the short version so nothing surprises you before then:

• Opus 5.5 - the default on every plan and the right starting point for most operator work: planning, judgment, anything a client reads.
• Sonnet 5 - the fast, cheaper option for high-volume interactive work where speed beats depth.
• Fable 5.1 - the top model, for the hardest long-running autonomous work. Opt-in only; reserve it for work that earns its cost.
• Haiku 4.5 - cheap and fast for bulk mechanical work, and the natural engine for subagents.

Switch with /model. /effort sets how hard the model thinks: low, medium, high, xhigh, or max. Both save your choice (per model, for /effort) as the new default, so if sessions suddenly feel different, check what you left them on; press s on either picker to make a choice session-only. The one habit to start today: turn effort down for renames and cleanups, up for planning and analysis. Maximum thinking on a file-rename task is pure waste.

Auto memory: what Claude remembers on its own

Claude Code now keeps its own notes per project, on by default. As you work, it writes what it learns - your preferences, project quirks, decisions - to a memory folder in your home directory, and loads the index back in at the start of each session in that project.

Practical upshots: saying "remember that we always exclude agencies under 10 people" stores it in auto memory and it survives /clear. The index auto-loads up to its first 200 lines, with deeper topic files pulled in on demand. Run /memory to see everything currently loaded and to manage or disable it.

Failure modes you can now name

• The kitchen-sink session: one session doing five unrelated jobs. Each pollutes the next. Fix: /clear between tasks.
• Correcting-over-and-over: the same fix attempted four times, conversation full of failures. Fix: two strikes, then /clear and one better prompt.
• Infinite exploration: "look around and tell me what you think" burns half the window on wandering. Fix: name the files and the question.

These three account for most of the "Claude got dumb" complaints you will ever hear. They are context problems, and now you can both diagnose and fix them - including for teammates, which is what makes you the local champion.

Do this now

Sources and further reading

• Claude Code documentation
• Monitoring usage - what your sessions consume