This is the version you keep open while you work. Pull it up mid-session with Claude when you want a quick answer: which command stretches the session limit, or what to do after the last reply. Nothing beyond what the source article covers — just the commands, thresholds, and habits, no new numbers.
The rule in one line: the session limit gets eaten mainly by re-reading the conversation, so cut the history, not just single messages.
Terms at a glance
| Term | What it means |
|---|---|
| Token | the smallest chunk of text the model reads and gets billed for (roughly a piece of a word) |
| Session | one continuous conversation with the model |
| Session limit | the usage ceiling that forces a wait for reset before you can keep working |
| Model | the AI engine itself that answers (Claude in its various versions, for instance) |
| Context window | how much text the model "sees" at once: system instructions, the whole conversation, tool calls and their results, files it has read — its current working memory |
| Claude Code | the tool Claude works in on coding and operational tasks; a context window on the order of a million tokens |
Where the cost comes from
- The model re-reads the whole conversation on every message. First question, first answer, second question, second answer, and so on up to the latest instruction. That's why the cost doesn't add up, it compounds.
- Re-reading history eats most of the budget. In one traced conversation of over a hundred messages, 98.5% of all tokens went purely to re-reading earlier history.
- Part of the window is spent before you type anything. The project instructions file, connected tools, and skills all load at startup. A typical overhead is around 8,000 tokens, but it can run as high as 62,000 on a fresh session.
- Context rot. The longer the session runs, the worse the quality gets — the model's attention spreads across more and more tokens, it loses the thread, and it edits files without reading them. Retrieval accuracy inside the window drops from 92% at 256,000 tokens to 78% at a million.
Five options after every reply
What you can do once the model finishes answering.
| Option | What it does | How to trigger it |
|---|---|---|
| Continue | you write another message; the history keeps growing | just send a new message |
| Rewind | you jump back to an earlier message and drop everything after it | /rewind |
| Clear | you start completely fresh | /clear |
| Compact | you summarize the session and replace the history with that summary | /compact |
| Delegate | you hand the task to a sub-agent: a fresh context window, only the result comes back | "spin up a sub-agent to verify this" |
The single most important habit is /rewind. When the model gets something wrong, most people type "that didn't work, try again" — but the failed attempt and the broken code stay in context and get re-read with every following message. /rewind jumps back to before the mistake and leaves the context clean.
Clearing, compacting, and summaries
- New task → /clear. You're starting something else, so start from zero. - Same task → /compact. Compacting rolls the conversation so far into a summary. - Watch out for auto-compact. Claude Code only triggers it on its own at around 95% of the window filled — too late. By then only about 20-30% of the original detail survives, and the summary gets written at the peak of context rot. - The manual version (usually the strongest lever). Before you hit the limit: 1. ask the model for a full summary — what's done and what's next; 2. run /clear; 3. paste the summary back in and keep working on a fresh session. - The catch with the manual version: since you're losing the conversation history, write your decisions down somewhere else — plan files, decision logs, task lists.
Sub-agents
- What a sub-agent is. A separate helper with its own, fresh context window: it does the task and sends only the result back to the main session. The grunt work happens outside your window, so it doesn't clutter the main conversation.
- How to delegate. Just say it: "spin up a sub-agent to verify this."
- A cheaper model. A sub-agent can run on a cheaper model, which cuts cost further at comparable quality.
- What to develop. A feel for which tasks are worth delegating.
Fewer tokens at the source
Tricks that cut usage before the conversation even starts.
| Trick | What it buys you |
|---|---|
| Convert files to markdown | converting from HTML trims about 90% of the tokens, from PDF roughly 65-70%, from docx about 33%. Exception: scans that need image-based text recognition, where the rule doesn't hold |
| Keep the instructions file lean | claude.md loads on every session; keep it under roughly 200 lines (about 2,000 tokens), and move the rest into skills and files read on demand |
| Start with a plan | planning before you build spends tokens up front but saves them later, since the model wanders into fewer dead ends |
| Watch the limit indicator | seeing how much session is left changes decisions: send the instruction now, or spin up a team of agents |
Checking the start of a session: in Claude Code, the /context command shows how much of the window is already spent before you've typed a word.
Why filling the whole window backfires
- A bigger window isn't a better result, just more room for context rot. The model's behavior doesn't change.
- Numbers from an analysis of 18,000 thinking blocks across 7,000 sessions: reasoning depth dropped 67% as sessions got longer, and editing files without reading them rose from 6% to 34%.
- An extreme case: one user, through bad habits, drove their token bill from $345 to $42,000 a month, while the quality of the work stayed flat.
- The productive stretch is the first ten to twenty percent of a session, while the model is freshest.
- When a session has gone off the rails, don't fight it. Open a new one and start clean.
Where to start
- Build the
/rewindreflex instead of "try again" — the habit the article recommends hardest. - Split work by command: new task →
/clear, same task →/compact. - Use the manual version before you hit the limit: summary →
/clear→ paste the summary back in, and write decisions down in files. - Hand grunt work to sub-agents, ideally on a cheaper model.
- If you're just starting out, stay on the smaller, two-hundred-thousand-token window and build the discipline: clear the session, log progress, delegate.