Claude Code usage is not a simple message counter. Ten short questions and ten debugging turns that read large files and build logs do not consume the same amount of input.
Two separate limits are easy to confuse:
- A usage limit controls how much Claude an account can use over a period.
- A context limit controls how much information the model can consider in one session request.
They are not the same quota. They still interact: as a session grows, prior conversation and work results can make each later request larger, so account usage may be consumed faster.
One message does not have a fixed size
A Claude Code request can include previous conversation, project instructions, the new prompt, and relevant results from files or commands.
First request
= startup instructions + first prompt
Request late in a debugging session
= startup instructions
+ prior conversation
+ files and command output
+ new promptThis is why perceived usage varies with session length, task complexity, model choice, and effort. Claude in other surfaces can also count toward the same account-level allowance, depending on the plan.
Identify which limit you reached
| Symptom | Likely cause | Where to check | Response |
|---|---|---|---|
| A reset time appears and work stops | Account usage limit | Settings → Usage | Wait or change the work pattern |
Prompt is too long |
Session context is full | /context |
Use /compact or /clear |
| Automatic compaction repeats | Context fills unusually fast | /context |
Inspect large output and rules |
| API cost is higher than expected | Long sessions or model cost | /cost, API console |
Review session and model choices |
/compact does not restore account usage. It summarizes the conversation to free context space; it does not refund tokens or reset an allowance.
Use /context to find what accumulated
Inspect the current session with:
/contextConversation history is only one contributor. Project instructions, files, command output, memory, rules, and tool information can all consume context. Deferred MCP definitions may reduce the startup footprint, but tools you use and their returned data still add information to the session.
If you see Prompt is too long, look for a large file or command result first. Also review unused MCP servers, long root instructions, and verbose build logs.
Use /clear when the task changes
A completed authentication investigation rarely needs to remain in context when the next task is a database migration.
/clearThis does not delete project files. It only removes the previous conversation and its context. Before clearing, put conclusions that matter into code, tests, an issue, or a short project note.
Use /compact when the same task continues
When you are still tracing one bug, the rejected hypotheses and failed fixes may matter. Compact the session while naming what must survive:
/compact Preserve the redirect cause, rejected hypotheses, changed files, and failing testsNaming what to preserve keeps the summary focused. If the context is already completely full, compaction itself may fail. In that case, Claude Code may need to drop recent messages or start a new session.
Root instructions also consume context
Claude Code loads CLAUDE.md as persistent project context. That is useful for commands and rules needed in most tasks, but it means unrelated material follows every request.
Anthropic’s guidance suggests keeping a CLAUDE.md concise and uses roughly 200 lines as a rule of thumb. That is not a product limit or a point where performance suddenly changes. Scope matters more than line count.
Keep repository-wide verification commands and non-obvious project rules at the root. Move one-off deployment procedures, stale incidents, and path-specific instructions closer to the work that needs them.
If Claude stops accepting prompts
- Check Settings → Usage to determine whether this is an account limit.
- If the session still runs, use
/contextto identify large contributors. - Use
/compactfor the same task or/clearfor a different task. - Reduce large logs, unnecessary tools, and stale persistent instructions.
API users can inspect session cost and tokens with /cost. Subscription users should not interpret /cost as their remaining plan allowance; the Usage screen is the relevant account view.
Clear finished work; compact ongoing work
Do not estimate Claude Code usage by counting messages. A long session carries more prior work into later turns.
First distinguish an account usage limit from a full context window. Clear finished work, compact continuing work, and use /context when you do not know what is taking space.