Reduce Claude token use: conversations, files and responses

Tokens are used beyond questions and answers: previous conversation, instructions, file contents and tool results can enter the model’s input. A short answer can still involve substantial reading.

Long conversations add context, but previous content is not always charged the same way again. Caching and compaction may apply, so token counts and billed amounts are not interchangeable.

Start a new conversation when changing to an unrelated topic if the old history will not help. If continuing the same problem, keeping necessary context may be better.

Before switching, ask for a summary of decisions, changed files, remaining work and verification results. Provide it with relevant file paths in the new conversation.

For large work, record results at each stage. Compare the effort of reexplaining and rereading material rather than automatically cutting conversations short.

If you know the location, give the file path and function name. “Check the date-conversion function in this file” narrows the search more than “fix this project.”

For long documents, specify the relevant section or provide the file and ask it to start there. A file alone does not automatically reduce usage; also specify the scope to read.

Keep enough context and constraints. Omitting key details and repeating work may cost more time and tokens.

If you only need results, ask for a format such as “Tell me the changes and checks only.” A summary might be three sentences; a comparison might be a table.

The effect varies. When inputs are large, shortening output alone may have little effect on total usage. Measure usage and adjust the largest of conversation, files and output first.

If you keep pasting report formats or review rules, collect them in a skill so shared instructions are easier to reuse and update.

Reading a skill also consumes tokens. Moving instructions to a file does not eliminate cost. Separate long references and tell the agent when to read them.

Put the use case and core steps in SKILL.md; separate long examples and details into references. Repeatable code can go in a scripts folder.

Organize material so it can be found. More important than the number of files is explaining which file to read in which situation.

Skills such as Caveman reduce unnecessary wording for concise answers. Try them when a task calls for results that are quick to read.

Savings vary by input, model and task. Do not interpret a listing’s claim as overall bill savings; compare accuracy and usage on your own work.

Some tasks become difficult to understand when explanation is cut. Retain the needed detail in beginner guidance or work requiring a review rationale.