Skip to content

cost-optimization.md

Latest commit

 

History

History
134 lines (79 loc) · 3.7 KB

cost-optimization.md

File metadata and controls

134 lines (79 loc) · 3.7 KB

Cost Optimization Tips

Reduce token usage and API costs without slowing down your workflow.

Check context usage with /context

At any point in a session, run /context to see how many tokens are in the current context window and how much has been spent so far. Use this to decide whether it is time to compact or clear.

/context

Show token count in the status line

Configure the status line to display token count and cumulative cost at a glance, so you never need to run /context manually. See claude-code-docs/statusline.md for the exact configuration.

With this set up, you can see context growth in real time and compact before it gets expensive.

Clear between tasks to reset context

Every task you run in the same session adds to the context window. Unrelated tasks pile up and increase cost without adding value.

/clear

Use /clear at the start of each new, unrelated task. Token cost resets to zero.

Use subagents for research to keep main context clean

Research tasks (reading documentation, exploring a large codebase, investigating a bug) can consume thousands of tokens. Running them in a subagent means the main session only receives the final summary.

Use a subagent to read through src/payments/ and summarize how the refund flow works.

The subagent's context is discarded after it reports back. Your main session stays lean.

Use Haiku for simple subagent tasks

When a subagent is doing straightforward work (reading a file, checking a value, running a search), you do not need the full Claude model. Configure it to use Haiku.

Use a subagent (model: haiku) to check if any files in src/config/ contain the string "DEBUG=true".

Haiku is significantly cheaper and fast enough for simple lookup tasks.

Scope file reads to the lines you need

Asking Claude to read a 2000-line file when you only need a 50-line function wastes tokens. Be specific.

Read lines 50-100 of src/api/users.ts

Or reference the function by name and let Claude find it efficiently:

Show me only the `validateToken` function in src/auth/token.ts

Compact with focus instructions

/compact without instructions compresses context but may drop details you still need. Focus instructions tell it what to preserve.

/compact focus on the migration plan and the schema changes we agreed on

This keeps the critical thread while dropping conversational filler and exploratory tangents.

Use /btw for quick questions

/btw asks a question without adding it to the main conversation history. It is invisible to future context.

/btw what does the --recursive flag do here?

Use it for quick lookups and clarifications that you do not need remembered for the rest of the session.

Set a budget cap with --max-budget-usd

For scripted or automated runs, set a hard cost ceiling so a runaway task cannot exceed your budget.

claude -p "analyze all endpoints for security issues" --max-budget-usd 2.00

The session stops when the budget is reached. Useful for batch jobs and CI pipelines.

Use /effort low for simple tasks

For tasks that do not require deep reasoning (renaming a variable, reformatting a file, adding a comment), tell Claude to use less effort.

/effort low
Rename the variable `tmp` to `userRecord` throughout src/models/user.ts

Less reasoning effort means fewer output tokens and lower cost.

Use fast mode for faster, cheaper output

/fast switches to a faster output mode on the same model. Responses are quicker and typically cheaper, suitable for tasks where you value speed over thoroughness.

/fast

Toggle it off with /fast again when you need careful, detailed analysis.