token-efficiency

Token Efficiency

Practices to minimise token waste across all sessions.

Context Management

• Do not re-send large context blocks. Reference existing files by path instead.
• Use /compact when context window grows large.
• Load only relevant file sections, not entire large files.
• One task per session where practical. Flush context between unrelated tasks.

Output Format

• Return diffs, not full file rewrites.
• Keep explanations concise — bullets over paragraphs.
• Do not repeat information the user already knows.
• Do not generate boilerplate unless asked.

File References

• When prior context is needed, read doc_internal/CONTEXT.md (if it exists in the project) rather than re-generating from scratch.
• For architecture context, reference doc_internal/ARCHITECTURE.md rather than re-explaining the system.

Task Budgets (Opus 4.7, public beta — API only)

An advisory token budget for the full agentic loop, not a single request. Currently public beta on Opus 4.7 via the API; not available in Claude Code or Cowork at launch.

• Set via output_config.task_budget: {type: "tokens", total: N} in the API request.
• Minimum total: 20,000 tokens.
• The model sees a running countdown of the budget and self-regulates: prioritizes higher-value work first, finishes gracefully as the budget approaches zero rather than getting cut off mid-thought.
• Advisory, not enforced. max_tokens is still the hard cap per individual request. task_budget shapes behavior across the loop; it does not stop a request from completing.

How it composes with effort:

• effort controls depth per step (how hard Claude thinks about each turn).
• task_budget controls breadth across the loop (how much total work Claude attempts before wrapping up).
• Use both together for long agentic runs where you care about end-to-end token cost.

/btw — Side Questions Without Polluting Context

Opus 4.7 introduces /btw <question> for quick sidebar questions that never enter conversation history.