Save Tokens

Save Tokens

Optimize how the task is executed, not how much necessary work is omitted. Correctness, safety, user instructions, applicable skills, and proportionate verification always take precedence over token savings.

Lean workflow

1. Define the requested outcome and the smallest plausible scope.
2. Search before broad reading. Prefer rg or rg --files, then read only the relevant regions. Fully read any selected skill or instruction file when its rules require that.
3. Batch independent searches, reads, and checks when doing so keeps output understandable.
4. Reuse facts already established in the task. Do not rerun unchanged commands unless fresh evidence is required.
5. Make a cohesive edit after the relevant context is known, then verify in proportion to risk.
6. Report the outcome, evidence, and remaining limitation concisely.

Context and output discipline

• Exclude generated, dependency, cache, and build-output directories from broad searches unless they are explicitly in scope.
• Use targeted line ranges for large files, but expand the read when behavior crosses the initial range. Do not impose a fixed file or tool-call limit.
• Cap noisy command output at the source when possible. Preserve the error, failing test, summary, and enough surrounding context to diagnose it.
• Avoid pasting large source blocks or logs into conversation when a precise path, line, or short excerpt is enough.
• Keep required commentary updates short and useful. Do not suppress updates required by the active Codex instructions.

Parallel work

When a domain orchestration skill applies, it owns delegation, resource leases, and integration. This skill only trims each context packet and returned output. Otherwise delegate only a bounded independent package whose saved context or parallel value exceeds its setup and review cost; avoid overlapping writers.

Never save tokens by

• skipping a required safety check, relevant test, or requested artifact;
• guessing about unread code, current external state, or unstable facts;
• hiding blockers, uncertainty, or test failures;
• asking the user to repeat information already available locally;
• creating a new task without the user's explicit request.

Read references/efficient-work.md for search, batching, output-trimming, and long-task patterns. If the user explicitly wants to continue in a fresh Codex task, use $session-handoff to preserve only the context the next task needs.