Twelve methods, fully open.

Do: Compare the current effort with one level lower on a representative evaluation set.

Measure: Track reasoning tokens, quality-pass rate, latency, and accepted-answer cost.

Watch: A lower setting is not a saving if it causes retries or misses important constraints.

Do: Set a task-appropriate maximum and ask for concise structure rather than an open-ended answer.

Measure: Compare output tokens, truncation rate, retries, and quality-pass rate.

Watch: A ceiling that truncates valid answers creates extra calls and can increase total cost.

Do: Map each requirement to one authoritative location, then remove only true duplicates in a paired test.

Measure: Compare input tokens and constraint-pass rate across at least three repeats per case.

Watch: Repetition can be useful when it fixes a measured failure; do not remove it on aesthetics alone.

Do: Remove one example, rerun the same evaluation set, and keep it removed only when quality remains inside the declared margin.

Measure: Track input-token reduction and per-example quality failures.

Watch: Deleting every example at once hides which one was valuable and makes regressions hard to diagnose.

Do: Choose a delimiter that cannot occur inside valid content, instruct the model to end with it, and configure the provider stop-sequence parameter.

Measure: Compare output tokens and incomplete-output rate, including delimiter-collision and escaping tests.

Watch: Stop support differs by model, and structured outputs are safer for complex JSON or content that may contain the delimiter.

Do: Use low verbosity for extraction, classification, routing, and other outputs where elaboration has no product value.

Measure: Compare output tokens, completeness, and retry rate.

Watch: Low verbosity can remove useful explanation from customer-facing or high-stakes answers.

Do: Add one short example that contains every required element and no optional commentary, then remove prose that the example makes redundant.

Measure: Compare median output tokens, required-element recall, and retry rate with and without the concise example.

Watch: The example adds input tokens on every uncached request, so keep it only when output or retry savings exceed that recurring cost.

Do: Tell the candidate to begin with the answer, omit greetings and conclusions, and never restate the task unless clarification is required.

Measure: Track output tokens, first-useful-token position, completeness, and the rate of confusingly abrupt answers.

Watch: Customer-facing explanations may need context or tone; do not remove framing that users demonstrably rely on.

Do: List the required answer elements in priority order and tell the candidate to add optional rationale only when budget remains.

Measure: Compare required-field recall, truncation failures, output tokens, and human preference on the same tasks.

Watch: Some reasoning-heavy tasks need explanation before a defensible conclusion; use the product acceptance rubric, not length alone.

Do: Give each allowed decision a stable compact label, ask for exactly one label, and map it to user-facing text in application code.

Measure: Compare output tokens, invalid-label rate, decision accuracy, and any repair calls against the prose response.

Watch: Compact labels hide nuance; keep a separate explanation field when downstream users need the reasoning.

Do: Replace generic adjectives such as clear, helpful, thorough, and professional with one testable sentence describing what the answer must let the reader do.

Measure: Compare input tokens, evaluator pass rate, clarification requests, and output length across a representative task set.

Watch: Do not collapse distinct safety, policy, or contractual requirements merely because their wording looks repetitive.

Do: Choose the smallest task-appropriate budget—for example three bullets or six fields—and state both the limit and the content that must fit inside it.

Measure: Track output tokens, budget violations, missing requirements, truncation, and retries against the current prompt.

Watch: Prompt budgets are soft controls; pair them with a safe API output ceiling for worst-case spend and leave headroom for valid edge cases.