Key answer
Specify relevant resource limits, distinguishing desired budgets from enforced stopping points. Time, cost, attempts and output depth differ; short answers may follow many searches, so word limits do not cap cost.
Full explanation
Define required artifacts first, time/cost ceilings, what counts as an attempt and what to deliver at the limit. Give deadline date/time zone; distinguish a revision round from a tool call.
Claude's task budgets are advisory token budgets for self-management. Managed Agents session budgets are checked between model requests; reaching a limit stops new requests while in-flight requests and associated tools may finish. Inspect both measurement scope and enforcement timing. Task budgets; session budgets
For example:
Deliver a usable one-page plan and key evidence, not a comprehensive review. Limit research to two rounds, each a batch of queries for current gaps. Start writing once evidence suffices; mark unresolved points after round two. Do not start paid services. Report completed scope and remaining gaps.
This limits research while retaining organization, checks and delivery; exhausting searches should not leave scattered material. For strict fees, set real platform limits. If a tool cannot read time/usage, identify estimates.
Use actual usage and output quality, not claims of effort. Before expanding budgets, identify the specific gap additional work should resolve so “one more try” does not become endless.