Two small controls point to different kinds of context waste
Anthropic’s Claude Code 2.1.261 release on September 4 adds `/skill-doctor`, which reports loaded skills that went unused and their context cost. It also adds `bashOutputMaxChars` and `taskOutputMaxChars` settings that can raise how much command or background-task output is included inline before the rest is saved to a file, up to 128K characters. These are separate controls: one helps inspect context that was loaded unnecessarily; the other changes how much tool output is brought into the active conversation.
That distinction matters because teams often use “the agent forgot” to describe several different failures. It may have received irrelevant instruction, missed a crucial line in truncated output, or simply been given an unclear task. More context is not automatically better. The useful response depends on where information was lost and whether it was needed at all.
Prune unused skills when their instructions are not needed
Begin with a representative set of sessions, not one unusual run. List which skills were enabled and what task each session was meant to complete. Inspect the skill report and ask whether the unused material was appropriate to load for that task. A skill can be unused in one session because the work never reached that step; that does not make it unnecessary for every workflow. Conversely, a broad default pack can add instructions that compete with the specific job.
Make the change narrowly. Separate always-needed guidance from task-specific skills, and test the common task plus an exception that should activate the specialized material. Compare whether the model follows the intended process, whether the person needs to repeat instructions, and whether any important rule disappears when a skill is no longer loaded. Keep the previous configuration until the tests pass. Do not remove a skill solely to reduce a context figure if its instructions govern a critical action.
Raise output caps only when the missing lines matter
A larger inline output limit can help when a command produces a report whose later section changes the decision. It can also flood the conversation with logs, duplicated records, or irrelevant content. Before changing the limit, inspect what the tool writes and how it behaves when output exceeds the current setting. If a concise summary, a filtered query, or a file inspection command can surface the needed evidence, that may be more reliable than importing a much larger stream.
Choose an explicit task and define the signal it needs. For example, a developer may need the final test summary and the failure excerpt, while a large build log contains thousands of lines unrelated to that decision. Adjust the command to preserve the relevant output where possible. If raising a cap is necessary, start with a bounded task and check whether the model can identify the relevant section without overlooking warnings or secrets.
Diagnose before combining the fixes
When an agent misses a requirement, reconstruct the path: what instruction was loaded, what command ran, what result was returned, and what the model saw. If a skill was present but irrelevant, prune or scope it. If the output was truncated at a critical point, improve the command or set a more suitable cap. If the data was present but the model ignored it, refine the task or test the model behavior. Avoid changing all three at once, or the next run will not tell you which adjustment helped.
Keep an operational record of the setting change, its owner, the tested workflow, and the rollback. A developer tool’s context controls are part of the workflow contract. They deserve a small test and a maintenance owner, especially when the tool is used across a team with shared settings.
A simple review habit beats a large default context
Once a month or after a meaningful workflow change, look at a sample of sessions and ask whether the right instructions and evidence arrived. Remove stale defaults, keep specialized guidance available only where it belongs, and revisit output limits when commands or logs change. This is not a hunt for the smallest possible context. It is a way to provide enough relevant information to do the task and avoid paying attention to noise.