context-budgetlisted

# Context budget Inventory the instructions, enabled tools, loaded skills and repeated output actually used by this session. Distinguish discoverable skill descriptions from skill bodies loaded on demand. Do not assume a fixed context window or a universal token cost per tool schema. Use provider usage/context information when available. Label character-based estimates as estimates; never present a sample calculation as measured savings. Prefer targeted searches and bounded reads. Keep task reports concise and preserve decisions, branch, open questions and failing checks when summarizing. Reuse a conversation when its context is relevant; avoid appending the same instructions to every turn. Expose only needed MCP servers and skills. Extra automatic model subprocesses consume real quota and can escape the office's accounting. Do not install a full marketplace plugin or auto-fix harness as an incidental optimization. RTK can compress supported command output; keep access to raw failure output and verify that compression preserves the information needed for the task.