Reduce agent context-window tokens

Context windows fill from more than chat history. Always-on instruction files, tool schemas, MCP output, and long sessions all compete. The cheapest wins are usually the ones you pay every turn — starting with AGENTS.md, CLAUDE.md, and Cursor rules.

1. Measure the always-on tax

Drop the instruction set into Clogmeter or use the AGENTS.md counter / Cursor rules counter. Note total tokens versus a 32k / 128k / 200k window before you rewrite prose.

2. Stop loading what the task does not need

• Switch Cursor rules from alwaysApply: true to globs or description triggers.
• Move long procedures into skills or docs the agent opens on demand.
• Keep one canonical rule file; delete or @-import duplicates (imports still cost tokens when loaded).

3. Clean the session, not only the files

• Between unrelated tasks, prefer a clear/new session over endlessly compacting a huge thread.
• Filter noisy tool/MCP output; return summaries instead of raw logs.
• Match model size to the job — smaller models for mechanical subagents.

4. Fix conflicts before micro-edits

Conflicting always/never lines teach the model to hedge and still spend tokens twice. Clogmeter flags duplicates and possible conflicts so you can keep one source of truth first — see also cut conflicting rules first.