Terminal Commands
Find commands that can flood context: tests, logs, Docker output, API responses, browser traces, and recursive directory reads.
Token cost checklist
Cut AI coding-agent token waste before buying more usage.
Claude Code, Codex, Cursor, browser agents, and CI agents get expensive when the workflow keeps feeding them avoidable output. Use this checklist to remove the common waste before scaling a team setup.
Command caps
Log hygiene
Agent rules
Fast fixes
Start with the places agents read too much.
Most cost waste is not mysterious. It comes from commands, logs, and status reports that are useful to a machine once, then expensive every time they are pasted back into context.
- Wrap any unknown-output command with
2>&1 | head -c 4000or a stricter cap. - Prefer summary tools such as
rtk git,rtk grep,rtk read,rtk diff,rtk test,rtk json, andrtk logwhen they fit. - Use
rgfor searches and point the agent at exact files instead of asking it to rediscover the whole repo. - Keep long generated reports out of the main transcript; write them to durable files and read summaries.
- After each savepoint, record the current state and the next command so the agent does not reread stale context.
What to inspect
The expensive surfaces are usually easy to name.
Instruction Files
Check AGENTS.md, CLAUDE.md, Cursor rules, and global instructions for command-output rules, rtk usage, and stop-loop rules.
Loop State
Look for agents rereading old reports, screenshots, ledgers, and raw logs instead of using a compact current-state file.
Tool-specific audits
Use the same cleanup discipline in the tool your team actually runs.
Claude Code token cost audit
Review the repo instructions, command output, and repeated scans that make Claude Code sessions heavier than needed.
Claude Code AuditCodex token cost audit
Clean up rtk usage, byte caps, loop-state files, and browser or API checks that flood Codex context.
Codex AuditAgent cost audit
Use the general audit for Cursor, browser agents, CI agents, or mixed coding-agent workflows.
Buy AuditDecision rule
Buy cleanup when the same waste repeats.
Small issue
Add command caps and a repo instruction. This is often enough for one-off work.
Recurring issue
Audit the workflow if agents repeatedly read huge files, logs, or generated artifacts.
Team issue
Roll out shared rules, templates, and token-cost checks across repos before usage grows.
Need the cleanup done?