Developer shares a 15-line SKILL.md to curb Claude Code token burn

On October 30, 2026, the AI section of DEV Community published a developer walkthrough describing a small configuration file meant to keep Claude Code agents from burning tokens inside unproductive loops.

The author runs several parallel Claude Code sessions on a single project each day, and reports that the cheapest fix they found was not a bigger usage limit. Instead, they wrote about 15 lines of guidance into a SKILL.md file that codifies two habits. First, brief the agent like a colleague rather than issuing commands: before launching a subtask, spell out four things clearly. Second, when a command fails, stop and rethink instead of retrying the same command unchanged; the article notes that an unattended agent will often repeat a failing command five times, with each retry consuming tokens, or keep re-checking a file that has not changed. Third, after a result comes back, the agent may ask at most three follow-up rounds, and must give a reason for each one. The author summarizes the underlying problem as the expensive part being not the work itself but the work that goes nowhere.

The SKILL.md is a plain-text file placed in the skills directory, after which the guidance is meant to be reusable across sessions. No follow-up coverage or independent confirmation of the effect was included in the report.