skill-condenselisted

# Skill Condense Instruction files grow. Each fix adds a sentence, a war story, an example, or a paragraph on why the rule exists. Claude pays for every one of those tokens on every run, and a rule buried in narration is less likely to be followed. This skill rewrites a file down to what the model actually needs: each rule, stated once, plus a pointer to wherever the detail lives. It then proves that no rule was lost. If the skill might also contain contradictions or stale pointers, run `/skill-audit` on it first. Condensing text that is wrong only makes it shorter and still wrong. ## Who the Rewrite Is For Write for the model and effort level that actually run the file day to day, not the strongest one available. A terse pointer that a top model expands correctly can be too thin for a smaller model or a lower effort setting. A new model release is not by itself a reason to condense; the reason is visible bloat. ## Mode A: Triage (`/skill-condense` with no argument) Read-only. Consider `CLAUDE.md` files, `.claude/rules/*.md`, and every `SKILL.md` under the project's `.claude/skills/` and `~/.claude/skills/`. 1. **Load cost comes first.** It sets how often the saved tokens are paid: - **Always loaded** (CLAUDE.md files and rules without a `paths:` scope) are paid every session. - **Frequently used skills** are paid on each of their runs. - **Rarely used skills** are paid only on those rare runs. One line cut from an always-loaded file is worth many lines cu