claude-md-compresslisted

# claude-md-compress ## Always run this skill in an Opus subagent Whoever triggers this skill delegates the actual work to a subagent (Agent tool) with `model: opus`. Never compress inline in the model currently running — Sonnet, Haiku, or any other loop worker that happens to hit this file hands the job off. This mirrors `.loopboard/LOOP.md` Rule 14, which routes New-task grooming to a subagent of the task's `groomer:` model instead of running it in the main loop. The subagent does everything below: read the file, compress, run the self-check, and — only on a pass — write the file and land it on a `task/**` branch (a PR is optional, Rule 7). ## Why this file is different from ordinary docs `CLAUDE.md` is not documentation. It is the project instruction file Claude Code loads into EVERY session in this repo, before any work starts — so its length is a fixed token cost paid on every conversation, and its content is what every agent working here believes about the project. It also `@`-includes other instruction files, so a dropped or reworded include line silently unloads a whole ruleset. "No information loss" is therefore a correctness requirement, not a style nicety. A dropped negation, a softened "NEVER", a lost scope qualifier, a mangled path, or a renumbered Non-negotiable is a behavioural change in every future session — including sessions that will never read the diff that caused it. **Lossless only.** This pass preserves meaning. Content that has gone obsolete —