CLAUDE.md Progressive Disclosurer

What it does

Turns "slim down my CLAUDE.md" into a measured, verifiable refactor:

• Profile first: /context share of the startup context, per-heading byte table (work in descending order), giant-line distribution, and downstream truncation limits (e.g. Codex project_doc_max_bytes, 32KiB default — anything beyond is silently invisible).
• Signal triage: the litmus is "if I delete this, will Claude get it wrong?" Anything inferable from the code, standard language conventions, or self-evident advice is an anti-signal — listed as a deletion candidate and removed only with user confirmation.
• Verbatim moves, compression forbidden: moving to Level 2 must copy text unchanged; 3+ sections use scripts/sink_sections.py with full-string verification and automatic rollback.
• Carrier selection matrix: decides whether a rule belongs in resident Level 1, a reference, a Skill, a Hook (blocking vs injecting), or ~/.claude/rules/ with paths: frontmatter.
• Three-part verification: reference existence (5a), content integrity (5b — an independent sub-agent audit is the default, forking banned), and line count explicitly excluded as a criterion (5c).
• Documents how the verifiers themselves lie: grep -F degrades to per-line OR on multi-line originals (false pass for lossy moves), symlinks silently skipped, regex metacharacters, bash vs zsh word splitting — all field-tested.

Who it's for

• Teams whose CLAUDE.md has grown so large that Claude keeps ignoring rules
• Anyone who wants to reorganize guidance without losing a single useful detail
• Users whose global ~/.claude/CLAUDE.md is polluted with project-specific paths
• People sharing the same doc with Codex or other tools that may truncate it silently

Examples

1. "Is my CLAUDE.md too long?" → the skill reframes (line count is a symptom, not a goal), then produces a per-section byte table and attacks the biggest contributor first.
2. Cleaning a 168KB global file → identifies it as 69% of session startup context, sinks 10 low-frequency SOP sections verbatim into ~/.claude/references/, and adds trigger-keyed index tables at the top and bottom.
3. Hunting fake pointers → a line says "see anti-patterns.md" but the content was never moved there; caught via distinctive-string grep -F plus a python3 whole-string containment check.

Install the skill I found on Claude Skill Mart.
Copy the daymade-claude-code/claude-md-progressive-disclosurer folder from the GitHub repo daymade/claude-code-skills into my ~/.claude/skills/claude-md-progressive-disclosurer/.
When it's done, tell me in one line what this skill can do.