Aiclean — Security skill for Claude Code

aiclean

Audit and clean up a Claude Code setup against Anthropic's current prompt-engineering guidance.

Most `~/.claude` directories drift the same way. They accumulate instructions written to patch weaknesses the models no longer have, plus skills installed once and never used. Both cost tokens on every session, and some of it actively degrades output — the guidance is explicit that older, over-prescriptive skills "can degrade output quality."

This skill finds it and asks before touching anything.

Install

Claude Code

/plugin marketplace add visualic/aiclean
/plugin install aiclean@visualic

Codex CLI and other harnesses

git clone https://github.com/visualic/aiclean.git ~/aiclean
cd ~/aiclean && ./install.sh

Links the skill into every harness home it finds (`~/.codex`, `~/.agents`, `~/.cursor`, `~/.gemini`, `~/.hermes`). Because it links rather than copies, `git pull` updates all of them at once. `./install.sh --list` shows what it would touch; `./install.sh --uninstall` removes it cleanly.

Use

Open a session and ask:

내 클로드 설정 점검해줘

or run `/aiclean` in Claude Code.

It gathers evidence, reports findings, and waits. Nothing is changed without your say-so, nothing is deleted — approved removals are moved to `~/.claude/skills-archive/` with the reason written down, so restoring is one `mv`.

What it checks

**CLAUDE.md** — instructions that current guidance says to remove (generic self-verification, "use subagents liberally", manual chain-of-thought, hard thresholds like "ANY task with 3+ steps", duplicate memory systems) and the ones most files are missing (response length, written-document length, task scope, correction narration, evidence behind progress claims).

**Skills and agents** — `SKILL.md` bodies large enough to cost tens of thousands of tokens per invocation; skills with no `description:`, which the router cannot see and which therefore can never be invoked; inert custom frontmatter keys; custom skills sha