chore(claude): disable redundant plugins and trim skills that restate model knowledge

[evansenter]

This is the third PR updating the setup for the Opus 5.5 generation. It covers plugins and skills. It follows #353, which removed every command and agent dependency on the plugins disabled here.

Plugins

Usage was measured by searching the 263 local session transcripts for skill and subagent invocations from each plugin.

Plugin	Sessions	Why it goes
commit-commands	101	Almost all came from /pr-create calling it; now explicit git/gh steps
pr-review-toolkit	94	/pr-review's fallback; the built-in code-review and simplify cover it
feature-dev	45	/work's explorer and architect phases, now done directly
superpowers	10	Also injected a long "YOU MUST" preamble into every session, pushing process skills (brainstorming, plan-writing, mandated verification) that the global CLAUDE.md now steers away from
ralph-loop	2	Superseded by the built-in /goal and /loop
claude-md-management	2	Overlaps the audit-docs agent
skill-creator	0	Still available as anthropic-skills:skill-creator
claude-code-setup, agent-sdk-dev, frontend-design	0	Unused

Kept: explanatory-output-style (the ★ Insight hook depends on it) and the four LSP plugins (rust-analyzer, gopls, pyright, typescript).

Disabling a plugin removes its skill and agent descriptions from every request and from the session-start context. To bring one back, add its enabledPlugins entry again.

Skills

• Removed rust-async-patterns: 524 lines of general Tokio and async-Rust material with nothing specific to these projects. The model already knows all of it.
• shell-scripting: its example local result="$(command)" is exactly what shellcheck SC2155 flags, which contradicts the skill's own "all .sh files must pass shellcheck" rule. It now shows the split form that bootstrap.sh already uses.
• summarize: keeps the summarize CLI usage, the part the model can't know. Drops the table of how to access each input type, the fixed output template, and the "~200 words" cap. Its description no longer fires on any "what is this"; it now targets links and media.

Docs: updated the shared-components table in docs/work-review-cycle.md, deferred here from #353's review.

make check passes, and settings.json is valid JSON. The diff only touches enabledPlugins.

🤖 Generated with Claude Code

https://claude.ai/code/session_01LBftoc1JCMLow9LDFZMADs

[claude]

Prompt: evansenter/dotfiles/.../claude-review.md

Code Review — Round 1

Summary
Turns off 10 plugins in settings.json (keeps explanatory-output-style and the four LSP plugins), deletes the general-purpose rust-async-patterns skill, fixes the SC2155 example in shell-scripting, and trims summarize down to the CLI usage. settings.json stays valid JSON after the trailing-comma change.

Checked:

• Leftover references: no command, agent, hook or skill in the repo still points at a disabled plugin or at rust-async-patterns.
  • docs/superpowers/plans/… still mentions superpowers: sub-skills. It is a historical plan doc that this PR does not touch.
  • The skill-creator mention in CLAUDE.md still works through anthropic-skills:skill-creator.
• Broken links after the skill deletion: bootstrap.sh links the whole skills/ directory (symlink_claude_dir "skills"), not each skill. So deleting the skill folder leaves no broken symlink in ~/.claude/skills/.
• Docs table: rfc-* agents no longer exist, so the audit-*, summarize-work change in docs/work-review-cycle.md is correct.
• shell-scripting: the new example local result; result="$(command)" is the form shellcheck accepts. It removes the conflict with the skill's own "must pass shellcheck" rule.
• summarize: the narrower description (it no longer fires on "what is this" or "what happened in this PR") is a deliberate choice explained in the PR body. The "Without the CLI" section still covers summarizing PRs and code by hand.

Verdict
APPROVE - No blocking findings. This is a config and docs cleanup with no behavior regressions.

---

Automated review by Claude Code