Claude Code plugins were supposed to make the tool smarter, but on current models a stack of add-ons can add token cost and slow simple tasks down. The fix is a leaner setup: audit what you have with /doctor prompt-audit, keep instructions short, and put hard limits in settings.json instead of in a prompt.
What Claude Code and plugins actually are
Claude Code is Anthropic's coding tool that runs in your terminal, the text-based command line on your computer. Plugins are add-on packages of instructions that give it extra capabilities. Tokens are the units of text the model reads and writes, and you pay by the token, so every instruction a plugin loads is a small recurring cost.
Earlier models needed step-by-step scaffolding to stay on track. Newer ones, such as Opus 5.5, reason well on their own, so redundant instructions can get in the way instead of helping.
Why more plugins can mean a bigger bill
The caution is strongest for big instruction packs. In some tests the popular Superpowers plugin made simple tasks cost more and run slower than plain Claude Code. Results vary by task: one 12-session test reported lower cost on some work, and a build-off had plain Claude Code finish fastest. Treat it as "overhead you should measure", not a verdict.
The Caveman plugin, built to make answers shorter, shows the same pattern. Its author's own eval found about a 2.9% median gap against typing "Answer concisely" into a normal prompt, which the author described as inside the noise. Monolithic packs such as ECC load hundreds of skills and agents at startup, which one creator's breakdown put at roughly 29,000 to 53,000 tokens before any task begins.
Creators are saying the same thing. Nate Herk's "new rules of Claude Code" video drew about 183K views, and Chase AI's "STOP Using These Claude Code Plugins" about 39K. On X, @WYZBen87 put the habit simply: delete the rules a new model no longer needs, because old prompts cost tokens and change nothing.
Keep rules short and specific
Anthropic's guidance on writing rules is plain: be specific ("use 2-space indent" works, "format nicely" does not), keep the file short, and skip the shouting. Capital letters and words like "CRITICAL" can make Claude over-apply a rule. Rules in prompts are guidance. Settings are enforcement.
Where auto mode helps and where you still decide
Auto mode acts as a default filter on risky commands. In an Anthropic study of 1,053 testers, it blocked one deliberately planted dangerous command 89% of the time, compared with 13.6% for human testers. That is one command in one study, so it is not a guarantee. A good rule of thumb is reversible versus irreversible. If Claude writes bad code, you can undo it. If it reads your .env file or deletes uncommitted work, you cannot.
Two things to do today
Run
/doctor prompt-auditin Claude Code (version 2.1.283 or later). It checks your CLAUDE.md, AGENTS.md, skills, agents and commands for stale facts, contradictions and wording written for older models. Nothing changes until you apply the edits. It audits instruction files and skills rather than installed plugins directly, so review your plugin list yourself too.Add a deny rule for
.envunderpermissionsin your project's.claude/settings.json. It is a hard block that applies before auto mode even evaluates the risk.
FAQ
Are Claude Code plugins bad?
No. Some are useful. The point is that large instruction packs can add cost and slow simple tasks on current models, so install fewer and measure the ones you keep.
What does /doctor prompt-audit do?
It scans your instruction files, skills, agents and commands for outdated or contradictory guidance and proposes edits. Nothing is changed until you apply them.
How do I stop Claude Code from reading my .env file?
Add a deny rule for it under permissions in .claude/settings.json. Settings-level rules are enforced, unlike instructions in a prompt or CLAUDE.md.
Does auto mode catch every dangerous command?
No. Anthropic's study found it blocked a single planted dangerous command 89% of the time, so keep human review for irreversible actions.
Companion episode:
Want more practical AI breakdowns for solo builders and small teams? Get them free at joebuildsai.com.

