code-simplifier

code-simplifier — Canonical Adoption

This skill is a canonical marketplace adoption of Anthropic's official code-simplifier plugin (https://github.com/anthropics/claude-plugins-official/tree/main/plugins/code-simplifier, Apache-2.0). The agent prompt body is preserved verbatim under references/agent-prompt.md; this SKILL.md is the marketplace wrapper that adapts it for multi-harness consumption and adds the audit rubric below.

Why this skill exists

Modern AI agents tend to ship code that works but accumulates:

• Dead abstractions, redundant try/catch, defensive code that hides intent
• Inconsistent naming, mixed arrow/function, missing return types
• Near-duplicate logic across services (e.g. parallel fetch/parse/validate paths)
• Silent security gaps (unvalidated inputs, leaked API keys, missing CSP, weak error boundaries) — the code passes typecheck but leaks in production
• Tight coupling that prevents extension without rewrites

code-simplifier exists to surface these issues with a functionality-preserving audit, and to refactor only what the user approves. It never changes behaviour.

Agent prompt (verbatim)

Load references/agent-prompt.md — the canonical Anthropic-authored system prompt for the simplifier. It defines the four non-negotiable guarantees: preserve functionality, apply project standards, enhance clarity, maintain balance, focus scope.

Audit rubric (marketplace extension)

When invoked with an explicit "audit" or "review" trigger (rather than auto-refine-after-write), the agent applies this five-axis rubric in addition to the four guarantees above:

1. Maintainability

• Single-responsibility per module; cohesive function bodies
• Naming: verbs for actions, nouns for data; consistent across the surface
• No copy-pasted logic — extract shared helpers
• No as any / @ts-ignore without a // reason: comment
• Public API surface minimal and stable

2. Security

• API keys never read from localStorage/sessionStorage and never logged
• User input sanitised before use in prompts, URLs, or DOM sinks
• dangerouslySetInnerHTML only with explicit, sanitised content
• WebSocket origins validated; CORS narrowed to known producers
• Generated URLs checked against an allow-list where they could exfiltrate
• Error messages scrubbed of secrets and stack-trace data

3. Extensibility

• Plugin/strategy seams instead of switch (provider) ladders
• Config via data (objects/arrays), not hard-coded branches
• Event-driven state changes (typed events) instead of implicit coupling
• Hook points explicit and documented
• Versioned contracts between services

4. Readability

• No nested ternaries; prefer if/else or switch for ≥3 branches
• Explicit return types on top-level functions
• Comments explain why, not what; obvious code is uncommented
• Magic numbers named; durations in named constants

5. Performance & correctness

• Async boundaries don't block UI; large work chunked
• Listeners and observers torn down on unmount
• Memos justified by measurement, not intuition
• Errors thrown early, not swallowed in catch {}

Workflow

1. Identify scope. Default = recently modified files in the current session. With explicit "full audit" trigger, walk the project's src/ tree.
2. Read project conventions before judging style:
   • AGENTS.md, CLAUDE.md, package.json (lint/format/prettier config)
   • Existing test files for tone
3. Produce a findings report (no edits yet) grouped by the five axes above. Each finding: file:line — axis — finding — proposed change (preserves behaviour).
4. Wait for approval before applying any refactor unless the agent is configured for autonomous mode (see references/agent-prompt.md).
5. Run lint + typecheck + tests after every refactor batch. No exceptions.
6. Document only significant changes — drop comments on mechanical renames.

Files

• references/agent-prompt.md — the Anthropic-authored agent prompt (verbatim).
• checklists/audit-checklist.md — per-axis checklist for self-review.
• MANIFEST.yaml — marketplace manifest (single-skill plugin, isolated).

License

Apache-2.0. Original work © Anthropic. Marketplace adoption © Claudius / iurFriend. See LICENSE.