SuperAPI AI Directory

simplify

Coding & development

View repository

Updated 10-16 02:04

About this skill

Simplify

Overview

Use this as a pre-review cleanup gate. The goal is not formal approval; it is to inspect changed code for reuse opportunities, code quality issues, and efficiency issues, then fix the worthwhile findings without changing behavior.

When to Use

  • The user wants a /simplify-style cleanup pass before reviewer or code-reviewer
  • A feature is complete and the diff likely contains duplication, rough edges, or avoidable inefficiency
  • You want to reduce avoidable reviewer comments before formal review
  • The change spans enough files or logic that one simplification pass is not enough

Do not use this skill for:

  • Tiny diffs where a single local cleanup is enough
  • Behavior changes, feature work, or bug fixes that alter requirements
  • Large architectural redesigns across untouched modules
  • Formal spec validation or security review

Place in Review Flow

Recommended placement:

implement -> simplify -> code-reviewer and/or reviewer -> receiving-code-review

Skip the full three-agent pass when:

  • the change is tiny and obvious
  • only one or two trivial lines changed
  • the user wants immediate formal review instead of cleanup

Workflow

Phase 1: Identify Changes

  • If the user gives a base/head range, use that
  • If staged changes exist, inspect git diff HEAD so both staged and unstaged changes are included
  • Otherwise use git diff
  • If there is no git diff, review the most recently modified files the user mentioned or that you edited earlier in the session
  • Keep a compact summary of touched files plus the full diff text for the reviewers

Phase 2: Read Local Standards

  • Check local instructions first such as AGENTS.md, repo conventions, and any relevant skills
  • Follow existing naming, formatting, and structure in the touched files
  • If local conventions conflict with generic cleanup preferences, local conventions win

Phase 3: Launch Three Review Agents in Parallel

For medium or large diffs, spawn three read-only subagents in parallel in a single round. Use explorer agents when possible because this is codebase analysis, not implementation.

Each agent must receive:

  • the full diff
  • the touched file list
  • the cleanup constraint: no behavior changes, no edits, findings only

The three review domains are:

  • Code reuse review
  • Code quality review
  • Efficiency review

Use the prompt templates in:

Important:

  • Review agents are read-only; they do not edit files
  • The main agent aggregates findings and performs the fixes
  • Do not spawn multiple worker agents to edit the same touched files in parallel

Phase 4: Fix the Worthwhile Issues

  • Wait for all reviewers to finish
  • Aggregate findings and remove duplicates
  • Fix issues directly if they improve the code without changing behavior
  • If a finding is a false positive or not worth the churn, note that and skip it without arguing at length

Phase 5: Verify

  • Run the narrowest meaningful verification for the changed area
  • Prefer targeted tests first, then lint/build if appropriate
  • If no automated verification exists, state that explicitly and mention residual risk
  • Briefly summarize what was fixed, or state that the code was already clean enough

Quick Reference

Goal Prefer Avoid
Readability Clear names, explicit branches, small coherent helpers Dense one-liners, clever tricks
Scope control Touched files, recent diff, user-specified paths Repo-wide cleanup without request
Safety Targeted verification after edits Assuming refactor is safe
Comments Intent and constraints Narrating obvious code
Agent usage Three parallel read-only reviewers, one controller applying fixes Multiple editing agents touching the same files

Common Mistakes

  • Turning simplification into redesign
  • Extracting abstractions too early
  • Renaming broadly across unrelated files
  • Optimizing for fewer lines instead of easier maintenance
  • Skipping verification because the change "looks mechanical"
  • Letting all three agents edit code instead of only reporting findings
  • Using this skill as a substitute for formal review or spec validation

Upstream Reference

This Codex skill adapts the /simplify flow used in Claude-based workflows into a Codex-native skill. See references/upstream-agent.md for the source notes and adaptation details.