code-simplification

Code Simplification Skill

When to trigger

Apply when:

• Features work and tests pass — simplification is a separate phase
• Code review flagged readability (deep nesting, long functions, unclear names)
• You're touching the area for an unrelated reason and the existing code obstructs understanding
• A new contributor (human or agent) was confused by the same code twice

Skip when:

• The code is already clean or follows the canonical pattern
• You don't fully understand what the code does (Chesterton's Fence — see below)
• Performance is critical and the "simplification" would change hot-path semantics
• Tests don't exist yet — write them first (test-driven-development skill)

The five principles

1. Preserve behavior exactly. Ask: "Does this produce the same output for every input — including the inputs the original author was paranoid about?"
2. Follow project conventions. Match the existing codebase, not your external preferences. The .rules/ directory and CLAUDE.md describe what's canonical here.
3. Prefer clarity over cleverness. Explicit beats compact when the reader has to "decompile" the line in their head.
4. Maintain balance. Don't over-inline (one giant function), don't over-extract (10 micro-helpers each used once).
5. Scope to what changed. Drive-by refactors in unrelated files inflate diffs and hide regressions. Open a follow-up PR.

Chesterton's Fence

"Don't remove a fence in the middle of a field until you understand why it was put there."

Before deleting code, comment, or guard clause:

1. Search for the issue or commit that added it (git log -S "term" --all, git blame).
2. Understand the original intent. A weird-looking check often guards against a real production bug from years ago.
3. If the original reason no longer applies, document the reasoning in the simplification commit. If you can't determine the reason, leave it alone — the cost of the fence is one read, the cost of a regression is much higher.

The process

1. Understand before touching — apply Chesterton's Fence above.
2. Identify opportunities
   • Deep nesting (>3 levels) — extract guard clauses, early returns, helper functions
   • Long functions (>50 lines per .rules/typescript.md) — split by concern
   • Generic names (data, tmp, result) — rename to describe role in this scope
   • Duplicated logic in 3+ places — extract (2 instances is coincidence, per CLAUDE.md DRY rule)
   • Redundant abstractions (one-liner wrappers, single-implementation interfaces)
3. Apply changes incrementally. One simplification → run tests → commit. Never batch multiple simplifications without tests between.
4. Keep refactoring separate from feature work. Different commits, ideally different PRs.
5. Verify the result is genuinely easier to understand. Read it cold. If you still need to think, the simplification didn't land.

Anti-rationalization — Simplification

Excuse	Counter
"It's working — leave it"	Working ≠ maintainable. Today's working-but-confusing code is tomorrow's bug-fix nightmare.
"Fewer lines = simpler"	Line count is a proxy, not a goal. A 5-line clever expression is often less readable than 12 obvious lines.
"I'll refactor while I add features"	This is how regressions arrive. Mixing refactor + feature changes makes the diff impossible to review and the regression impossible to bisect.
"Types are self-documenting"	Types describe shape, not intent. function process(x: User): Result says nothing about what process does.
"Just one tiny drive-by cleanup"	Drive-by edits in unrelated files inflate the diff and hide regressions in unrelated systems. Open a follow-up PR.
"I don't need to understand it to refactor it"	Yes you do. See Chesterton's Fence.

Red flags

• Tests had to be modified to pass — you changed behavior, not just shape
• "Simplified" code is longer or harder to follow than the original
• You removed error handling because it "looked redundant"
• You're refactoring code you can't fully explain
• Multiple simplifications batched into one commit without tests between them
• The diff touches files unrelated to the original task

Verification checklist

• All tests pass unmodified: pnpm lint && pnpm typecheck && pnpm test
• No errors swallowed, no try/catch removed unless replaced with explicit non-erroring path
• Each simplification is a separate commit with the test run in between
• Diff scope matches the original task — no unrelated drive-bys
• Names follow project conventions; no data/tmp/result introduced
• Function size, file size, and complexity ESLint gates still pass (max-lines-per-function: 50, max-lines: 400, complexity: 10)
• Read the result cold. If still confusing, the simplification didn't land — revert and try again.