Imported from jankneumann/agentic-coding-tools (skills/simplify/SKILL.md). Install upstream with npx skills add jankneumann/agentic-coding-tools --skill simplify. Copyright stays with the author.

Simplify

Inspect a focused diff, file, or module for behavior-preserving simplifications: dead code, deep nesting, long functions, premature abstractions, generic names, and isomorphic DRY extracts. The goal is fewer moving parts and faster comprehension — not stylistic preference and not fewer lines for their own sake.

This skill is read → pin → edit: it reviews first, applies a coverage gate, writes characterization tests when needed, gates each candidate against Chesterton's Fence, applies changes one pattern at a time, and dual-runs the suite. Large surface areas are deferred (Rule of 500).

Primary invoke: /simplify
Invocation mode: manual only — operators (or explicit human request) run this skill. Autopilot and implement-feature do not auto-run simplify by default.

When to Use

• After a feature is green and the implementation feels heavier than needed
• During review when readability / complexity is flagged without a behavior bug
• When tech-debt reports local Long Method, Deep Nesting, or local Duplication
• When consolidating duplicated logic that should share one helper
• Optional polish after /implement-feature or /iterate-on-implementation (separate refactor commits)

When NOT to Use

Scope

• Run on the current diff, a specified file/module, or a tech-debt finding ID.
• Production edits only after the coverage gate passes.
• Characterization commits may add tests; simplify commits must not change assertion bodies.
• Single-PR / small-batch changes. Cross-cutting refactors follow Rule of 500 or escalate.

Principles

1. Preserve behavior exactly

Same inputs → same outputs, errors, side effects, and ordering. If unsure, do not change it.

2. Tests are the isomorphism proof

Observable equivalence is proven by behavioral tests, not by agent confidence. Prefer state-based tests (inputs/outputs) over interaction mocks so structure can move without rewriting tests — see test-driven-development (Beyoncé Rule, DAMP tests).

3. Follow project conventions

Match neighboring code: imports, naming, error handling, typing depth. External “clever” idioms that fight the codebase are churn, not simplification.

4. Clarity over cleverness and line count

A short nested ternary is not simpler than an explicit branch. Over-inlining that erases a useful name is a failure mode.

5. Scope to what you intended

No drive-by refactors outside the named surface unless the operator broadens scope.

Chesterton's Fence — Pre-Simplification Check

Before removing or refactoring any non-trivial piece of code, answer all three. If any answer is "I don't know," stop and investigate.

1. Why does this exist? git blame, introducing commit message, callers (grep), tests that pin it.
2. What problem does it still solve? Rate limits, retries, ordering, error masking, security boundaries — load-bearing fences stay.
3. What non-obvious invariants does it preserve? Idempotency, transactional boundaries, timezone normalization, injection defense.

If (2) is "nothing — reason is gone," the fence may come down. Otherwise leave it (or document why in a # CHESTERTON: kept because … comment).

Coverage Gate (required)

Surface under edit
       │
       ▼
 Existing state-based tests pin inputs/outputs/errors/side effects?
       │
  yes  │  no
       │   └──► CHARACTERIZE first:
       │         • Write tests that pass on CURRENT code (green-on-baseline)
       │         • Prefer real impl / fakes over interaction mocks
       │         • Commit: test(<scope>): pin behavior for <surface>
       │         • Only then proceed to production edits
       ▼
 Continue to candidate list / Chesterton / edits

If you cannot pin the surface, you cannot simplify it. Hope is not a dual-run.

Characterization tests are not a license to change behavior later in the same PR — they freeze today's behavior so refactors cannot silently drift.

Rule of 500

Simplifications that touch more than 500 lines OR more than 5 files SHALL NOT be done by hand.

When exceeded:

• (a) Automate — codemod / AST tool (libcst, ts-morph, jscodeshift) with reviewable automation.
• (b) Split — one module / one pattern / one PR; repeat.
• (c) Escalate — /plan-feature for design + review gates.

Mechanical check (recommended):

python3 "<skill-base-dir>/scripts/check_scope.py" --base <baseline-sha>
# oversized? re-run with --allow-codemod only when a real codemod produced the diff

Pattern catalog

Local clarity (existing)

Isomorphic structure (added)

Rebalance note: Inlining premature abstractions is still valid for single-use abstractions that are not extension points. Extracting real duplication is the dual — do not “inline” away a helper that names a real domain concept used in multiple places.

Workflow

0. Scope

Identify target: git diff, path, module, or tech-debt finding ID. Record the baseline SHA (tip before any simplify production edit; after characterization commits if those land first).

1. Understand (Chesterton's Fence)

Blame, callers, existing tests, edge cases. Read project conventions (AGENTS.md / CLAUDE.md / neighboring modules).

2. Coverage gate

Pin or characterize (see above). Run characterization tests and confirm green on baseline.

3. Candidate list

List opportunities by pattern. Drop any that fail Chesterton's Fence; note fences kept.

4. Rule of 500

Group remaining work. Automate, split, or escalate if over budget.

5. Apply incrementally

For each remaining candidate:

1. Make one simplification.
2. Run the targeted suite (then broader suite if targeted is green).
3. If red → revert that simplification; re-evaluate.
4. Commit: refactor(<scope>): <pattern> — <brief> (e.g. refactor(parser): extract guard clauses from validate_input).

Never mix feat / fix with simplify polish in the same commit.

6. Dual-run verify

# Recommended mechanical dual-run (writes simplify-report.json by default).
# Prefer a project-local interpreter so detached worktrees resolve tools;
# the script also symlinks .venv / node_modules from the main repo when present.
python3 "<skill-base-dir>/scripts/verify_behavior_preservation.py" \
  --baseline <baseline-sha> \
  --test-cmd "python3 -m pytest -q"   # or: .venv/bin/python -m pytest / npm test

# Assertion contract on the simplify range (should be clean for expectation bodies).
# --base MUST be the tip AFTER characterization commits.
python3 "<skill-base-dir>/scripts/check_test_contract.py" --base <baseline-sha>
python3 "<skill-base-dir>/scripts/check_scope.py" --base <baseline-sha>

Source-contribution-only example (this monorepo, not portable to consumers): skills/.venv/bin/python -m pytest -q skills/tests/simplify/

Manual equivalent: run the same suite on <baseline-sha> and on HEAD; both must pass.

7. Report

Summarize: patterns applied, fences kept, characterization tests added, dual-run evidence (commands + exit codes or report path), Rule of 500 status.

Script helpers

Scripts live in <skill-base-dir>/scripts/ (installed copy under .claude/skills/simplify/scripts/ or .agents/skills/simplify/scripts/). They use only the standard library plus git.

check_test_contract.py expects --base at the tip after characterization commits. Within that range, any +/- assertion line (including deleted test files) is a contract break.

verify_behavior_preservation.py takes a trusted --test-cmd shell string (e.g. pytest -q). Both SHAs are checked out via temporary detached worktrees so a dirty working tree cannot skew results.

Language sketches (clarity, not prescription)

Python — guard clauses

# Before: nested happy path
def process(data):
    if data is not None:
        if data.is_valid():
            return do_work(data)
        raise ValueError("invalid")
    raise TypeError("missing")

# After
def process(data):
    if data is None:
        raise TypeError("missing")
    if not data.is_valid():
        raise ValueError("invalid")
    return do_work(data)

TypeScript — redundant boolean

// Before
function isValid(input: string): boolean {
  if (input.length > 0 && input.length < 100) return true;
  return false;
}
// After
function isValid(input: string): boolean {
  return input.length > 0 && input.length < 100;
}

Prefer project idioms when they conflict with these sketches.

Handoffs

Common Rationalizations

Red Flags

• Production simplify commits without a coverage-gate decision (existing pins or new characterization tests).
• A simplify PR that changes test assertion bodies to go green.
• Diff over 500 lines or 5 files with no codemod / split plan (Rule of 500 violation).
• Removed code with no blame / caller investigation recorded.
• feat/fix mixed into the same commit as a clarity refactor.
• Autopilot or implement silently running simplify without operator request.
• Inlined helper that deleted a comment documenting a non-obvious invariant (fence lost).
• Isomorphic extract landed without tests covering all rewritten call sites.

Verification

1. Cite each pattern catalog entry applied in the PR/report.
2. For removed/renamed/inlined constructs, cite blame or introducing commit (Chesterton's Fence).
3. Confirm coverage gate: either list existing pinning tests or show the characterization commit (test(...): pin behavior…) that is green on baseline.
4. Confirm dual-run: suite green on baseline SHA and on HEAD (attach simplify-report.json from verify_behavior_preservation.py when used).
5. Confirm assertion contract: check_test_contract.py --base <baseline> exits 0 for the simplify range (characterization commits may add tests; simplify commits must not mutate expectation bodies).
6. Confirm scope: check_scope.py --base <baseline> exits 0, or --allow-codemod with the codemod named in the report.
7. Confirm git diff <baseline>..HEAD --stat (or report) shows intentional surface only — no unrelated drive-by files.