under200

name

under200

description

Enforce NASA-style "Power of Ten" software engineering discipline on any code being written or modified, with a hard guideline that no source file exceeds 200 lines so a reviewer can read it top-to-bottom in one editor screen without jumping around. Use this skill whenever writing new source files, refactoring existing ones, reviewing code, splitting modules, or setting up a new project — even if the user does not explicitly say "under 200" or "NASA". If the user mentions file length, file size, "this file is too long", complexity, readability, code-review friendliness, or asks to set up linters/formatters/CI, apply this skill. Also apply proactively whenever any source file being edited is approaching or exceeds 200 lines.

under200

A discipline skill that adapts NASA's Power of Ten rules (Gerard J. Holzmann, JPL) into general-purpose software engineering practice. The central, non-negotiable rule is the 200-line file ceiling. Every other rule exists to make that ceiling achievable without smuggling complexity elsewhere.

Why 200 lines

A reviewer should be able to open a file, scroll once, and have read the whole thing. Files longer than that force the reader into the jump-around-and-hold-state-in-your-head mode that hides bugs, defeats code review, and resists onboarding. 200 lines is roughly two editor screens at a comfortable font size — enough room for a real unit of code, short enough that nothing is hiding.

This is a ceiling, not a target. Most files should be much smaller. If a file is at 195 lines, it is already overdue for a split.

The rules

Apply these in priority order. The earlier rules win when they conflict with later ones.

1. Files stay under 200 lines (hard ceiling)

• Count all lines including imports, blank lines, and comments. The reviewer has to scroll past them too.
• Generated files, lockfiles, fixtures, and migrations are exempt — they are not read top-to-bottom.
• When a file crosses 200 lines, split it before adding the feature that pushed it over, not after. Splitting under deadline pressure produces worse seams than splitting deliberately.
• Prefer splits along meaning (one concept per file), not along line count (don't just cut at line 200).

2. Functions stay under ~60 lines

One printed page. If a function does not fit on a screen, the reader cannot hold its control flow in working memory. Extract helpers — but extract along meaningful boundaries, not arbitrarily.

3. Control flow stays simple

• No goto. No setjmp/longjmp analogues (uncontrolled exception-as-control-flow).
• Recursion only when bounded by a known finite input (parsing a fixed-depth AST is fine; recursing over user input is not).
• Every loop must have a provable upper bound on iterations. while (true) is acceptable only with a clear exit invariant documented in code.

4. Validate at boundaries, trust internals

• Check the return value of every fallible call. If a return value is intentionally ignored, make that explicit (_ = … in Go, void cast in C, # noqa with rationale in Python).
• Validate inputs at system boundaries (HTTP handlers, CLI entry points, deserialization). Do not re-validate inside internal helpers — that is defensive bloat and inflates line counts.

5. Smallest possible scope

• Declare variables at the narrowest scope where they are used.
• Prefer module-private over package-public; package-public over global.
• No mutable globals unless the runtime forces them (and then document why).

6. Limit metaprogramming and indirection

• Preprocessor macros (C/C++), decorators that rewrite ASTs, runtime monkey-patching, and deeply nested generics all hide control flow from the reader. Use them only when the alternative is materially worse.
• Function pointers, callbacks, and dynamic dispatch are fine — but each layer of indirection costs the reader something. Spend the budget where it pays off.

7. Compile/lint clean

• All warnings enabled. Treat warnings as errors in CI.
• Static analyzers and linters configured and passing. If a rule fires, fix the code; do not blanket-disable the rule.

When this skill triggers

Apply it whenever you are:

• Writing a new source file → start with the structure that will keep it under 200 lines.
• Editing an existing file → check the line count first. If it is over 150, plan the split before adding code.
• Reviewing code → flag any file over 200 lines and any function over 60.
• Setting up a new project, CI, or pre-commit config → configure the linter rules in the next section.
• Asked to "clean up" / "refactor" / "make this more readable" → the line ceiling is usually the fastest win.

Linter enforcement

If the project has a linter, configure it to enforce these rules. A rule that lives only in a skill rots; a rule that fails CI does not. When adding a new linter rule to an existing codebase that violates it, use the linter's per-file ignore mechanism for the existing offenders rather than relaxing the global threshold — that way new code is held to the standard and old code surfaces as known debt.

Concrete rule mappings (use these exact rules when the corresponding linter is present):

• ESLint / TypeScript: max-lines: ["error", { max: 200, skipBlankLines: false, skipComments: false }], max-lines-per-function: ["error", 60], complexity: ["error", 10], max-depth: ["error", 4].
• Biome: noExcessiveLinesPerFunction, noExcessiveCognitiveComplexity. File-length cap is not built in — add a pre-commit check.
• Ruff / Python: select PLR0915 (too-many-statements), C901 (complexity), PLR0912 (too-many-branches). For file length use a pre-commit hook (check-added-large-files is for bytes; file-line cap needs a custom hook).
• Pylint: max-module-lines=200, max-line-length=…, max-args, max-locals, max-branches.
• golangci-lint: enable funlen (lines: 60), gocyclo (min-complexity: 10), lll. File length: enable filelength linter or add a CI check.
• Clippy / Rust: clippy::too_many_lines (configure in clippy.toml with too-many-lines-threshold = 60), clippy::cognitive_complexity.
• RuboCop: Metrics/ClassLength: Max: 200, Metrics/MethodLength: Max: 60, Metrics/CyclomaticComplexity, Metrics/AbcSize.
• SwiftLint: file_length: 200, function_body_length: 60, cyclomatic_complexity.
• ktlint / detekt: detekt LongMethod (threshold 60), LargeClass (threshold 200), ComplexMethod.

If the project has no linter, add a minimal pre-commit hook that fails when any tracked source file exceeds 200 lines. A POSIX one-liner is enough:

git diff --cached --name-only --diff-filter=ACM \
  | grep -E '\.(ts|tsx|js|jsx|py|go|rs|rb|swift|kt|java|c|h|cpp)$' \
  | xargs -I{} sh -c 'lines=$(wc -l < "{}"); [ "$lines" -le 200 ] || { echo "{}: $lines lines (>200)"; exit 1; }'

Wire it into Husky, pre-commit, lefthook, or whatever the project already uses. Do not invent a new hook framework.

How to apply during a task

1. Before editing a file, run wc -l <file>. If it is already ≥ 150 lines, plan how to split before adding code.
2. After editing, re-check. If you crossed 200, split now — do not defer.
3. When splitting, cut along the seam where coupling is weakest (one concept per file). If no such seam exists, the function/class is doing too much and needs decomposition first.
4. If the user pushes back on a split ("just add it here, it's fine"), surface the cost ("this file will be 240 lines after the change, which makes review harder") and let them override. The rule is a strong default, not a religion — but the override should be conscious.
5. When configuring a new project, add the linter rules from the table above in the same PR as the project scaffolding. Retrofitting limits later is painful.

What this skill is not

• Not a style guide (indentation, naming, import order). Use the project's existing formatter for that.
• Not a substitute for tests, types, or review.
• Not a license to split files arbitrarily — a 50-line file that is half of a coherent concept is worse than a 220-line file that is whole. Split meaning, not lines.

One-line summary

If a reviewer cannot read your file top-to-bottom in one scroll, the file is wrong — fix the file, and let the linter remember the rule so you don't have to.