writing-rules

Writing Rules

Scope: covers .claude/rules/ file authoring — a Claude-Code-specific concept (path-scoped, always-loaded instruction files with paths: glob frontmatter). Codex CLI and Antigravity have no exact .claude/rules/ equivalent; their always-on project instructions live in the hierarchical memory file (AGENTS.md for Codex, GEMINI.md for Antigravity — both can be pointed at the canonical AGENTS.md; see [[nlpm:conventions-codex]] / [[nlpm:conventions-antigravity]]). The bold-imperative-plus-rationale writing technique here applies to any tool's instruction files. For CLAUDE.md / AGENTS.md conventions, see [[writing-plugins]]. For system prompts generally, see [[writing-prompts]].

1. The Golden Format

Every rule should have three parts:

**Use X, not Y.** Without X, [concrete bad thing happens]. Y causes [specific problem] because [mechanism].

Part	Purpose	Example
Imperative	What to do	Use Result<T, AppError> for all API handler returns.
Consequence	What goes wrong without it	Without it, errors propagate as 500s with no context.
Mechanism	Why it fails	Raw panics bypass the error middleware and crash the worker.

One-Line Rules (when mechanism is obvious)

**Use `const`/`let`, never `var`.** `var` hoists to function scope, causing stale-reference bugs.

Multi-Line Rules (when mechanism needs explanation)

**Use database transactions for multi-table writes.** Without transactions, partial writes leave the database in an inconsistent state. The ORM's `save()` method does not auto-wrap related writes -- you must explicitly call `db.transaction()`.

2. Positive Framing (Pink Elephant Effect)

Claude fixates on prohibited things. Saying "Don't use X" makes Claude think about X.

Before (negative framing -- score 60)

- Don't use var
- Don't mutate function parameters
- Don't use console.log in production code

After (positive framing -- score 90)

- **Use `const` for all bindings; use `let` only when reassignment is required.**
- **Return new objects instead of mutating function parameters.**
- **Use the `logger` service for all logging.** `console.log` is stripped in production builds.

Conversion Pattern

Negative (avoid)	Positive (use instead)
Don't use X	Use Y (where Y is the correct alternative)
Never do X	Always do Y
Avoid X because...	Use Y because... (flip the rationale)
X is deprecated	Use Y, which replaced X in version N

3. Enforceability Test

Before writing a rule, ask: "Can I check compliance in a 30-second code review?" If no, it is not a rule.

Enforceable (specific, testable)

Rule	Test
Use Result<T, AppError> for all API handler returns.	Grep for handler functions, check return types
All API endpoints require @auth decorator.	Grep for route definitions, check for decorator
Database queries use parameterized statements, not string concatenation.	Grep for SQL strings, check for + or template literals

Not Enforceable (subjective, unmeasurable)

Rule	Why it fails
"Write clean, maintainable code"	What is "clean"? No objective test.
"Keep functions small"	How small? 10 lines? 20? 50?
"Use meaningful variable names"	"Meaningful" is subjective.
"Follow best practices"	Which practices? Says nothing specific.

Making Vague Rules Enforceable

Vague	Enforceable version
"Keep functions small"	Functions must be under 40 lines. Reference: enforced by eslint max-lines-per-function
"Use meaningful names"	Variable names must be >= 3 characters except loop indices (i, j, k).
"Handle errors properly"	Every catch block must either re-throw, log + return error response, or call reportError().

4. Budget Discipline

All rules across .claude/rules/ must total under 500 lines. Every line costs tokens on every Claude interaction -- rules are always loaded.

Token Cost

Rule lines	Approx tokens per interaction	Annual cost at 100 interactions/day
100	~400	Negligible
300	~1,200	Noticeable
500	~2,000	Budget line
800+	~3,200+	Over budget -- consolidate

Line Reduction Strategies

Strategy	Example	Lines saved
Defer to linter	"Reference: enforced by pnpm lint" instead of re-stating lint rules	10-30
Merge related rules	Combine 3 files about error handling into 1	15-25
Delete training knowledge	Remove rules Claude follows without being told	5-15
Use tables instead of lists	10 rules as list = 20 lines; as table = 12 lines	5-10

Rules Claude Already Follows (safe to delete)

These are part of Claude's training and do not need rules:

• "Use descriptive variable names" (Claude already does this)
• "Add comments to complex code" (Claude already does this)
• "Handle null/undefined checks" (Claude already does this)
• "Use async/await instead of callbacks" (Claude already prefers this)

Only write rules for things specific to your project that Claude would not know.

5. Path Scoping

Rules without path scoping apply to every file -- expensive and often wrong.

---
paths: ["src/api/**/*.ts"]
---

Scoping Strategy

Rule type	Scope	Example paths
API conventions	API routes only	src/api/**/*.ts, src/routes/**/*.ts
Database rules	Data layer only	src/db/**/*.ts, src/models/**/*.ts
Test conventions	Test files only	**/*.test.ts, **/*.spec.ts
Universal rules	No scope (apply everywhere)	(omit paths field)

Rule: if a rule mentions a specific directory, technology, or layer -- scope it.

Cost Impact of Scoping

Scenario	Token cost
Unscoped: 200-line rules file loaded on every interaction	800 tokens always
Scoped: same rules split into 4 files with path scoping	200 tokens per interaction (only relevant rules load)

6. Conflict Prevention

Two rules must never contradict. If they could, put them in the same file with explicit conditions.

Bad (separate files, contradictory)

rules/api.md:

**Return raw JSON objects from API handlers.**

rules/error-handling.md:

**Wrap all returns in Result<T, AppError>.**

Good (same file, explicit conditions)

rules/api-returns.md:

**Return `Result<T, AppError>` from API handler functions.** This ensures consistent error formatting through the error middleware.

**Return raw JSON from internal service functions.** Services are called by handlers, not directly by clients, so they do not need the Result wrapper.

Conflict Detection Checklist

Before adding a new rule, check:

1. Search all existing rules for the same keywords
2. Does any existing rule say the opposite?
3. Does any existing rule cover a broader case that includes yours?
4. If conflict found: merge into the same file with explicit conditions

7. Worked Example

Before (score 45/100) -- 800 lines, 12 files

.claude/rules/
  naming.md          (80 lines -- mostly restates ESLint rules)
  errors.md          (90 lines -- contradicts exceptions.md)
  exceptions.md      (70 lines -- contradicts errors.md)
  logging.md         (60 lines -- unscoped, only relevant to src/api/)
  testing.md         (85 lines -- includes Jest tutorial content)
  database.md        (95 lines -- unscoped, only relevant to src/db/)
  api.md             (70 lines -- overlaps with errors.md)
  security.md        (55 lines -- restates OWASP basics Claude already knows)
  performance.md     (45 lines -- vague advice like "write fast code")
  imports.md         (30 lines -- restates ESLint import rules)
  comments.md        (25 lines -- Claude already adds good comments)
  types.md           (95 lines -- half is TypeScript tutorial)
Total: 800 lines, 12 files

After (score 92/100) -- 180 lines, 4 files

.claude/rules/
  api.md             (55 lines, scoped to src/api/**)
  database.md        (45 lines, scoped to src/db/**)
  testing.md         (40 lines, scoped to **/*.test.ts)
  universal.md       (40 lines, unscoped -- truly universal rules)
Total: 180 lines, 4 files

What was removed:

• naming.md: deleted (ESLint handles this, Claude defaults are fine)
• errors.md + exceptions.md: merged into api.md with explicit conditions
• logging.md: merged into api.md, scoped to src/api/**
• security.md: deleted (Claude already knows OWASP basics)
• performance.md: deleted (vague, unenforceable)
• imports.md: deleted (ESLint handles this)
• comments.md: deleted (Claude already writes good comments)
• types.md: reduced to 10 lines of project-specific type rules in universal.md

Savings: 800 -> 180 lines = 78% reduction. Token cost per interaction dropped from ~3,200 to ~720.

8. Quality Checklist

Before shipping rules, verify:

• Every rule follows the golden format: imperative + consequence + mechanism
• Positive framing (no "Don't..." as the primary instruction)
• Every rule passes the 30-second enforceability test
• Total across all rule files < 500 lines
• Rules scoped via paths: frontmatter (e.g., paths: ["src/api/**/*.ts"]) when not universally applicable
• No contradictions between rule files
• No rules that Claude follows by default from training
• No rules that a linter already enforces (reference the linter instead)