Agent Skills: Crafting Rules

Crafting Rules

Overview

Rules are markdown files with optional YAML frontmatter, injected into the system prompt to guide agent behavior. Scope them with filters or leave unconditional for global standards.

Field Reference

| Field | Type | Category | Purpose | | ---------- | ------------------ | ---------- | ---------------------------------------------------------- | | globs | string[] | Legacy | Apply when any file in context matches a pattern | | keywords | string[] | Legacy | Apply when the user's latest prompt matches a keyword | | tools | string[] | Legacy | Apply when any listed tool ID is available | | model | string[] | Runtime | Match against the current LLM model ID | | agent | string[] | Runtime | Match against the current agent type (e.g., programmer) | | command | string[] | Runtime | Match against the current slash command (e.g., /plan) | | project | string[] | Runtime | Match against detected project tags (e.g., node, rust) | | branch | string[] | Runtime | Match against git branch name (supports glob patterns) | | os | string[] | Runtime | Match against OS (linux, darwin, win32) | | ci | boolean | Runtime | Match against CI environment (true = in CI) | | match | 'any' | 'all' | Combinator | any (default): OR logic. all: AND logic. |

• All fields are optional; no frontmatter means the rule always applies.
• With match: any (default), the rule applies if ANY declared condition matches.
• With match: all, the rule applies only if ALL declared conditions match.
• When a runtime value is unavailable (e.g., no git repo), that dimension is a non-match.

Rule Format

---
globs:
  - '**/*.ts'
keywords:
  - 'vitest'
model:
  - claude-sonnet-4
agent:
  - programmer
branch:
  - feature/*
match: any
---

# Rule Title

- Write rules as concrete, actionable instructions.

Matching Strategy

• Use globs when the rule is about code in specific files/directories.
• Use keywords when the rule is about a topic that may not include files.
• Use tools when the rule depends on specific MCP tools being available.
• Use runtime filters (model, agent, command, project, branch, os, ci) to scope rules to specific environments or workflows.
• Use match: all when you need every declared condition to be true (AND logic).
• Use match: any (or omit match) when any single condition should trigger (OR logic).
• Use no filters for global standards (tone, structure, safety, commit conventions).

Important constraints:

• Keyword matching is case-insensitive word-boundary prefix matching (e.g., test matches tests and testing).
• Branch patterns support globs via minimatch (e.g., feature/*, release/**).
• Missing runtime context (e.g., no git repo for branch) counts as a non-match for that dimension.

Keyword Selection

Keywords use case-insensitive word-boundary prefix matching — short or generic words over-match.

Denylist: generic nouns (code, file, project, repo, bug, issue, change), common verbs (add, update, remove, fix, make, create, implement), over-broad topics (testing, performance, security, deployment, database, api), single-token abbreviations (ci, cd, db, ui, ux).

Allowlist: tool/framework names (vitest, jest, pytest, playwright, cypress, eslint, prettier, typescript, terraform, kubernetes), compound phrases (unit test, integration test, snapshot test, lint rule, error boundary, api endpoint, rest api), high-intent verbs (refactor, rollback, migrate, deprecate).

Audit checklist:

• Would this keyword appear in prompts where the rule should NOT apply?
• Is it likely to appear as part of another word due to prefix matching?
• Can you scope via globs instead?
• Prefer globs (file-scoped) over denylisted keywords.
• Replace generic keywords with compound phrases or tool names that capture intent.

Extracting Rules from Patterns

Signals a pattern should become a rule:

• Explicit: "always do X", "never do Y", "remember to...", "from now on..."
• Corrections: the user fixes the same agent behavior more than once
• Preferences: consistent style/process guidance (tests, commits, PRs, error handling)
• Frustration indicators: "I told you before", "again"

Analysis questions:

• Is this recurring, or a one-off for the current task?
• Does this apply to specific files, or all work?
• Is there an existing rule/config that already encodes this (AGENTS.md, lint config, Prettier, etc.)?
• Would this conflict with project conventions?

Workflow:

1. Identify the behavioral delta (what should change?).
2. Determine scope (globs, keywords, or always-on).
3. Check existing rules/configs for overlap/conflict.
4. Draft the minimal rule (one concept per rule when practical).
5. Choose location (global vs project).

Conversation extraction examples:

• User: "Use early returns instead of nested if/else" -> always-on code style rule.
• User: "In unit tests, always use describe/it blocks" -> prefer glob-scoped rule (e.g., **/*.{test,spec}.*, **/__tests__/**); if prompt-scoped, use allowlisted keywords like unit test, vitest, jest (avoid test/testing).
• User repeatedly fixes import ordering -> glob-scoped rule for the relevant languages/files.

Storage Location

• ~/.config/opencode/rules/: personal preferences you want across projects.
• .opencode/rules/: project/team conventions and repo-specific behavior.

Writing Guidelines

• Use imperative voice: "Do X", "Prefer Y", "Avoid Z".
• Make rules executable: instructions the agent can follow.
• Stay minimal: avoid restating generic best practices.
• Prefer examples over prose when a pattern is subtle.

Examples

Glob-based: TypeScript conventions

---
globs:
  - '**/*.ts'
  - '**/*.tsx'
---

# TypeScript

- Prefer `type` over `interface` unless you need declaration merging.
- Avoid `any`; use `unknown` and narrow.

Keyword-based: unit test guidance (allowlisted terms)

---
keywords:
  - 'unit test'
  - 'integration test'
  - 'vitest'
  - 'jest'
---

# Unit Tests

- Follow Arrange-Act-Assert.
- Name tests: `it('should <expected> when <condition>')`.

Unconditional: always-on standards

# Code Style

- Prefer early returns over deep nesting.
- Extract magic numbers to named constants.

Runtime filters with match: all: feature branch development

---
agent:
  - programmer
branch:
  - feature/*
os:
  - linux
  - darwin
ci: false
match: all
---

# Feature Branch Dev

- Create atomic commits with clear messages.
- Run tests before pushing.

Common Mistakes

• Using denylisted keywords: test fires on nearly every prompt — use unit test or globs instead.
• Forgetting match: all: Two filters with default OR means EITHER triggers — add match: all for AND logic.
• Overloading a single rule: 6+ dimensions are hard to reason about — split into focused rules.
• Duplicating lint/formatter config: Check Prettier/ESLint before adding a style rule.
• Using ci as a keyword: Prefix-matches circuit, citizen — use ci: true boolean filter instead.