agent-add-rule（Agent 添加规则）

Add Rule — Place Agent Instructions Correctly

Add a new rule or convention to the right location in the progressive disclosure structure.

Context Spectrum

Static (root CLAUDE.md)      — loaded every conversation. Token cost paid on every conversation.
Semi-dynamic (docs/agents/)  — linked from root. Token cost paid only when Claude reads it.
Fully dynamic (skills)       — metadata only in context. Token cost paid only when triggered.

Workflow

Step 1: Ask

Ask the user: "What rule or convention do you want to add?"

Accept free text. If the user already provided it (e.g., /agent-add-rule always use snake_case for database columns), skip this step.

Step 2: Analyze Current Structure

Read:

• Root CLAUDE.md
• List files in docs/agents/
• List available skills

Understand what already exists so you don't duplicate or contradict.

Step 3: Classify

Apply this decision tree:

Does the agent consistently get this wrong WITHOUT being told?
├── NO → Skip. The rule costs tokens and is not needed.
│
├── YES → Does it apply to EVERY task?
│   ├── YES → Root CLAUDE.md (static)
│   │         Examples: package manager, multi-tenancy, project scripts
│   │
│   └── NO → docs/agents/ file (semi-dynamic)
│             Examples: lint rules, test thresholds, API conventions
│
└── Is it a repeatable workflow or procedural knowledge?
    ├── YES → Skill (fully dynamic)
    │         Examples: deployment process, PR review checklist, migration procedure
    │
    └── NO → Skip. The rule is not needed.

Key questions to ask the user:

1. "Does the agent consistently get this wrong?" — If no, consider skipping
2. "Does this apply to every task or just some?" — Static vs semi-dynamic
3. "Is this a rule or a workflow?" — docs/agents/ vs skill
4. "Will this change frequently?" — Skills are easier to evolve independently

Step 4: Recommend

Present the recommended placement with reasoning:

Recommendation: Add to docs/agents/guardrails.md

Reasoning:
- This is a data handling rule, not a universal workflow rule
- It applies only when working with the database
- guardrails.md already covers data isolation patterns
- Adding to root would cost tokens on every conversation unnecessarily

Step 5: Confirm

Ask the user to confirm or override. If they override, respect their choice but note the trade-off:

• Moving to root: "This adds 1-3 lines to every conversation's context"
• Moving to docs/agents/: "This is visible only when explicitly linked and read"
• Moving to skill: "This loads only when the skill is triggered by a matching keyword"

Step 6: Write

Based on confirmed placement:

If root CLAUDE.md:

• Add the rule under the appropriate section (Key Rules, Workflow, etc.)
• This adds 1-2 lines to every conversation's context
• Keep it concise — 1-2 lines max

If existing docs/agents/ file:

• Read the target file
• Add the rule under the appropriate section
• Keep consistent formatting with existing content

If new docs/agents/ file:

• Create the file with a clear heading and the rule
• Update root CLAUDE.md links section with a new entry including routing signal
• Example: - API Conventions (docs/agents/api-conventions.md) — REST patterns, error response format, pagination

If skill:

• Tell the user to run /agent-skill-creator to scaffold it
• Provide the rule content as input for the skill body

Examples

Example 1: Universal Rule → Root

User: "Always use pnpm, never npm"

Classification: Agent gets this wrong without being told + applies to every task → Root

Action: Add to Key Rules section in CLAUDE.md

Example 2: Topic-Specific Rule → docs/agents/

User: "API responses must always include a requestId field"

Classification: Agent sometimes misses this + only applies to API work → Semi-dynamic

Action: Add to docs/agents/guardrails.md or create docs/agents/api-conventions.md

Example 3: Complex Workflow → Skill

User: "When deploying, always run migrations first, then build, then deploy to staging, verify, then production"

Classification: Repeatable multi-step procedure → Fully dynamic (skill)

Action: Suggest /agent-skill-creator to create a deployment skill

Example 4: Unnecessary Rule → Challenge

User: "Always use const instead of let"

Classification: ESLint already enforces this → Not needed

Response: "ESLint already enforces this via the prefer-const rule. Adding it to agent instructions would cost tokens without benefit. Skip?"

Positive Trigger

User: "Add a new convention that API responses must include a request ID and put it in the right agent config location."

Expected behavior: Use agent-add-rule guidance to classify placement, confirm with the user, and apply the rule in the appropriate location.

Non-Trigger

User: "Implement a feature flag system for staged rollouts in our backend service."

Expected behavior: Do not prioritize agent-add-rule; use an implementation-focused skill/workflow instead.

Principles

• Validate necessity: Every rule costs tokens. Only add rules the agent consistently gets wrong without being told.
• No duplication: If ESLint, TypeScript, or another tool already enforces it, don't add it
• Routing signals matter: When adding to docs/agents/, update the root CLAUDE.md link description so Claude knows when to follow it
• One level deep: Never cross-reference between docs/agents/ files. All links go from root

Troubleshooting

Skill Does Not Trigger

• Error: The skill is not selected when expected.
• Cause: Request wording does not clearly match the description trigger conditions.
• Solution: Rephrase with explicit domain/task keywords from the description and retry.

Guidance Conflicts With Another Skill

• Error: Instructions from multiple skills conflict in one task.
• Cause: Overlapping scope across loaded skills.
• Solution: State which skill is authoritative for the current step and apply that workflow first.

Output Is Too Generic

• Error: Result lacks concrete, actionable detail.
• Cause: Task input omitted context, constraints, or target format.
• Solution: Add specific constraints (environment, scope, format, success criteria) and rerun.