Claude Instruction Router

Claude has at least seven places an instruction can live, and most setups put everything in one of them. Paste your rules and each one is routed to where it actually works: CLAUDE.md, a Skill, a hook, a Project, your personal preferences, a connector, a subagent, or the bin.

Paste the instructions you give Claude (a CLAUDE.md, Project instructions, custom instructions, or a list of rules) one per line. Each one is routed to the place it belongs: CLAUDE.md, a Skill, a hook, a Project, your personal preferences, a connector, a subagent, or the bin. Runs in your browser; nothing is sent anywhere.

Your instructions (one per line)

11 instructions · 7 can leave your always-loaded context

Hook3

It's a rule that must hold every time. Instructions are advice the model can miss; a hook runs as code and can block the action.

Claude Code: .claude/settings.json → hooks (PreToolUse / PostToolUse / Stop)

Never commit directly to main.a hard 'never' around an action
Always run prettier after editing a file.something that must run automatically
Never read or print the contents of .env files.protects secrets or production

CLAUDE.md2

It's a fact about this codebase that applies to most sessions and that the agent can't reliably infer from the files.

Claude Code: CLAUDE.md (or AGENTS.md) at the repo root, or .claude/rules/ for path-specific rules

Use pnpm, not npm. Run pnpm test before saying you're done.a build or tooling command for this repo
API handlers live in src/server/routes and must validate input with zod.a fact about this codebase

Personal preferences1

It's about you, not a project. Put it once at the user level so it follows you everywhere instead of being copied into every project.

claude.ai: Settings → Profile (personal preferences) or a Style · Claude Code: ~/.claude/CLAUDE.md

I prefer short answers with no emojis.about you, not a project; a personal style preference

Project instructions + knowledge1

It's context for one body of work in chat (a client, a product, a voice). Projects keep it scoped, and project files are cached so re-use costs less.

claude.ai: Projects → Instructions, with reference files in project knowledge

Our brand voice is warm, plain-spoken, and never uses jargon.domain context for a body of chat work

Skill1

It's a multi-step procedure needed only for one kind of task. A skill loads only when the task comes up, so it costs nothing the rest of the time.

SKILL.md in .claude/skills/<name>/ (Claude Code) or uploaded under Settings → Capabilities (claude.ai)

When cutting a release: bump the version, update CHANGELOG.md, tag, and push the tag.only applies to one kind of task

MCP server / connector1

It needs live access to a system (tickets, a database, chat). No amount of instruction text gives Claude access; a connector does.

claude.ai: Settings → Connectors · Claude Code: .mcp.json

Check the Linear ticket for acceptance criteria before starting.needs live access to an external system; mentions an external system

Subagent1

It's a separate job with its own context (independent review, broad research). A subagent keeps that work out of the main conversation.

Claude Code: .claude/agents/<name>.md

Get an independent review of security-sensitive diffs before merging.separate job that benefits from its own context

Delete it1

The model already does this by default, or the rule is too vague to change behavior. It costs tokens every session and buries the rules that matter.

Nowhere

Write clean, readable code and follow best practices.generic advice the model already follows

Keyword heuristics, not a model: treat low-confidence lines as a prompt to decide, not a verdict. The reasoning behind each destination: Projects vs Skills vs CLAUDE.md vs hooks. Build the pieces with the SKILL.md Builder, Hooks Builder, and CLAUDE.md Generator.

100 of 102 run entirely in your browser (Weather Widget uses Open-Meteo; What's My IP asks this site's server).

Share on X Share on LinkedIn

Saved you time? Buy me a coffee

How to use the Claude Instruction Router

Paste your CLAUDE.md, Project instructions, custom instructions, or any list of rules, one per line.
Review where each rule was routed and the reason given.
Decide on low-confidence lines yourself; they had mixed or weak signals.
Copy the routing plan, then move rules with the SKILL.md Builder, Hooks Builder, or CLAUDE.md Generator.

When to reach for it

Putting a rule in the wrong place fails quietly. A 'never commit to main' line in CLAUDE.md is advice that the model can miss late in a long session; the same rule as a hook is code that blocks the commit. A twelve-step release procedure in CLAUDE.md is loaded into every session even though you release once a week; as a Skill it loads only when you release. Personal preferences copied into every Project drift apart; at the profile level they're written once.

The router looks for the signals that decide placement: hard always/never rules around an action (hook), procedures for one kind of task (Skill), facts about a codebase (CLAUDE.md), things about you (preferences), domain context for chat work (Project), live access to another system (connector), and generic advice the model already follows (delete). It's a keyword heuristic, so it explains each decision and flags the ones it isn't sure about.

Frequently asked questions

What's the difference between a Claude Project and a Skill?

A Project is a workspace in claude.ai: instructions and reference files that apply to every chat inside it. A Skill is a packaged procedure that Claude loads only when a task calls for it, in any chat or project. Use a Project for context about a body of work and a Skill for how to do a specific kind of task.

Should a rule go in CLAUDE.md or a hook?

If breaking the rule once is acceptable, CLAUDE.md is fine. If it must never happen (committing to main, reading .env, skipping the formatter), use a hook. CLAUDE.md is loaded as context and followed most of the time; a hook runs as code on every matching tool call.

What's the difference between CLAUDE.md and a Skill?

CLAUDE.md loads at the start of every Claude Code session, so it should hold short facts most sessions need. A Skill loads only when its description matches the task, so it can hold long procedures without costing anything the rest of the time.

Where do I put personal preferences like 'no emojis'?

At the user level, not in each project: your profile preferences or a Style in claude.ai, and ~/.claude/CLAUDE.md in Claude Code. Then they follow you everywhere and live in one place.

How to use the Claude Instruction Router

Paste your CLAUDE.md, Project instructions, custom instructions, or any list of rules, one per line.

Review where each rule was routed and the reason given.

Decide on low-confidence lines yourself; they had mixed or weak signals.

Copy the routing plan, then move rules with the SKILL.md Builder, Hooks Builder, or CLAUDE.md Generator.

When to reach for it

Frequently asked questions

What's the difference between a Claude Project and a Skill?

Should a rule go in CLAUDE.md or a hook?

What's the difference between CLAUDE.md and a Skill?

Where do I put personal preferences like 'no emojis'?

At the user level, not in each project: your profile preferences or a Style in claude.ai, and ~/.claude/CLAUDE.md in Claude Code. Then they follow you everywhere and live in one place.

Claude Instruction Router

How to use the Claude Instruction Router

When to reach for it

Frequently asked questions

Read more

Related tools

Claude Instruction Router

How to use the Claude Instruction Router

When to reach for it

Frequently asked questions

Read more

Related tools