modes-pattern

---

name: modes-pattern description: Sets up the router + modes multi-agent pattern in a Claude Code project. The pattern is a dispatch contract where the orchestrator (main-thread Claude) routes every domain request to a specialist (subagent OR forked skill) running a per-task playbook in isolated context. Use this skill PROACTIVELY whenever the user wants to build a "modes" system in any project, scaffold a router skill, set up specialist subagents in .claude/agents/, organize per-task playbooks in .claude/modes/, design a routed dispatcher for any domain with repeat task types, or asks "how do I set up modes?", "build me a router skill", "scaffold a multi-agent dispatch pattern", "make Claude route my requests to specialists", "set up the modes pattern", "build a _shared.md prefix", "add a new mode to my router", "how do I do routed subagents", "make a routed multi-agent system", "what's the modes pattern". Also fire when reading or extending an existing .claude/modes/ directory and the user needs the architecture reference. Domain-agnostic — works for any project where one operator dispatches repeat tasks across a fixed set of categories. Covers BOTH implementation bases: agents (.claude/agents/*.md + Agent() calls) and skills (context: fork + agent: general-purpose + Skill() calls). Includes scaffolding scripts plus templates for every file in the pattern. Don't answer modes-setup questions from training-data guesses; load this skill first.

The modes pattern

A reusable architecture for building multi-agent dispatch systems on top of Claude Code. One operator, many repeat task types, deterministic routing, isolated execution per task, runtime enforcement of cross-task rules.

This skill teaches Claude Code how to scaffold and extend the pattern correctly. Read the workflow below. Pull in references on demand for deep dives.

Trigger handling

When the user expresses any of the trigger intents in the description, do not answer from training-data guesses about "multi-agent patterns." Load this skill's references (below) and follow the workflow. The pattern has specific files, specific frontmatter fields, specific dispatch semantics. Generic advice will produce a system that does not actually route.

The pattern in 60 seconds

USER TYPES SOMETHING
        ↓
Router skill auto-loads on description-matched phrase
        ↓
Orchestrator follows the dispatch contract:
  1. Read _shared.md from disk
  2. Read the matched mode file from disk
  3. Compose prompt = shared + mode + user_ask
  4. Call Agent() or Skill() to dispatch
        ↓
Specialist runs in ISOLATED context with its own:
  - tool allowlist
  - MCP scoping
  - system prompt (identity)
  - playbook (mode body)
        ↓
Writes artifact to a deterministic output path
        ↓
Returns summary; specialist context discarded

The orchestrator's job stays tiny. Every domain rule lives in a file loaded fresh on each run.

The two implementation bases

Before scaffolding, decide which base fits the project. The pattern's structure is identical; the primitives differ.

Aspect	Agent base	Skill base
Execution primitive	Agent(subagent_type=...)	Skill(skill=...) with context: fork + agent: general-purpose
Specialist identity	Agent body (.claude/agents/<name>.md) — stable across all modes that specialist serves	Skill body — carries identity AND playbook together
Per-task playbook	Mode file in .claude/modes/	Skill body
Tool scoping	Per-specialty via agent tools allowlist	Per-mode via skill allowed-tools (finer-grained)
MCP scoping	Per-specialty via agent mcpServers	Per-mode via allowed-tools listing mcp__* entries
Persistent memory	memory: project (runtime-managed)	Convention (skill body reads/writes a known path)
Best for	Modes that share state per specialty, heavy fan-out, strict MCP isolation	Portable projects, mostly single-shot modes, per-mode tuning

Pick agent base when modes will run thousands of times, share persistent state per specialty, or need runtime-level MCP isolation. Pick skill base when the project is lightweight, modes are single-shot, or finer per-mode scoping matters more than per-specialty identity reuse. Both can coexist in one project.

Full comparison: ${CLAUDE_SKILL_DIR}/references/skill-base.md.

Setup workflow

Step 1: Confirm scope and base with the user

Before writing files, confirm:

• The domain the system serves (custom to the project — e.g. research, content production, support workflows, ops runbooks, data extraction, anything with repeat task types).
• The list of mode names (one per task type). 3–10 is typical; do not exceed 12 without splitting domains.
• The list of specialist names if using agent base (often fewer than modes; one specialist can serve many modes).
• The output root path (.work/, .manager/, .artifacts/, etc.).
• The base (agent / skill / both). Default to agent base when in doubt — it scales better and the runtime memory feature pays back as the project matures.

Step 2: Scaffold the directory layout

Run the bundled scaffolding script:

bash ${CLAUDE_SKILL_DIR}/scripts/scaffold.sh <project-root> <base>

Where <base> is agent, skill, or both. The script creates the canonical directory tree and copies templates into place. See ${CLAUDE_SKILL_DIR}/scripts/scaffold.sh for what it does; do not reinvent it.

The canonical layout (agent base):

.claude/
├── agents/           # Specialist identities (Layer 4)
├── modes/            # _shared.md + per-task playbooks (Layers 2, 3)
├── skills/<router>/  # Router skill (Layer 1)
├── hooks/            # Runtime enforcement (Layer 11)
├── config/           # Tunable YAML (Layer 7)
├── context/          # Operator source-of-truth files (Layer 6)
├── tools/            # Project-local CLIs (Layer 10)
├── agent-memory/     # Persistent per-specialist state (Layer 5)
└── settings.json     # Permissions, hook registrations (Layer 12)
<output-root>/        # Artifacts (Layer 8)
.mcp.json             # MCP server registry (Layer 9)

For skill base, replace agents/ and modes/ with the mode-skills going directly into .claude/skills/<mode>/SKILL.md.

Step 3: Write _shared.md

Single file at .claude/modes/_shared.md. Prepended to every dispatch. Carries:

1. Operator profile — who the system is for; hard scope, out-of-scope rules.
2. Source-of-truth file paths — what's in .claude/context/.
3. Tool inventory — exact mcp__server__tool identifiers and what each is for. Specialists hallucinate tool names without this.
4. Cross-cutting style or policy rules — whatever rules apply to every artifact regardless of mode (tone constraints, banned terms, citation requirements, format guarantees, anything the project cares about enforcing uniformly).
5. Hard rules — non-negotiables that apply to every mode.

Target 100–200 lines. If it grows past that, you are putting per-task content in the wrong file (move to mode files).

Template: ${CLAUDE_SKILL_DIR}/templates/_shared.md.tmpl. Adapt placeholders to the domain.

Step 4: Write the first mode file

Every mode follows the five-section template:

# Mode: <mode-name>

**Subagent**: `<specialist-id>`
**Trigger phrases**: "<phrase 1>", "<phrase 2>"
**Output**: <output path template>
**Approval gate**: required (only if mode has side effects)

## Inputs
- **Required**: <param>
- **Optional**: <param> (default: ...)

## Steps
1. Read context files from `_shared.md`.
2. Call <exact tool name> with <args>.
3. <Domain logic>.
4. <STOP condition if input is out of scope>.
5. Write artifact to the output path.

## Output schema
<literal markdown — actual headers, table columns, placeholders>

## Quality bar
- <pass/fail rule 1>
- <pass/fail rule 2>

Template: ${CLAUDE_SKILL_DIR}/templates/mode.md.tmpl. Full section-by-section spec: ${CLAUDE_SKILL_DIR}/references/architecture.md → "Layer 3: Mode files."

Step 5: Write the specialist agent (agent base only)

Skip this step if using skill base.

---
name: <specialist-id>
description: <when to delegate>
tools: [<tools the specialist needs>]
disallowedTools: [<tools to deny>]
model: sonnet
maxTurns: 30
mcpServers: [<servers connected at startup>]
memory: project
skills: [<preloaded skill bodies>]
---

<system prompt body — persistent identity, tool-usage rules, output formats, hard constraints>

The agent body is stable across every dispatch this specialist receives. Per-task playbooks belong in mode files, not the agent body.

Template: ${CLAUDE_SKILL_DIR}/templates/agent.md.tmpl. Full frontmatter spec: ${CLAUDE_SKILL_DIR}/references/architecture.md → "Layer 4: Specialist agents."

Step 6: Write the router skill

The router auto-loads on description-matched phrases and forces the dispatch contract every turn. Description must pack every trigger phrase from every mode header — coverage matters more than brevity.

Frontmatter:

---
name: <router-name>
description: Routes <domain> requests to the right specialist. ALWAYS use this skill when the user expresses any <domain> intent — "<phrase 1>", "<phrase 2>", ... [every phrase from every mode]. Do NOT answer <domain> questions directly in the main thread. Skip ONLY when the user is editing the system itself.
---

Body has four parts: routing table, dispatch contract, chained-dispatch table, "when NOT to trigger" list.

Template: ${CLAUDE_SKILL_DIR}/templates/router-skill.md.tmpl. Full spec: ${CLAUDE_SKILL_DIR}/references/architecture.md → "Layer 1: Router skill."

Step 7: Register hooks in settings.json

Hooks are the third enforcement layer. They fire on runtime events and block bad output even when the specialist drifts.

Minimum useful set:

• SessionStart → setup-tools-path.sh — prepends .claude/tools/ to PATH so project scripts resolve as bare commands.
• (Optional) PreToolUse (Write|Edit) → artifact-content-guard.sh — example mechanism for any project-defined policy that should block writes (banned phrases, secret patterns, schema invalidity, missing headers, etc.). The template ships as a .tmpl because the policy is project-specific; the operator adapts the CHECK section before registering it.
• SubagentStop → dispatch-audit.sh — logs every dispatch (mode, specialist, duration, artifact path).

Templates: ${CLAUDE_SKILL_DIR}/templates/hooks/*.sh.tmpl. Full spec: ${CLAUDE_SKILL_DIR}/references/architecture.md → "Layer 11: Hooks."

Register in settings.json:

{
  "hooks": {
    "SessionStart": [{ "hooks": [{ "type": "command", "command": "$CLAUDE_PROJECT_DIR/.claude/hooks/setup-tools-path.sh" }] }],
    "SubagentStop": [{ "hooks": [{ "type": "command", "command": "$CLAUDE_PROJECT_DIR/.claude/hooks/dispatch-audit.sh" }] }]
  }
}

Step 8: Test the dispatch

Have the user type one of the router skill's trigger phrases. Verify:

1. Router skill auto-loads (shown in available-skills list).
2. Orchestrator reads _shared.md and the matched mode file (visible in tool calls).
3. The right specialist spawns with description="<mode>".
4. The artifact lands at the templated output path.
5. Hooks fire on dispatch end (audit log). Any optional content-guard hooks the operator added fire on writes.

If any step fails, see ${CLAUDE_SKILL_DIR}/references/dispatch.md → "Debugging."

Extending an existing system

Use the rituals when adding pieces. Each ritual is small and predictable. See ${CLAUDE_SKILL_DIR}/references/rituals.md for full step-by-steps:

• Adding a new mode
• Adding a new specialist
• Adding a new hook
• Adding a new MCP server
• Adding a new config value

The system is right when adding a task type is one new mode file plus one routing-table row, and nothing else.

When NOT to invoke this skill

• The user is asking a one-off question about an existing modes system that does not require architectural setup or extension.
• The user is doing routine engineering on a project that already has modes (debugging, refactoring, adding a script). Use the existing modes; don't redesign them.
• The user is asking conceptual questions about general multi-agent architectures unrelated to Claude Code's Agent() / Skill() primitives.

References

Load on demand based on what the user is doing.

• ${CLAUDE_SKILL_DIR}/references/architecture.md — full spec for all 12 layers (router, _shared.md, modes, agents, agent-memory, context, config, output root, MCP, tools, hooks, settings). Load when writing or extending any layer.
• ${CLAUDE_SKILL_DIR}/references/dispatch.md — dispatch contract, end-to-end walkthrough, Agent() call surface, chained dispatch, fresh-context implications, observability and debugging. Load when implementing the orchestrator's routing logic.
• ${CLAUDE_SKILL_DIR}/references/skill-base.md — the alternative implementation where skills replace agents as the execution base. Load when the user wants the portable / lightweight implementation OR is deciding between bases.
• ${CLAUDE_SKILL_DIR}/references/rituals.md — step-by-step procedures for adding modes, specialists, hooks, MCP servers, config values. Load when extending an existing system.

Templates

In ${CLAUDE_SKILL_DIR}/templates/:

• router-skill.md.tmpl — the router skill
• _shared.md.tmpl — the shared prefix
• mode.md.tmpl — one mode file
• agent.md.tmpl — one specialist agent (agent base only)
• mode-skill.md.tmpl — one mode-skill (skill base only)
• hooks/setup-tools-path.sh.tmpl (always-on, generic)
• hooks/dispatch-audit.sh.tmpl (always-on, generic)
• hooks/artifact-content-guard.sh.tmpl (optional — adapt the CHECK section to whatever policy the project needs)

Copy and adapt placeholders; do not invent fields.

Hard rules carried into every modes system you build

1. Read _shared.md and the mode file fresh from disk on every dispatch. Not cached, not paraphrased, not shortcut.
2. Name tools by exact identifier (mcp__server__tool or local CLI name). Never describe a tool by prose in a Steps block.
3. Required inputs are collected by the orchestrator before composing. Specialists run in fresh context and cannot ask back.
4. Side effects are gated. Modes that push to a public artifact, deploy, or send declare **Approval gate**: required and the router waits for explicit operator confirmation.
5. Hooks enforce quality bars deterministically. Prompts are advisory; hooks fire even when the specialist drifts.
6. Specialists return partial data on tool failure. Never crash a dispatch — surface the failure at the top of the artifact and continue.
7. Output paths are deterministic. Re-running a mode overwrites cleanly; the operator never gets duplicate artifacts at random paths.