/atelier-orchestrator

name: atelier-orchestrator
description: >
  Skill routing and workflow orchestration. Selects Inline Plan or Spec-backed Plan, routes
   to the correct workflow skill, and manages transitions between phases. Use when starting
  any conversation or task to determine which planning mode and skill apply.
user-invocable: false

<EXTREMELY-IMPORTANT> If you think there is even a 1% chance a skill might apply to what you are doing, you ABSOLUTELY MUST invoke the skill.

IF A SKILL APPLIES TO YOUR TASK, YOU DO NOT HAVE A CHOICE. YOU MUST USE IT.

This is not negotiable. This is not optional. You cannot rationalize your way out of this. </EXTREMELY-IMPORTANT>

Planning Workflow

You have skills. You MUST use them. Not "should." Not "when convenient." MUST.

Invoke relevant skills BEFORE any response or action. Even a 1% chance a skill might apply means you invoke it. If an invoked skill turns out to be wrong for the situation, you don't need to follow it. But you must check.

The One Rule

**When an implementation choice exists, never write code until the human has reviewed and approved a written plan.** Exact mechanical changes may use the exception below.

The plan may be inline in the conversation or backed by persisted spec artifacts. Select the mode before invoking an artifact-producing skill, state the choice with one brief reason, and let the human override it.

Governing Principles

Invoking a skill selects relevant guidance. It does not make every pattern in every loaded skill part of the solution. Existing source, explicit requirements, and the smallest correct change determine which patterns apply.

Begin with the current design and the minimum behavior-preserving change. Product, infrastructure, and architectural expansion require explicit scope. Keep **ponytail** active throughout development when it is available or required by the repository. Keep **writing-clearly-and-concisely** active whenever writing or editing prose for humans. Before returning or persisting that prose, apply **humanizer** in embedded mode.

Planning Modes

Inline Plan (default)

Use for bounded, well-understood work, including ordinary features, bug fixes, refactors, configuration changes, and multi-file changes. Route to **spec-plan** with Inline mode explicitly selected. It presents a concise plan in conversation, creates no planning artifacts or tracker entries, and stops for approval. Approval ends the `spec-plan` invocation. A later, explicit implementation request may implement the plan directly; do not invoke `spec-implement`, `spec-finish`, or `code-subagents` for Inline work.

Direct Inline implementation must:

• Avoid main/master unless the human explicitly permits it
• Stay within the approved scope and files
• Run every validation item in the approved plan
• Review the final diff for unintended changes
• Do not commit, push, or open a PR without explicit human instruction

When in doubt, choose Inline Plan. File count and estimated duration alone do not make work spec-worthy.

Spec-backed Plan

Use when at least one concrete signal exists:

• The human explicitly requests a spec or durable planning artifacts
• Requirements or boundaries need discovery before planning
• The work materially changes architecture, public contracts, data models, or cross-system behavior
• Multiple dependent subsystems need decomposition and dependency tracking
• The plan must survive the conversation for handoff, audit, or multi-person coordination

Route through **spec-brainstorm**. This mode produces:

design.md    ← spec-brainstorm (requirements + research + architecture)
plan.json    ← spec-plan (tasks, dependencies)

Each skill produces only its declared output and stops. The human starts the next phase with a separate request.

Do not infer substantial work from size alone. If no concrete signal applies, use Inline Plan.

Skill Routing

atelier-orchestrator → Select and announce planning mode
spec-brainstorm   → Spec-backed discovery + design → only design.md → stop
spec-plan         → Inline conversation or only plan.json → stop
spec-implement    → Execute approved Spec-backed Plans; track tasks
spec-finish       → Post-implementation validation for Spec-backed Plans
spec-loop         → Run an approved Spec-backed Plan unattended to a draft PR
code-subagents    → Parallel dispatch for Spec-backed Plan tasks

Ordinary bounded work

spec-plan (Inline mode) → approval → STOP
new implementation request → implement directly

Substantial work or explicit spec request

spec-brainstorm → design.md → STOP
new spec-plan request → plan.json → STOP
new spec-implement request → implementation → spec-finish

Truly mechanical change

A typo or exact single-line replacement may skip planning when there is no implementation choice to review. If any choice exists, use an Inline Plan.

Delegation

Delegate bounded, read-only work to a subagent when it would otherwise flood the main context: wide codebase research, finding comparable code, or independent review. Ask for a concise, file:line-cited summary, and read the key cited files yourself before deciding. Do not delegate interviews or design dialogue with the human, lookups of a few files, or stateful debug loops.

Delegate implementation too. Once the design is settled, implementation batches go to **Keymaker** subagents through **code-subagents**; the main conversation edits code directly only for trivial or mechanical changes and debugging loops.

Every subagent launch must explicitly select an agent. Match the work to the descriptions of Atelier's agents (Sentinel, Oracle, Architect, Keymaker) and choose one. Never launch without a selection or rely on the harness default. Exception: design-sensitive implementation may use the harness's general agent. If Atelier's agents are not installed, explicitly choose the closest available agent by description and state the choice in the output.

Parent