Orchestrator Lite Cloud

View saved version →

---
name: orchestrator-lite-cloud
description: Run a bounded, evidence-led research workflow in Cloud Work for Japanese or English requests. Route automatically for independent, read-heavy research, especially comparisons or synthesis across multiple named primary sources that require citations or an explicit evidence boundary; honour FORCE and OFF phrases; optionally use one read-only Cloud Researcher when a visible child-agent capability creates concrete value.
---

# Orchestrator Lite Cloud

Use this Skill for bounded research that benefits from an explicit evidence plan. It is a routing and quality-control workflow, not a requirement to delegate.

Respond in the language used by the user unless the user explicitly requests another language. The workflow and quality rules are identical for Japanese and English requests.

## 1. Select the mode

Classify the user's current request before researching. OFF overrides FORCE; FORCE overrides AUTO.

| Mode | Trigger | Required behaviour |
|---|---|---|
| OFF | The user says `Orchestratorなし`, `このTaskではSkillを使わない`, `Do not use the Orchestrator`, or `Do not use this Skill`. | Work directly. Do not describe the task as orchestrated, do not create a Researcher plan, and do not invoke a child agent. |
| FORCE | The user says `Orchestratorを使って`, `Use the Orchestrator`, or explicitly invokes `orchestrator-lite-cloud`. | Use this workflow. A separate Researcher is still optional. |
| AUTO | Default. | Use this workflow when independent, read-heavy research can materially improve the outcome, including bounded comparison or synthesis across multiple named primary sources. Keep simple, direct tasks in Main. |

For AUTO, positive routing signals include:

- comparing, synthesizing, or validating two or more named primary sources, papers, standards, official documents, or supplied links;
- reconciling source-by-source definitions, claims, disagreements, or Unknowns;
- requiring claim-level citations together with an explicit source boundary and stopping condition; or
- independently validating a consequential factual premise before a decision.

Citation, quotation, or document-formatting Skills may support this workflow. They do not replace the bounded research brief when the positive AUTO signals above make evidence planning material.

For AUTO, keep work in Main when the request is answerable from one supplied item, needs tightly coupled reasoning or editing, has a small/clear fact check, or would not gain materially from a separate evidence pass. Do not turn ordinary task execution into orchestration merely because research is possible.

## 2. Open a bounded research brief

When FORCE applies, or AUTO selects this workflow, state a compact brief before substantive research:

1. **Objective** — what must be established or produced.
2. **Decision** — what concrete decision, design choice, or recommendation the research will inform. State `None — explanatory research only` when applicable.
3. **Sources in scope** — supplied files/links first; then the specific source classes or named sources to examine. Prefer primary sources for external facts (official documentation, original papers, standards, filings, source data, or direct statements).
4. **Completion criteria** — the claims/questions that must be resolved, the evidence needed for each, and the intended output. Include an explicit stopping boundary.

Reuse supplied material before retrieving duplicates. Treat secondary sources as leads or context unless the needed primary source is unavailable; mark that limitation. Separate verified facts, reasoned inferences, and unresolved items.

## 3. Decide whether a Researcher is justified

Main owns the task by default. Delegate only when **all** conditions hold:

- the active Work environment visibly exposes a native Cloud child-agent capability;
- a read-only evidence pass can proceed independently from Main's analysis;
- the separate pass has a concrete benefit, such as checking a distinct prima

Preview of the first 4,000 characters. Read full instructions →