aistudio

name	aistudio
description	Use for creating or modifying AI Studio app and workflow project resources, including agentic apps, workflows, business objects/BO nodes, policies, tools, connectors, approvals, document schemas, and function templates. Route requests, load the relevant bundled prompt references, and use the bundled aistudio CLI for local work.

AI Studio

Copyright: Copyright (c) 2026 Oracle and/or its affiliates License: UPL (Universal Permissive License)

Use this skill to create or modify AI Studio app and workflow project resources, including agentic apps and related local project resources.

Start Here

1. Classify the requested artifact before choosing prompts or commands.
2. Before creating any new artifact, resolve package layout from the current project. If multiple app packages exist and the current user request did not explicitly name one, ask for the target app package and stop before any discovery, planning, generation, test sync, or create command.
3. Read references/prompts/index.md to see available prompt references.
4. Load only the prompt files relevant to the task.
5. Keep the shell working directory at the AI Studio project root. Do not cd into the skill directory to run commands.
6. Run the bundled CLI by path, for example node .agents/skills/aistudio/scripts/aistudio.js <command> ... from the project root. If the skill is installed elsewhere, use that skill directory's scripts/aistudio.js path while keeping cwd at the project root.
7. Treat prompt examples that start with aistudio as shorthand for the bundled script path. Do not search PATH, install a global CLI, or run which aistudio / command -v aistudio unless the user explicitly asks for global CLI setup.
8. Run init only when the user explicitly asks to initialize or scaffold a blank project. Do not run init as a prerequisite for creating apps, workflows, or other artifacts in an existing workspace.

Artifact Routing

Natural-language artifact phrases are primary. File extensions, file paths, and explicit command names are override signals.

User intent	Artifact	Primary reference
workflow, workflow node, step, edge, connect, wire, tool node, add a tool to a workflow	.wf workflow	workflow-vibe.md; node details in workflow-node-prompts/
app, agentic app, app panel, app communication, app template, app action	.app app	app-ingestion.md, app-best-practices.md, app-vibe-master.md plus app-specific prompts
business object	.bo Business Object	business-object-builder.md
business object tool, BO tool, deeplink tool, document tool, email tool, external REST tool, MCP tool, connector tool	.tool Tool	tools-builder.md
deeplink or deep link source artifact	.dl Deeplink	artifact-conventions.md plus CLI help
topic or instructions	.topic Topic	artifact-conventions.md plus CLI help
agent	.agent Agent	artifact-conventions.md plus CLI help
connector definition or generated connector from OpenAPI/MCP spec	.connectorDefinition Connector Definition	connector-definition-cli-compat.md
connector instance or configured connector	.connectorInstance Connector Instance	connector-instance-cli-compat.md
approval process	.approval Approval Process	approval-process-builder.md
policy or policy store	.policy Policy Store	policy-store-builder.md
policy template	.policyTemplate Policy Template	policy-store-builder.md and policy-store-cli-compat.md

If wording could still mean more than one artifact, ask one targeted question: Tool artifact, source artifact, or workflow node?

Workflow And App Boundary

Before building or modifying a workflow, decide whether it is standalone or app-backed.

• Standalone workflow intent: the user asks for a workflow, automation, node, or debug flow without mentioning an app.
• App-backed workflow intent: the user says the workflow is for an app, app agent, app panel, app communication, app template, target agent, InitDisplay, InitActions, InitCommunications, FillParameters, SendCommunication, AI Apps, Agentic App, or app-compatible workflow.
• For standalone workflows, read workflow-vibe.md and the relevant workflow-node-prompts/* files. Do not add app-stage routing or aiAppsCompatibleFlag.
• For app-backed workflows, read both workflow references and app-vibe-master.md / app-vibe-templates.md as needed. App configuration determines which app-stage paths are required.
• Only set or change aiAppsCompatibleFlag when app intent is explicit.
• When app intent is explicit, the workflow is incomplete until the required app-stage paths are real and routable. See the app prompt references for InitDisplay, InitActions, InitCommunications, FillParameters, and SendCommunication requirements.
• For app-backed workflows, route on $context.$app.$OraMessageHint in the graph, usually with a SWITCH node. Do not put multiple app-stage branches into one LLM/AGENT prompt.
• For app-backed workflows with shared recordable data, shared fetch/guard nodes must run before the app-stage router; only terminal LLM/AGENT nodes need to be stage-specific.
• A single-node app-backed workflow is invalid when it must handle more than one app message hint. For InitDisplay plus Query, create separate terminal nodes behind the OraMessageHint router.

Reference Loading

• Workflow authoring: read workflow-vibe.md; for node work, also read workflow-node-prompts/index.md and the relevant node prompt files. When a workflow is created or materially edited, also read workflow-test-authoring.md before concluding so the automatic test sync handoff is visible.
• Workflow debugging: read workflow-debug.md.
• Workflow test authoring: read workflow-test-authoring.md. For scripted multi-turn creation, editing, recording, execution, or diagnosis, also read workflow-conversation-test-authoring.md; it extends rather than replaces the single-turn contract. The semantic judge section owns the model-authored judge flow and the read-only get-workflow-test-judge-context contract retrieval command. Treat needs-model-test-data as a nonterminal continuation and complete its compaction/apply flow before continuing workflow or app sync.
• Distinguish preview from execution. Requests like "show", "inspect", "preview", "display", "what would be generated", or "get the sync plan" are read-only plan requests: run only the relevant get-*-test-sync-plan command and explain the plan. Requests like "run", "execute", "generate", "regenerate", "sync", "create/update the tests", or "complete the sync plan" are execution requests: complete the workflow or app test sync loop from the relevant prompt reference. Do not stop after printing a plan when executable missing or out-of-sync actions remain and no user decision is required.
• Final-response checkpoint for test-sync execution requests: a latest workflow plan with finalSummaryAllowed: false or coverageExecution.nextActionIds, or a latest app plan with required backing workflow actions, prohibits a final response. Execute the listed next action or required workflow command instead. A summary of pending counts is not a valid execution result. Follow the detailed completion and blocker rules in workflow-test-authoring.md and app-test-authoring.md.
• Workflow optimization: when the user asks about model costs, token usage, model comparison, cost reduction, BO payload size, prompt verbosity, or which model to use on a node, read workflow-test-authoring.md for the optimization command reference. Do not run optimization commands automatically during normal workflow creation or editing.
• App intake for new app work, material app edits, and app-backed workflow creation: read app-ingestion.md before asking questions or starting implementation. A confirmed BO or other data source does not complete intake by itself.
• App authoring: read app-best-practices.md and ; add , , , or only when relevant.

Resources

• App samples live under resources/app-samples/. Read resources/app-samples/index.md before using app samples; it lists each available sample file and what it demonstrates.

Global CLI Rules

• Prefer bundled CLI commands over direct JSON edits when a command supports the requested operation.
• Invoke the bundled CLI by path from the project root, such as node .agents/skills/aistudio/scripts/aistudio.js <command> .... Bare aistudio <command> examples in references are shorthand only.
• Never use npm run or npm test to execute workflow tests, optimization sweeps, or any aistudio command. The app-pkg/package.json and root package.json npm scripts are for CI/CD pipelines only. Always invoke the CLI binary directly.
• Never pass --judge-provider remote to run-workflow-test, run-workflow-tests, or any optimization sweep command unless the user explicitly requests remote judging. This applies regardless of workspace type; the presence of app-pkg/package.json with remote-judge npm scripts does not change this default.
• Treat data source and evaluation mode as independent. A user request for file mode means pass --data-source file only. For an ordinary test run, omit --evaluation-mode so each test uses its configured mode, and omit --judge-provider so local judging remains the default. Pass --evaluation-mode deterministic only when the user explicitly requests deterministic-only execution or when test-authoring guidance requires transient validation of one affected test. Never use deterministic execution as a final confirmation after local judge attachment.
• When an authenticated AI Studio command, including a normal draft save, fails while sandboxed and the host keystore or credential store may be unavailable, do not conclude that authentication or the operation failed. Retry the same command outside the sandbox so it can access the host credential store, requesting escalation if required. Enter the AUTH_REQUIRED recovery flow only if the unsandboxed retry also returns AUTH_REQUIRED.
• For basic/dev-mode pre-production or PD environments, normal artifact commands never prompt for the FA password. If any CLI command returns JSON with code: "AUTH_REQUIRED", do not treat that JSON as the final result and do not simply report the manual command. Pause the original command and recover authentication first:
  • In a real foreground terminal, run the returned interactiveCommand so the user can type into the hidden prompt, then retry the original command.
  • In background terminal UIs that cannot expose hidden password prompts, say: "This UI cannot expose the hidden password prompt." Then present both recovery choices: the user can run the returned in their own terminal, or the user can explicitly authorize passing a password they provide to the returned using .

Local Layout

Artifact	Default location
Workflow	src/workflows/*.wf
App	src/apps/*.app
Agent	src/agents/*.agent
Business Object	src/businessObjects/*.bo
Deeplink	src/deeplinks/*.dl
Topic	src/topics/*.topic
Tool	src/tools/*.tool
Connector Definition	src/connectorDefinitions/*.connectorDefinition
Connector Instance	src/connectorInstances/*.connectorInstance
Approval Process	src/approvals/*.approval
Policy Store	src/policies/*.policy
Policy Template	src/policyTemplates/*.policyTemplate
Document Schema	src/documentSchemas/*.documentSchema
Function Template	src/functions/*.function

init is for explicit blank-project setup only; artifact creation commands write their own target files and should not be preceded by init in an existing project.

Verification

• After material workflow edits, run node .agents/skills/aistudio/scripts/aistudio.js do-prettify-workflow --file <workflow-file> after structural batches, then run node .agents/skills/aistudio/scripts/aistudio.js validate-workflow --file <workflow-file>.
• For condition, switch, code, return, or other expression-bearing workflow node edits, use file-backed JSON patch inputs such as --inputs-patch @.debug/<name>-inputs.json; do not inline JSON containing empty-string literals, quotes, $context, {{...}}, comparison operators, or expression syntax through the shell. For condition guards, prefer quote-free existence checks such as {{!!...}} when equivalent. After modifying a guard expression, read back or validate the workflow before recording tests so shell quoting did not strip literals. If live recording reaches an unexpected fallback, inspect guard expressions before reporting an environment blocker.
• After a successful workflow create or material workflow edit, do not stop at prettify/validation. Unless the user explicitly opted out or asked only for analysis/planning, continue with the automatic workflow test sync loop from workflow-test-authoring.md, starting with node .agents/skills/aistudio/scripts/aistudio.js get-workflow-test-sync-plan --file <workflow-file>. This applies to new workflow files, node additions, node edits, edge rewiring, prompt/code changes, behavior-changing metadata changes, and direct workflow-file repairs when the CLI did not support the operation.
• After a successful app create or material app edit, run the app flow from app-test-authoring.md for apps with top-level agent containers. That prompt owns scope selection and the detailed sequence. An app-wide request completes every required backing workflow and app action, runs do-sync-app-tests --file <app-file>, and finishes with get-app-test-final-summary --file <app-file>. A request naming one panel, one app test, or one backing workflow together with its panel completes only that workflow and matching panel action, runs only run-app-test --test-file <target-app-test>, and does not call the app-wide sync or final-summary commands. do-sync-app-tests does not generate, record, or run workflow tests. If the app is only an empty shell with no top-level agent containers, report that app tests will be created after a panel is added; do not fabricate tests for an untestable shell.
• When standalone workflow tests finish, run node .agents/skills/aistudio/scripts/aistudio.js get-workflow-test-final-summary --file <workflow-file> and use that command output as the evidence source for the final response section named exactly after your workflow creation, change, or sync-outcome summary. Preserve the test counts, workflow suite report, , and . Do not hand assemble them from suite JSON or rerun tests merely to refresh reports.

Scope

This is the base AI Studio skill. It should stay concise and point to bundled prompt references for artifact-specific details.