From OpenAI Developers
Use when a developer needs a ChatGPT Apps submission import file from an MCP server. Inspect the implementation, generate chatgpt-app-submission.json with app info, tool-hint justifications, positive and negative test cases, then report review findings and outputSchema warnings.
How this skill is triggered — by the user, by Claude, or both
Slash command
/openai-developers:chatgpt-app-submissionThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Use this skill when a developer needs a `chatgpt-app-submission.json` file for a ChatGPT Apps submission. The file is uploaded in the Apps submission form to fill out parts of App Info, MCP Server, and Testing.
Use this skill when a developer needs a chatgpt-app-submission.json file for a ChatGPT Apps submission. The file is uploaded in the Apps submission form to fill out parts of App Info, MCP Server, and Testing.
readOnlyHint, openWorldHint, and destructiveHint annotations, and whether it declares outputSchema.chatgpt-app-submission.json in the current working directory.outputSchema warnings in the final response, and explain what the developer should do with each finding before submission.Do not infer behavior from the tool name alone. Use the real tool implementation and declared annotations. If a tool calls into another module or API client, inspect enough of that path to know whether it reads, writes, deletes, sends, publishes, or changes external state.
Suggest app info from source-of-truth project metadata and the tool behavior you inspected. Keep it plain-language and submission-review-facing.
display_name: use the product or app name from repo metadata, package metadata, README, manifest, or existing configuration. Keep it short enough for the submission form.subtitle: summarize what the app does in one short functional phrase, not marketing copy. It must be 30 characters or less.description: describe concrete user value and the main workflows the tools support.category: choose one of BUSINESS, COLLABORATION, DESIGN, DEVELOPER_TOOLS, EDUCATION, ENTERTAINMENT, FINANCE, FOOD, LIFESTYLE, NEWS, PRODUCTIVITY, SHOPPING, or TRAVEL.Use the Apps SDK review meanings:
readOnlyHint: true only when the tool strictly fetches, looks up, lists, retrieves, or computes data without changing state. false if it can create, update, delete, send, enqueue, run jobs, write logs, start workflows, or otherwise mutate state.destructiveHint: true if the tool can delete, overwrite, send irreversible messages or transactions, revoke access, or perform destructive admin actions, including via some modes or parameters. Otherwise false.openWorldHint: true if the tool can change publicly visible internet state or external third-party systems, such as sending emails or messages, posting/publishing content, creating public tickets/issues, pushing code/content, or submitting external forms. false if it only operates in closed/private systems.ChatGPT Apps submissions require every tool to set all three hints explicitly. Missing or null hints are submission blockers, even if MCP clients may have protocol-level defaults.
If a hint is missing, null, or behaviorally wrong, stop before writing the JSON and ask the developer for approval to update source. List each affected tool, the current or missing hint, the behavior you observed, and the recommended explicit value. If approval arrives, make the smallest source change and continue; otherwise report the blocked mismatch instead of generating misleading JSON.
While inspecting exposed MCP tools, record each tool whose descriptor or source definition omits outputSchema or sets it to null. Missing outputSchema is not a blocker for generating chatgpt-app-submission.json, and the submission JSON does not include output schemas.
Do not infer or invent output schemas for this warning. Use the actual MCP tool descriptor or source definition. In the final response, include a concise warning for any missing tools: Add an outputSchema so models can use this tool's results more reliably. See https://modelcontextprotocol.io/specification/draft/server/tools#tool. Include the affected tool names. If every tool declares outputSchema, do not include an outputSchema warning.
Check tool names, tool descriptions, and widget CSP metadata while inspecting the app.
If tool names, descriptions, or CSP values are missing or inconsistent with behavior, ask for approval before editing source. If approval does not arrive or the edit is ambiguous, keep the JSON truthful and report the finding.
Generate at least five positive test cases and at least three negative test cases.
tools_triggered.Write exactly one JSON file named chatgpt-app-submission.json:
{
"$schema": "https://developers.openai.com/apps-sdk/schemas/chatgpt-app-submission.v1.json",
"schema_version": 1,
"app_info": {
"display_name": "Example App",
"subtitle": "Find and update records",
"description": "Example App helps users find records, inspect details, and update workspace data through ChatGPT.",
"category": "PRODUCTIVITY"
},
"tools": {
"tool_name": {
"annotations": {
"readOnlyHint": true,
"openWorldHint": false,
"destructiveHint": false
},
"justifications": {
"read_only_justification": "Only retrieves matching records and does not modify data.",
"open_world_justification": "Does not write to public internet state or third-party systems.",
"destructive_justification": "Does not delete, overwrite, revoke access, or perform irreversible actions."
}
}
},
"test_cases": [
{
"description": "Find records that match a specific user request.",
"user_prompt": "Find my open records for this week.",
"file_attachment_urls": null,
"tools_triggered": "tool_name",
"expected_output": "Returns matching records with enough detail for the user to choose the next action.",
"expected_output_url": null
}
],
"negative_test_cases": [
{
"description": "Do not trigger for unrelated calendar requests.",
"user_prompt": "What meetings do I have tomorrow?",
"file_attachment_urls": null,
"tools_triggered": null,
"expected_output": "The app should not be invoked because the request is outside its supported workflows.",
"expected_output_url": null
}
]
}
$schema identifies the import file shape for editors and importers; Claude Code does not need to fetch it. tools is required. app_info, test_cases, and negative_test_cases are optional in the schema, but generate them whenever the repo contains enough information. Do not include review-check findings in this JSON file.
Report these checks in the final response after writing chatgpt-app-submission.json. Do not write them into the JSON file.
For each finding, explain the practical next step: update source, update submission copy, narrow CSP, remove or justify a sensitive input, or manually review before submitting. If there are no findings, say that these checks did not find obvious issues from source inspection.
After writing the file, summarize the app info fields generated, number of tools covered, positive test case count, and negative test case count. Then include a Review findings section with any sensitive data solicitation, tool data use, tool naming, weak CSP findings, or missing outputSchema warnings and what to do with each one. If generation is blocked, lead with the exact missing hints or source ambiguity.
npx claudepluginhub openai/openai-developers-for-claude --plugin openai-developersCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.