Two packages from one listing source

Keep every word of the listing (names, descriptions, prompts, test cases, URLs, colours) in one file, for example plugin/listing.json, and generate both packages from it with a script. Put the store's name in the shared description as a placeholder ({app} → "ChatGPT" or "Claude"). Re-run the generator in CI with a --check flag that fails when a package is out of date.

The ChatGPT ZIP

Upload at platform.openai.com/plugins. The ZIP's root holds:

plugin.json                 $schema https://agent-plugins.org/schemas/1.0.0/plugin.schema.json
mcp.json                    the server and its auth
skills/get-started/SKILL.md an onboarding skill (name must match the folder)
assets/icon.png icon-dark.png logo.png

Shapes and limits are in references/chatgpt-package.md. The checks that the upload enforces and that bit us, all caught by scripts/check_chatgpt_package.py:

• review.test_cases.positive[].tools_triggered is one string, names joined by ", ", not a list ("Input should be a valid string").
• mcp.json auth needs authorizationUrl and tokenUrl beside client: {mode: "cimd"} ("Declare both authorizationUrl and tokenUrl for OAuth"), even though discovery gives them.
• Brand colours need contrast: dark vs #212121 at least 2:1 ("Dark brand color must have at least 2:1 contrast against #212121"); check the light one against white too.
• The demo video is required to submit for review. Leave an empty demo_recording_url out (the upload rejects it), and tell the user they still need the video link before submitting: in plugin.json, or pasted in the portal's Supporting content step.
• Exactly five positive and three negative test cases for submission; developerName equal to your verified identity; displayName and shortDescription at most 30 characters; up to three defaultPrompts of at most 128.
• No commerce in the plugin: purchases stay on your website.

python3 scripts/check_chatgpt_package.py upsc-grade-chatgpt.zip           # or a folder
python3 scripts/check_chatgpt_package.py dist/chatgpt --server https://api.example.com

--server also fetches your authorization-server metadata and checks the ZIP's authorizationUrl and tokenUrl match it.

Claude

Claude has two listings that can point at the same server:

• The connector (claude.ai/directory/manage → Submit → MCP connector): typed into the portal. Generate its text (one-liner, description, use cases, test steps) from the listing source into a markdown file you copy from; see submit-to-claude.
• A plugin bundle (optional): a public GitHub repo with .claude-plugin/plugin.json, .mcp.json ({"mcpServers": {"<name>": {"type": "http", "url": "https://…/mcp"}}}), a skill, a README of at least 40 words that says what it does, how to use it and what data it sends, and a LICENSE. No symlinks, no secrets.

Words that fit both

• ChatGPT: shortDescription 30 characters ("Your exam second brain"), longDescription up to 4,000, capabilities up to 20 lines of 120.
• Claude: one-liner up to 200, but the directory card shows about 90 before cutting it; description up to 2,000.
• Claims must match what the tools do: reviewers compare them. Count things on real data (a "20,000 answers" claim that counts empty rows is wrong), and say where data comes from ("from the official key wherever one is published").
• Leave affiliation disclaimers and marketing superlatives out of the one-liner; keep facts.