How to use the generate-manifest ADK Recipes agent skill

What the generate-manifest skill makes the agent do

The skill is written for recipes under core/, contrib/ or plugins/ in the adk-recipes repository. Given a recipe path such as core/python/rag-agent-search, the agent reads the live schema at .github/schemas/manifest-schema.json rather than relying on memory.

It then scans the recipe's code and README files and infers only what the code supports. Team and owner fields are left as placeholders for a human to fill in.

• Reads pyproject.toml or requirements.txt to set the language field
• Checks for a runnable entry point, such as an if __name__ == "__main__" block or an adk web or adk run invocation, to decide between standalone and module
• Counts Agent( constructor calls to choose single or multi agent
• Drafts a description of at most 15 words from README.md and AGENTS.md

How to install the generate-manifest skill

Download the skill ZIP from this page. You can install it for a single project or for your user account.

For a project, unzip the file into the .claude/skills/ folder inside that project. For all your projects, unzip it into ~/.claude/skills/. In the Claude apps, you can instead upload the ZIP under Settings → Capabilities → Skills.

• Project scope: the project's .claude/skills/ folder
• User scope: ~/.claude/skills/
• Claude apps: Settings → Capabilities → Skills, then upload the ZIP

How to trigger the skill with a natural request

You do not need a special command. Ask for a manifest for a recipe and give the path relative to the repo root. The skill's description covers requests to create or generate a manifest.yaml for a recipe under core/, contrib/ or plugins/.

If you leave out the path, the skill tells the agent to ask you for it before it does anything else.

• Generate a manifest.yaml for core/python/rag-agent-search
• Create a manifest for the new recipe at contrib/python/my-recipe
• Draft a manifest.yaml for plugins/retail/store-ops and check it against the schema

Review the agent's inferences before you commit

The skill asks the agent to state what it found for each inferred field, such as type, deployable and the architecture block, and to cite the file or line behind each conclusion. Read those citations first, because a wrong guess lands in the file otherwise.

The owner fields are intentionally invalid until a person replaces them. The ownership.team and ownership.poc values are fixed placeholder strings, and the agent is told never to invent a team name or GitHub ID. Dependencies and tags are commented out for the author to decide.

• Check type: standalone needs a real entry point; module is used only when there is no way to run the recipe alone
• Check that language matches the file types in the recipe
• Check the description and rewrite it if needed; it keeps a TODO comment
• Replace the team and poc placeholders yourself
• Add license only if the recipe explicitly declares one

Validate the manifest from the repo root

After writing manifest.yaml, the skill runs the validator from the repo root. Do not change into tools/ first, and do not pass --active. Use a path relative to the repo root, not an absolute path.

Two kinds of failure matter. Schema errors such as a missing required field, an invalid enum value, a description under 10 characters or malformed YAML must be fixed. Placeholder errors on ownership.team and ownership.poc are expected until a human fills them in.

Safety notes and requirements

The skill runs no scripts and uses no network access, according to its listing. It works by reading and writing files in the repository.

It depends on fixed paths in the repo. It needs .github/schemas/manifest-schema.json and .github/policy.yml, and an agent runtime that can read and write files in the repo. The policy file holds the recipe_size_limits used when deciding whether to set large.

• Runs scripts: no
• Uses network: no
• Needs access to the adk-recipes repository files
• Needs an agent that can read and write files in the repo

Tips for better results

Give the most specific path you can, and point the agent at a recipe that already has a README or AGENTS.md, because those files drive the description draft. Compare the output with existing manifests in core/ and contrib/, which the skill says the output should resemble in style.

Expect to edit the output. The skill's own guidance is that its inferences need human review.

• Keep the recipe's README.md and AGENTS.md accurate, since they feed the description
• Ask for one recipe at a time
• Treat the large and deployable fields as opt-in and omit them unless the criteria clearly apply