Skip to main content
An App Definition is a document with a published schema and an offline checker. That makes it an unusually good target for an AI agent: it can write, check its own work, and be told precisely what it got wrong, all without a database, a model API or a deploy. This section has a tuned starting prompt for each of the common agents. They differ, because the tools differ. Read this page first, because the reasoning behind all of them is the same.

​
Your workspace already briefs the agent

That writes CLAUDE.md and AGENTS.md into the workspace root, with the same generated text in both. Two conventions want two filenames, and a workspace that taught two agents different rules would produce source only one of them could maintain.
This is the single most important thing on the page. The instructions are already there. Your prompt does not need to re-explain Cordango, and it should not try.
What those files teach is the workflow, not the vocabulary: how to find out, rather than a copy of the schema. That is deliberate, and it is what stops an agent’s context filling up with 212 KB of JSON Schema.

​
The rules they install

Every prompt in this section leans on these, so they are worth reading once. Discover, do not guess. cordango inspect for what exists, cordango vocabulary for what may be written. Both answer from the installed build, so they cannot drift from what the CLI will accept. Never read the schema, never read the binary. If cordango vocabulary cannot answer a question, that is a gap worth reporting rather than working around. Check constantly. cordango check calls no model and touches no database. It is free. Run it after every change. One aggregate per change. It keeps diffs reviewable, and it is what --scope enforces. Stable keys. Renaming an entity or field key is not supported, because nothing rewrites the references to it. Add and remove deliberately instead. The agent does not accept its own work. cordango apply writes to the working tree. A person reads git diff and stages it.

​
Every command speaks JSON

--json works on all of them. cordango help --json in particular is how an agent discovers the command surface without a human pasting a README into its context.

​
The shape of a good prompt

Short, and about the job. The workspace supplies the rest.
Note what is absent. No explanation of what Cordango is, no YAML example, no schema excerpt, no list of field types. All of that is one cordango vocabulary field away, and an agent that looks it up gets the version that matches the installed CLI.

​
Pick your agent

Claude Code

Reads CLAUDE.md automatically. The path of least resistance.

Codex

Reads AGENTS.md. Same workspace, same instructions.

Cursor

Needs a rules file pointing at the workspace instructions.

GLM

Strongest with explicit, sequenced instructions.