Build from a brief

You can ask a coding assistant to build an ATFGraph application from a written brief. The bundled coding-assistant skills explain how to write, test and deploy Python nodes. These Markdown instructions are different from runtime skills: the @app.skill() functions your application exposes for traversals to call.

Four bundled skills support vNext: atfgraph-vnext builds a project, atfgraph-vnext-use calls registered nodes, atfgraph-vnext-personal installs a local node, and atfgraph-vnext-topology designs a graph without building it. The installer installs all four. It leaves existing directories alone unless you pass --replace, and keeps any existing skill named atfgraph separate.

Install for Claude Code

Run these commands from the repository root. An installed client carries the same skill files under share/atfgraph/resources/skills.

python3 scripts/install-skills.py --harness claude
export SDK_ROOT="$PWD"                              # this checkout
export SERVER_ROOT="$PWD"                           # the same checkout holds the server, if you run one locally

Start Claude Code in your application's directory and ask it to use atfgraph-vnext. These commands are for a first installation; inspect an existing destination before replacing it.

Install for Codex

Run the Codex installer from the same repository directory:

python3 scripts/install-skills.py --harness codex
export SDK_ROOT="$PWD"
export SERVER_ROOT="$PWD"

Open a new Codex session in your application's directory and invoke $atfgraph-vnext. The assistant follows the skill's instructions to write the application, test it and prepare deployment.

Give it a brief

Build an AI travel consultant that understands a brief, clarifies missing details, and creates a travel plan with flights, accommodation and activities. Keep the traveller's brief, saved offers and latest proposal in session memory. Explain the recommendations and answer follow-ups about fares and alternative hotels from saved evidence. Refresh searches only when needed or explicitly requested. Include offline tests, Python client and CLI examples, and prepare Docker Compose deployment.

Expect a project you can inspect, tests the assistant has actually run, and a demonstrated run with a recorded path. Tests with fixture data establish offline behavior. Live results also require working provider configuration, and a deployed application needs healthy workers.

Extend the application with the Python guide. Use the CLI and deployment guides to install, call and host it.