Extension Authoring Skill
Use the Flow Steward Extension Authoring skill when asking a coding agent to build or review an extension. It guides the agent through the target application's schemas, provider runtime, tests, and packaging. It is separate from the skills registered inside Flow Steward.
Install the Flow Steward Agent Plugin to get this developer skill
and the workflow/operations skill together. For an optional standalone installation,
download the individual skill ZIP, or
read SKILL.md before installing it.
The ZIP contains the canonical flow-steward-extension-authoring/ instructions.
The individual download copies the same canonical leaf as the plugin and requires
inspection of the actual target release's contracts. It does not contain credentials,
install dependencies or deploy an extension automatically.
Before you start
Have a compatible Flow Steward application source checkout, a working Development authoring environment, and Docker available for its tooling. A Compact production container or this documentation repository is not the extension development toolchain. Follow Local Development and the target checkout's instructions.
Decide which provider to integrate, what the extension should do, and which Flow Steward version it must support. Start with mocked provider responses; keep live credentials out of prompts and committed fixtures. The Example Showcase extension is a good reference to point the agent at: it implements every extension feature in one bundle.
The skill is also how you add custom logic to workflows: Flow Steward has no inline code step, so ask the agent for an extension whose action you then call with the Extension tools step (see Custom code).
Install in your coding agent
For the complete plugin, follow Install and verify
and use the client’s discovered skill names. Claude plugin installs expose
/flow-steward:flow-steward-extension-authoring. For the individual ZIP, unzip
the download. From the project directory your agent will use, create the
appropriate skills directory and copy flow-steward-extension-authoring/ into it:
| Agent | Project-local destination | Invoke |
|---|---|---|
| Claude Code | .claude/skills/flow-steward-extension-authoring/SKILL.md | /flow-steward-extension-authoring followed by the task |
| Codex | .agents/skills/flow-steward-extension-authoring/SKILL.md | $flow-steward-extension-authoring followed by the task |
| Cursor | .cursor/skills/flow-steward-extension-authoring/SKILL.md | Ask the agent to use flow-steward-extension-authoring for the task |
For example, after extracting to your Downloads directory:
# Run in the project where Codex will work.
mkdir -p .agents/skills
cp -R ~/Downloads/flow-steward-extension-authoring .agents/skills/
For Claude Code or Cursor, substitute the destination directory from the table.
If an older copy exists, review and replace that copy rather than creating a
nested flow-steward-extension-authoring/flow-steward-extension-authoring/ directory. Keep one
active version of the skill in the agent's search paths. Start a new agent
session if the installed skill is not discovered.
These are local coding-agent installation instructions. Cloud sessions must have access to the skill and application checkout in their own environment; a folder on your laptop is not automatically available there. See the official Claude Code skills documentation, Codex skills documentation, and Cursor skills documentation for discovery and account-level installation behavior.
Give the agent a concrete task
For Codex, start with a request such as:
Use $flow-steward-extension-authoring to create com.acme.crm-sync for the Flow Steward
checkout in this workspace. Target the checked-out release. Implement lookup
of CRM records and an authenticated incoming webhook that starts a workflow.
Use mocked provider responses first. Explain each configuration field and
where an operator obtains its value. Validate and package for Linux AMD64
and ARM64, including compatible hash-locked third-party wheels when needed.
Report the Python/ABI target, archives and checks actually run on each
architecture, plus missing verification. Do not deploy or publish it.
Use Claude Code's slash invocation or the same natural-language request in Cursor. Replace the example provider, operations, authentication requirements, and target release with your own requirements. The skill also supports a review request: identify your existing bundle and ask for contract validation, missing tests, and operator-instruction gaps.
What you should receive
- An extension bundle with its manifests, runtime, referenced files, and focused tests; portable third-party wheels cover both Linux AMD64 and ARM64 for the target tooling Python/ABI. Native executables need separately targeted archives.
- Configuration instructions, required provider permissions, a first successful operation, and troubleshooting guidance.
- The target Flow Steward revision, Python/ABI and architecture targets, archive paths/digests, commands actually run, and results.
- An actual verification report identifying which architectures were tested and what remains unverified. Packaging/wheel coverage alone does not prove successful installation on both architectures or a live provider integration.
Check the delivered wheels/ folder yourself: every requirement needs an
_x86_64 and an _aarch64 wheel, or one py3-none-any wheel.
extensions validate passes a bundle that has only one architecture; only
lock-deps and bundle --with-wheels run for both. Follow
Python Dependencies And Wheel Distribution
for locks, platform defaults and missing-wheel handling, and
Bundle And Install for
separate native runtime targets.
The skill does not grant the agent access to a provider account, approve an installation, or guarantee that generated code is correct. Review the output and run the target's validation before installing a package. For manual authoring, continue with Build An Extension; for runtime and schema details, see Extensions.
Updating the skill
Replace the installed folder with a reviewed new download when adopting a new Flow Steward release. Keep local changes under version control. The agent must still compare contracts with the target checkout: a skill update and an application release are independent artifacts. For a full plugin installation, follow the plugin update steps, refresh the client, and verify both discovered skills against the reviewed package.