Globus SDK Skill

A Claude Code / Agent Skill that teaches Claude how to write correct, modern Python against the Globus platform — Transfer, Flows, Compute, Search, and Auth.

It steers the model toward the current globus-sdk v4+ APIs (the GlobusApp / UserApp / ClientApp auth model) and away from deprecated patterns like fair_research_login, globus-automate-client, funcx, and hand-rolled NativeAppAuthClient OAuth flows.

Why

LLMs frequently emit outdated Globus code, because much of the public training data predates the v4 SDK and the 2024 deprecations. This skill loads up-to-date, opinionated reference material on demand so generated code uses the supported interfaces.

New to the project? Start here

If you just joined and want this working in a few minutes:

1. Install Claude Code (or another Agent Skills–compatible client).

2. Install the skill — copy the source folder into your personal skills directory:

   git clone https://github.com/ryanchard/globus-skill.git
   cp -r globus-skill/globus-sdk ~/.claude/skills/

   You should now have ~/.claude/skills/globus-sdk/SKILL.md.

3. Try it — start a new Claude Code session and ask:

   "Write a Python script that transfers a file between two Globus collections."

   Claude will automatically load this skill and produce modern globus-sdk v4 code. You're done.

To actually run the generated code you'll also need a Globus account and a registered Client ID — see Registering a Client ID in globus-sdk/SKILL.md. The skill itself needs none of that; it's just guidance Claude reads.

You only need the build step below if you edit the skill and want to share a new .skill bundle.

What it covers

Service	Reference	Highlights
Auth	references/auth.md	UserApp / ClientApp, scopes, consent handling, GlobusAppConfig
Transfer	references/transfer.md	TransferData, DeleteData, filter rules, task monitoring, ConsentRequired
Flows	references/flows.md	FlowsClient / SpecificFlowClient, flow definitions, input schemas, reuse
Compute	references/compute.md	Executor, writing/registering functions, batch & ShellFunction
Search	references/search.md	SearchClient, indices, ingest, query, use within Flows

The top-level SKILL.md holds the trigger conditions, the deprecated-pattern warnings, and quick-start auth snippets; Claude reads the relevant reference file(s) before writing code for a given service.

Install

A skill is just a directory containing a SKILL.md. Drop it into your skills directory and Claude Code picks it up.

Option A — install the prebuilt bundle

# .skill files are zip archives; extract into your personal skills dir
unzip dist/globus-sdk.skill -d ~/.claude/skills/

Option B — install from source

git clone https://github.com/ryanchard/globus-skill.git
cp -r globus-skill/globus-sdk ~/.claude/skills/

Either way you should end up with ~/.claude/skills/globus-sdk/SKILL.md. Restart Claude Code (or start a new session) and ask it to write some Globus code — it will trigger automatically when you mention Globus, globus-sdk, Transfer/Flows/Compute/Search, or research data movement between endpoints.

For a project-local install instead, copy globus-sdk/ into .claude/skills/ inside your project.

Repo layout

globus-skill/
├── globus-sdk/              # editable skill source (the canonical copy)
│   ├── SKILL.md
│   └── references/
│       ├── auth.md
│       ├── transfer.md
│       ├── flows.md
│       ├── compute.md
│       └── search.md
├── dist/
│   └── globus-sdk.skill     # prebuilt bundle, regenerated by the build script
└── scripts/
    └── build-skill.sh       # packages globus-sdk/ -> dist/globus-sdk.skill

Develop & rebuild

The skill is the globus-sdk/ folder — editing it is just editing Markdown. Building the .skill bundle is nothing more than zipping that folder.

Edit the Markdown under globus-sdk/, then repackage:

./scripts/build-skill.sh

That script is a thin wrapper around a single zip command — equivalent to:

zip -r dist/globus-sdk.skill globus-sdk

The result, dist/globus-sdk.skill, is a zip archive whose top-level entry is globus-sdk/. Anyone can install it by unzipping into ~/.claude/skills/. After editing, commit both the source change and the regenerated bundle.

Writing tips

• Keep SKILL.md's description rich in trigger terms — it's how Claude decides to load the skill.
• Put detailed, scannable, code-first guidance in the per-service reference files; keep SKILL.md lean.
• Prefer showing the supported pattern over merely naming the deprecated one.

License

MIT