DerivaML Skills para Claude Code
informatics-isi-edu/deriva-ml-skills
30 skills para ML workflows completos em catálogos Deriva
Ver no GitHubPlugin de skills para Claude Code que oferece 30 comandos para orquestração completa de workflows de ML sobre catálogos Deriva. Cobre ciclo de vida de datasets, execuções, experimentos, gestão de features, configurações Hydra-zen e troubleshooting. Requer deriva-ml-mcp server e plugin deriva-skills.
Instalação
npx convez-ai informatics-isi-edu/deriva-ml-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Orquestração de ciclo de vida de datasets e experimentos ML
- Configuração e validação de projetos com Hydra-zen
- Troubleshooting de execuções e gestão de assets ML
- Integração com catálogos Deriva para rastreabilidade
Veredito

Solução robusta e bem-arquitetada para equipes que usam Deriva e precisam automatizar workflows ML em Claude Code. Exige setup prévio (deriva-ml-mcp server + plugins) e conhecimento técnico sólido; recompensa com cobertura completa do ciclo de vida ML.
README
A tradução para o português deste README ainda está na fila.
DerivaML Skills Plugin
Claude Code skills plugin for DerivaML ML workflows. Provides 30 skills covering the full ML development cycle on top of Deriva catalogs: dataset / execution / experiment lifecycles, features, asset management, Hydra-zen configs, model development, project setup validation, and execution-specific troubleshooting.
The plugin requires the deriva-skills plugin (for generic Deriva catalog operations the DerivaML skills cross-reference) and a deriva-mcp-core server with the deriva-ml-mcp plugin loaded (for the deriva_ml_* MCP tools). The install procedure below brings in both plugins; the MCP server you set up separately.
Installation
Install via the deriva-plugins marketplace:
# Add the marketplace (one-time) — covers both deriva and deriva-ml
/plugin marketplace add informatics-isi-edu/deriva-plugins
# Install both plugins — deriva-ml assumes deriva is loaded for cross-references
/plugin install deriva
/plugin install deriva-ml
You also need a Deriva MCP server with the deriva-ml-mcp plugin loaded. See the deriva-mcp-core deployment docs.
Migrating from the old per-repo marketplaces? Earlier versions were installed via
/plugin marketplace add informatics-isi-edu/deriva-skillsand/plugin marketplace add informatics-isi-edu/deriva-ml-skills. Both single-plugin marketplaces have been retired. To migrate, first remove both old caches:/plugin marketplace remove deriva-plugins # was the deriva-skills marketplace /plugin marketplace remove deriva-ml-plugins # was the deriva-ml-skills marketplaceNote the name collision: the new unified marketplace is also internally named
deriva-plugins, so the oldderiva-pluginscache must be removed before re-adding from the new repo. Then run the three commands above.
Updating
Enable "autoUpdate": true in ~/.claude/settings.json for the deriva-plugins marketplace and restart Claude Code; new versions will be picked up automatically.
For checking what's currently installed and walking through updates, the versioning content lives in two troubleshooting skills:
/deriva:troubleshoot-deriva-errors(deriva-skills) — "Versioning and updates" section covers the foundation: deriva-py, deriva-mcp-core, the deriva plugin/deriva-ml:troubleshoot-execution(this plugin) — "Versioning and updates" section covers the DerivaML layer: deriva-ml, deriva-ml-mcp, deriva-ml-skills
Check the foundation first; the DerivaML stack depends on it.
Available Skills
The plugin ships two layers of skills with different invocation models. User commands are tools you invoke directly (or that fire when you ask Claude something matching them); auto-invoked guides are background disciplines that watch what you're doing and inject framing before mistakes happen.
User commands
Invoke directly with /deriva-ml:<skill-name>, or by asking Claude something that maps to one of them.
| Category | Command | Description |
|---|---|---|
| Lifecycle (auto-fires too) | /deriva-ml:dataset-lifecycle | Create, populate, split, version, browse, download datasets |
/deriva-ml:execution-lifecycle | Pre-flight validation, run experiments, execution provenance, state machine | |
/deriva-ml:experiment-lifecycle | The seven-phase experiment cycle: hypothesis → configure → identify assets → run → update assets → evaluate → repeat | |
| Datasets | /deriva-ml:debug-bag-contents | Diagnose missing data in dataset bag exports |
| Features | /deriva-ml:create-feature | Create features with vocabulary or scalar value domains; add per-row values |
/deriva-ml:compare-model-runs | Rank/compare metrics across model training executions | |
| Assets | /deriva-ml:work-with-assets | File assets — upload, download, provenance, types |
/deriva-ml:manage-deriva-storage | Storage Manager (a built-in Deriva-ML-Apps app) | |
| Schema / curation | /deriva-ml:schema-evolution-impact | Impact analysis before schema or data changes — which datasets, features, executions reference the target |
| Experiments / configs | /deriva-ml:configure-experiment | Set up DerivaML experiment project structure |
/deriva-ml:write-hydra-config | Write and validate hydra-zen config files | |
| Models | /deriva-ml:new-model | Scaffold a new model function and wire it into configs/workflows |
/deriva-ml:model-development-workflow | End-to-end model development workflow | |
| Notebooks | /deriva-ml:setup-notebook-environment | Set up Jupyter environment for DerivaML |
/deriva-ml:run-notebook | Develop and run notebooks with execution tracking | |
| Project setup | /deriva-ml:setup-derivaml-project | Bootstrap a new DerivaML project: repo init, pyproject.toml template, gh install, coding conventions |
/deriva-ml:validate-project-setup | Validate the project conforms to the deriva-ml-model-template shape | |
| Apps + visualization | /deriva-ml:create-web-app | Build and register custom web apps for DerivaML data |
/deriva-ml:browse-erd | Launch the ERD browser for the connected catalog | |
| Help / orientation | /deriva-ml:help | General orientation: what is DerivaML, how to use it, where to start |
| Troubleshooting | /deriva-ml:troubleshoot-execution | Execution-lifecycle errors AND DerivaML versioning/updates |
Auto-invoked guides (background disciplines, not commands)
These are skills that Claude loads on its own when the situation calls for them. They do not appear in the /deriva-ml: slash-command picker, and you should not type them as commands — they're internal disciplines that "look over the ML developer's shoulder" to inject the right framing before mistakes happen, capture decisions as they're made, and route catalog mutations through the right provenance-preserving paths.
| Skill | When Claude loads it |
|---|---|
deriva-ml-context | Always — establishes the DerivaML domain frame, the five abstractions, and the inheritance-with-override rule that governs when to use a deriva-ml surface vs the underlying deriva surface |
dataset-lifecycle | When the user is working with datasets (creating, splitting, versioning, browsing) — guides through the lifecycle phases proactively |
execution-lifecycle | When the user is running experiments — guides through the state machine and upload discipline |
experiment-lifecycle | When the user is designing or running an experiment cycle — names the seven phases and the cross-step disciplines |
create-feature | When the user is creating or working with features — guides through the assess/design/create phases and the feature-vs-column decision |
run-notebook | When the user is creating, developing, or running a DerivaML Jupyter notebook — guides through the three-stage development cycle and run_notebook() provenance machinery (also slash-typable) |
capture-tacit-knowledge | After significant decisions in any phase — captures rationale to tacit-knowledge.md |
generate-descriptions | When creating any DerivaML entity without a user-supplied description |
You'll see the auto-invoked skills' effects in Claude's behavior (it asks "should we name a hypothesis first?" when starting an experiment; it captures the decision to tacit-knowledge.md; it walks the dataset-creation phases in order); you won't see them as commands you can invoke.
Which plugin's skill to reach for
The two plugins are designed to be used together. As a rough guide:
deriva-skills covers generic Deriva catalog operations — what works on any Deriva catalog, ML or otherwise:
- Schema and table operations
- Generic vocabulary CRUD (e.g., your project's
Sample_TypeorTissue_Type) - Querying / browsing catalog data
- Chaise display annotations — interactive MCP-tool path (
/deriva:customize-display). The type-safe Python builder path (/deriva-ml:use-annotation-builders) lives in this plugin because it requires thederiva-mlPython package. - Generic catalog troubleshooting (auth, RIDs, missing records)
- Loading row data and uploading assets via
deriva-upload-cli
deriva-ml-skills (this plugin) covers the DerivaML domain layer — the abstractions DerivaML adds on top of a catalog:
- The five DerivaML abstractions: Datasets, Workflows, Executions, Features, Assets
- The four DerivaML built-in vocabularies:
Dataset_Type,Workflow_Type,Asset_Type,Execution_Status_Type - Experiment configuration (Hydra-zen), model development, training workflows
- Execution-state-machine debugging
Override rule. When you're working in a deriva-ml catalog, everything that applies in a Deriva catalog applies here too — by default reach for
deriva-skills. But if a deriva-ml surface exists for the operation — a/deriva-ml:<skill>, aderiva_ml_*MCP tool/prompt/resource, or a deriva-ml Python object — prefer it over the equivalent deriva surface (/deriva:<skill>,deriva-mcp-coretool, orderiva-pycall). The five abstractions above are where the override mostly lands; everywhere else the deriva default applies. The auto-invokedderiva-ml-contextskill carries the rule plugin-wide; ADR-0001 records the design decision.
Development
Load the plugin from a local path without installing:
claude --plugin-dir /path/to/deriva-ml-skills
Related Projects
deriva-skills— Companion plugin for generic Deriva catalog operations (required — install alongside this one)deriva-mcp-core— Core MCP framework + generic Deriva catalog toolsderiva-ml-mcp— DerivaML MCP plugin (loaded by deriva-mcp-core); required for these skills' MCP toolsderiva-ml— Core Python library for ML workflows on Derivaderiva-py— Python SDK for Deriva scientific data management
License
Apache 2.0