TAO Skill Bank Capability Discovery
Query TAO Skill Bank workflows, models, platforms, and AutoML support from packaged manifests.
The skill limits declared tools to Read and Bash and mainly reads packaged manifests to produce capability information; no malware, credential theft, or covert exfiltration is shown. However, it executes scripts from an external path without documenting data flow, user confirmation, dependency integrity, rollback, or abnormal-output boundaries, so points are deducted.
SKILL.md provides consistent routing for general capability, model, and AutoML questions and defines an AutoML schema gate. However, it depends on an external tao-skill-bank path and supplies no bundled scripts, version pinning, abnormal-input handling, or diagnosable failure guidance. Under static calibration, evidence is insufficient for more than 7/10.
Trigger phrases and scenarios for general capabilities, model lists, and AutoML queries are reasonably clear, and the dependency is declared. Non-fit boundaries, input/output limits, platform differences, Chinese-language support, and reachability in mainland-China networks are not established, so points are deducted.
Front matter provides name, description, Apache-2.0 licensing, author, and version; skill-card.md adds use case, risks, and governance context, while repository files provide publication and security channels. However, SKILL.md lacks the recommended Instructions and Examples sections and offers little troubleshooting, changelog, or skill-specific maintenance ownership, so points are deducted.
The documented scripts are designated as the source of truth and cover application, data, platform, model-action, and AutoML discovery, plausibly completing the core task. However, the packaged dependency is not supplied, representative outputs and key-path completeness are not directly demonstrated, and static calibration limits the score to at most 7; 6 is awarded.
A pinned revision, evaluation file, and BENCHMARK.md are present, with reported results for security, correctness, discoverability, and effectiveness. However, coverage is limited to one evaluation task, there is no committed test suite for this skill, and the report is not independently reproducible from the supplied material, so only 4/10 is awarded.
- Execution depends on external TAO Skill Bank helper scripts; verify the path, script version, manifest completeness, and dependency provenance before deployment.
- Failure messaging, rollback behavior, and user-confirmation requirements are not documented.
- AutoML claims depend on a parseable train.schema.json; missing or invalid schemas should not be described as supported.
- The benchmark contains only one positive task and does not establish negative-trigger behavior, boundary coverage, or cross-environment stability.
What does this skill do, and when should you use it?
This skill answers what TAO Skill Bank can do, which models and workflows are available, and what supports AutoML. It runs packaged Python helper scripts that generate responses from application, data, model, AutoML, and platform manifests. General capability questions and model or AutoML questions use separate query paths. AutoML support depends on a valid packaged training schema for the model.
Runs scripts/list_tao_capabilities.py for general capability questions and uses its text output as the answer source. The reported coverage includes top-level workflows under applications/ and data/, supported execution platforms, and available model actions such as training, evaluation, inference, export, and TensorRT engine generation when present in the packaged schema manifest. Runs scripts/list_tao_models.py --scope all for model and action queries. Runs the same script with --scope automl, or scripts/list_automl_support.py, for AutoML queries; a model is not described as AutoML-supported when its skills/models/<network>/schemas/train.schema.json is missing or invalid JSON.
- A TAO user wants an inventory of the Skill Bank's application and data workflows. Use the general capability query.
- A developer needs to verify which actions a TAO model supports, such as training, evaluation, inference, or export. Use the model-list query.
- A team is assessing AutoML coverage and needs to identify models with valid packaged training schemas. Use AutoML mode.
- A maintainer needs manifest-based answers instead of manually scanning model directories. Use this skill's packaged helper scripts.
What are this skill's strengths and limitations?
- Uses packaged application, data, model, platform, and AutoML manifests as the source of truth.
- Separates general capability, model-action, and AutoML query paths.
- Defines a clear exclusion rule for models with missing or invalid training schemas.
- The standalone SKILL.md declares Apache-2.0 licensing and version 0.1.0.
- Requires the packaged TAO Skill Bank helper scripts; the SKILL.md alone cannot perform the queries.
- The source provides no actual helper-script output examples, test results, or platform inventory.
- This skill discovers capabilities rather than executing TAO training, deployment, or inference workflows.
- The source does not document cost, performance, or external-service availability.
How do you install this skill?
Install the skill from NVIDIA's collection with: npx skills add nvidia/skills --skill tao-list-capabilities --yes. To target Codex: npx skills add nvidia/skills --skill tao-list-capabilities --agent codex --yes. The source does not document separate Python dependency installation.
How do you use this skill?
Ask the agent questions such as “what can TAO Skill Bank do”, “list TAO models”, “which TAO workflows are available”, or “what supports AutoML”. For general capabilities, run ${TAO_SKILL_BANK_PATH:-~/tao-skills-external}/scripts/list_tao_capabilities.py --skill-bank ${TAO_SKILL_BANK_PATH:-~/tao-skills-external} --format text. For models, use list_tao_models.py --scope all --format text. For AutoML, use list_tao_models.py --scope automl --format text.