Jetson Target Knowledge Base Generator
Scans a Jetson BSP and source tree to create a target-specific knowledge snapshot.
The skill explicitly limits discovery to read-only BSP/source inspection and forbids URL fetching. However, it overwrites the destination KB without confirmation and specifies no backup, rollback, or concurrency protection; it may also expose local document paths in the generated file, so points are deducted.
Inputs, bounded scans, missing-path behavior, and failure messages are described clearly. However, there is no executable implementation, dependency pinning, abnormal-input coverage, or end-to-end test suite. Under static calibration the score remains below 10.
Audience, invocation timing, inputs, output location, and non-fit cases are reasonably clear, with documented degradation when source or documents are absent. Chinese-language interaction and mainland-China environment fit are not addressed, and triggering depends on neighboring skills and profile conventions, so points are deducted.
The document has strong structure with overview, procedure, limitations, prerequisites, troubleshooting, references, and a version. It lacks dedicated Examples/Instructions sections, a changelog, and a clear maintenance/update path. License metadata is also inconsistent across the supplied materials: SKILL.md says Apache-2.0, repository metadata says NOASSERTION, and the skill card states dual licensing.
The stated goal is a directly readable per-target Markdown snapshot, with a defined output structure and confirmation summary. The supplied benchmark reports some positive-task accuracy, but it covers only two internal tasks and provides no statically verifiable representative output; generated results still require review, so the static score is conservatively limited.
The fixed revision, benchmark report, and evals.json make the intended behavior auditable. There is no independent third-party reproduction or committed test suite and real CI covering this skill's key paths, so the score is kept below 5.
- The skill overwrites target-platform/<profile-stem>.md without confirmation and provides no rollback mechanism; add backup or atomic-write safeguards before production use.
- Generated output records local document paths and existence status; verify that these paths cannot disclose sensitive information.
- License declarations are inconsistent; maintainers should clarify the applicable license for SKILL.md and its relationship to the repository license.
- The benchmark contains only two positive tasks and no negative tasks, so false triggering, boundary inputs, and mainland-China usability remain weakly validated.
What does this skill do, and when should you use it?
This skill supports NVIDIA Jetson Linux BSP workflows by generating a Markdown snapshot for the active target profile. It records BSP directory layout, source-tree structure, chip-family device-tree matches, nvpmodel variants, and referenced documents. The snapshot is stored beside the target profile and can be regenerated after the underlying data changes. It is useful for engineers who need a concise structural reference, but it does not edit profile YAML or perform a full recursive scan.
Resolves the active target profile and requires an existing bsp_image.root_path. It lists top-level BSP directories, checks canonical subtrees, verifies the active flash_config, and filters nvpmodel files by module ID and SKU. When source.root_path is available, it checks the specified source subtrees and lists matching device-tree files for the known chip family, capped at 30 entries. It checks whether local documents.* paths exist and records URLs verbatim without fetching them. It then writes or overwrites target-platform/<profile-stem>.md with a date-stamped snapshot and prints a short summary.
- A Jetson engineer has prepared a new BSP profile and needs a quick record of rootfs, bootloader, source, and related directories.
- A developer has updated the BSP source tree and wants to inspect the current kernel, out-of-tree driver, and device-tree layout.
- A flashing workflow needs confirmation that the active flash configuration and SKU-specific nvpmodel files are present.
- A team maintains several Jetson target profiles and wants a separate, reproducible reference snapshot for each one.
What are this skill's strengths and limitations?
- Produces one Markdown snapshot covering BSP layout, source layout, device trees, nvpmodel files, and document references.
- Designed for repeatable runs and overwrites the previous snapshot on each run.
- Uses constrained directory listings and targeted filters instead of recursively scanning a large BSP.
- Keeps document handling offline by recording URLs without fetching them.
- Requires a configured and existing bsp_image.root_path.
- Does not provide a complete recursive file index or read arbitrary file contents.
- The chip-family map covers only p3701, p3767, and p3834; other module IDs skip the chip-prefixed device-tree filter.
- The supplied source provides no single-skill test suite, platform coverage, or performance evidence.
How do you install this skill?
The README provides this collection-level installation command: npx skills add nvidia/skills. To install this skill directly, use npx skills add nvidia/skills --skill jetson-generate-kb --yes. The supplied source does not document other installation methods or version requirements.
How do you use this skill?
Prepare an active target profile with a bsp_image.root_path that points to an existing BSP directory; source.root_path and documents.* are optional. After the skill is loaded, ask the agent: “Generate the target knowledge base for the active Jetson profile.” It writes target-platform/<profile-stem>.md and refuses to generate the file when the BSP root is missing or does not exist.