agentsclimarketplace

Jetson link docs

Skill NVIDIA/skills/skills/jetson-link-docs

Agent Skills for NVIDIA products — install into Claude Code, Codex, and other coding agents to run Physical AI, robotics, simulation, CUDA, and RAG workflows end to end.

Install
npx -y skills add NVIDIA/skills --skill jetson-link-docs

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

What its author says it does

Copied from the file, not written here

Bind pre-downloaded Jetson reference docs (developer guide, design guide, pinmux, schematics) into the active profile documents block. Use after staging docs on disk; not for downloading.

The file declares its own license as Apache-2.0. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

13.1 KB, as published. Nobody here has run it

jetson-link-docs

Overview

This skill writes the documents: block of the active Jetson / IGX target-platform profile YAML so downstream skills (/jetson-generate-kb, /jetson-customize-pinmux, camera / pcie / uphy, etc.) can resolve doc paths by name. It walks the user through every document slot in the profile schema, tries to auto-bind each slot to a file under <documents.root_path>/ via case-insensitive glob matching, and writes the resulting paths back into the active profile.

Scope is registering pointers only — this skill does not fetch or download. The files must already exist on disk under <documents.root_path>/.

When to invoke

  • After /jetson-init-target finishes and the user has documents on disk to register.
  • The user wants to add, change, or remove document references on an existing profile.
  • A downstream skill (e.g. jetson-generate-kb) reports "no documents recorded" and the user wants to fix that.

Procedure

Resolve the active target

Resolve the active profile + <workspace> per the contract in ../../context/target-platform-contract.md. Cache the loaded profile in memory — this skill mutates it in the "Write the documents: block back to the profile" step.

Load the document-slot schema

Load ../../references/platform_template.yaml. Parse the documents: block. Each per-document field is marked <OPTIONAL: description>. Use the marker description as prompt text verbatim. Match markers with the regex ^<(REQUIRED|OPTIONAL|DERIVED):\s*(.*)>$ after YAML parsing strips surrounding quotes.

Skip custom_carrier_schematic and custom_carrier_pinmux_xls entirely when the active profile has no custom_carrier: block — both are meaningless without one. This filter applies through the "Scan and auto-match" and "Manual prompts for unmatched fields" steps.

Resolve documents.root_path

Default: <workspace>/Documents. If the profile already records documents.root_path, use it. Otherwise, if <workspace>/Documents/ exists, use it (the field is omitted from the written profile — downstream skills fall back to the workspace default). If neither is available, prompt the user for an absolute path, or accept Enter / cancel to skip the auto-scan. A user-provided path that doesn't exist is treated as skipped (warn, don't refuse — the field is OPTIONAL); manual prompts in the "Manual prompts for unmatched fields" step still run.

Resolve the product token

Read the Product Token column from ../../references/bsp-platforms-catalogue.md for the row matching reference_devkit.name. The token is a case-insensitive glob fragment (e.g. *orin*nano*, *agx*thor*) consumed by the fallback patterns in the "Scan and auto-match" step.

If reference_devkit.name has no row in the catalogue, log a warning and proceed without a product-token fallback — the "Scan and auto-match" step still works with strictly SKU-keyed matching.

For custom carriers, derive <custom-token> from custom_carrier.name using this recipe: lowercase, replace each space with *, wrap in * on both ends. E.g. "Acme Vision X1" → *acme*vision*x1*.

Scan and auto-match

Skip this step entirely if documents.root_path did not resolve in the "Resolve documents.root_path" step (no scan target → no auto-suggest; fall through to manual prompts in the "Manual prompts for unmatched fields" step).

Scan the directory once (one level deep) and try to auto-match each remaining <OPTIONAL:…> field using the case-insensitive globs below. Use the lower-case module.id / carrier.id / custom_carrier.id strings from the profile in the SKU column.

FieldSKU glob (primary)Product-token glob (fallback)
bsp_developer_guide*developer*guide*.pdf, *BSP*guide*.pdf(no fallback — pattern is product-agnostic)
soc_tech_ref_manual*TRM*.pdf, *tech*ref*manual*.pdf(no fallback — same)
module_data_sheet*<module.id>*data*sheet*.pdf, *<module.id>*datasheet*.pdf<token>data*sheet*.pdf, <token>datasheet*.pdf
module_design_guide*<module.id>*design*guide*.pdf, *<module.id>*PDG*.pdf<token>design*guide*.pdf, <token>PDG*.pdf
module_thermal_design_guide*<module.id>*thermal*.pdf (covers "Thermal Design Guide" / "TDG")<token>thermal*.pdf
module_schematic*<module.id>*schem*.pdf<token>schem*.pdf
carrier_board_spec*<carrier.id>*board*spec*.pdf, *<carrier.id>*spec*.pdf<token>carrier*spec*.pdf
carrier_schematic*<carrier.id>*schem*.pdf<token>carrier*schem*.pdf
custom_carrier_schematic*<custom_carrier.id>*schem*.pdf (only if custom carrier)<custom-token>schem*.pdf (only if custom carrier)
ref_devkit_pinmux_xls*<carrier.id>*pinmux*.xls* (matches .xls, .xlsx, .xlsm)<token>pinmux*.xls*
custom_carrier_pinmux_xls*<custom_carrier.id>*pinmux*.xls* (only if custom carrier)<custom-token>pinmux*.xls* (only if custom carrier)

<token> is the catalogue-resolved product token; <custom-token> is derived from custom_carrier.name per the "Resolve the product token" step. Tokens already include leading/trailing *, so the table does not repeat them.

Match policy per field

For each field that has auto-match results:

  • Take the union of hits across the SKU glob and the product- token glob, then deduplicate by absolute path — a file matched by both globs counts once.
  • Exactly 1 unique hit → show the path and prompt use this? (yes/no, default yes). On yes, record it and skip the manual prompt for that field. On no, fall through to the manual prompt in the "Manual prompts for unmatched fields" step.
  • 0 hits → skip auto-suggest entirely for that field; fall through to the "Manual prompts for unmatched fields" step.
  • 2+ unique hits → present them as a numbered list in the "Manual prompts for unmatched fields" step so the user can pick by number rather than typing a path; include a skip / NA option. Never silently bind a multi-hit candidate.
  • Never silently bind without user confirmation — wrong-schematic / wrong-pinmux bindings are real and costly.

If documents.root_path is folder-organised one level deeper than flat (NVIDIA archives often are: Schematics/, Design-Guides/, Pinmux/, etc.), the file globs may return zero hits even when the right documents exist. v0.2 only scans one level deep — when 0 hits is suspicious (documents.root_path exists but no fields auto-bound), surface the limitation to the user and offer to fall through to manual prompts.

Manual prompts for unmatched fields

For every field that wasn't auto-bound (and wasn't filtered out in the "Load the document-slot schema" step), prompt using the marker description from the "Load the document-slot schema" step as prompt text, in document order. Accept Enter and NA interchangeably as "skip this field". When the "Match policy per field" step produced 2+ candidate hits for a field, present them as a numbered list with a skip / NA option rather than asking for a free-text path.

Validate that user-provided paths exist on disk (warn if not, but do not refuse — the user may be recording a planned path). URLs (values starting with http://, https://, or ftp://) are accepted verbatim and not validated.

Write the documents: block back to the profile

Edit target-platform/<active>.yaml in place. Preserve all other top-level blocks (reference_devkit:, custom_carrier:, bsp_image:, source:) and their comments verbatim. Write only the fields the user provided — omit skipped / NA fields entirely (no NA placeholders, no empty keys).

Edge behavior: when every field was skipped (including documents.root_path), drop the documents: block entirely from the profile — never write documents: {} or a block of NA values. When only documents.root_path was provided (no per-document binding), record it alone — the path has value as a hint for future re-runs. On re-run with an existing documents: block, merge: existing bindings are preserved unless the user picks a new file or NA; newly bound fields are added.

Confirm

Print a summary:

  • Profile path written.
  • documents.root_path — resolved value (or "default — omitted").
  • Auto-bound fields: count + per-field one-line list.
  • Manually entered fields: count + list.
  • Skipped fields: count.
  • A reminder that jetson-generate-kb re-reads documents.* and should be re-run if a KB exists.

If a downstream skill triggered this run, tell the user to re-issue their original request; do not silently re-trigger it.

Gotchas

  • Active profile must exist. This skill writes back to whichever profile is active. If no profile is active, refuse and route to jetson-set-target / jetson-init-target.
  • Product-token globs are intentionally broad. A token like *orin*nano* matches both Orin-Nano-specific docs and combined Orin-NX/Nano docs (e.g. Jetson-Orin-NX-Nano-Design-Guide_…). That is usually correct for module-side docs (NVIDIA ships combined manuals), but verify on schematic / pinmux / spec fields where wrong-product binding is costly.
  • Update bsp-platforms-catalogue.md when adding new product rows. The Product Token column is consumed by the "Resolve the product token" step; a missing token degrades the auto-scan to SKU-only matching (the skill warns and continues, but doc-rich documents.root_path scans will degrade silently from "5 auto-binds" to "fewer auto- binds").
  • Use a round-tripping YAML loader. the "Write the documents: block back to the profile" step mutates an existing YAML file. Plain yaml.safe_load + yaml.safe_dump loses comments, block ordering, and quoting style — use ruamel.yaml or equivalent so hand-edited fields and comments survive.
  • Re-runnable. Re-running merges new bindings; existing bindings are preserved unless the user explicitly changes them. Safe to invoke as part of a profile refresh.

Prerequisites

  • Active target profile resolved per ../../context/target-platform-contract.md.
  • Documents available either under the recorded documents.root_path, under default <workspace>/Documents/, or as user-provided paths / URLs during manual prompts. A missing root only disables auto-scan; it is not a hard prerequisite.
  • ruamel.yaml or another round-tripping YAML writer for the profile edit step.

Limitations

  • Registers pointers only; never downloads, copies, or renames files.
  • Per-document field set is fixed to the schema in ../../references/platform_template.yaml — no ad-hoc keys.
  • Glob matching is filename-only; bad filenames in documents.root_path will under-bind and require manual selection.

Troubleshooting

  • documents.root_path missing — auto-scan is skipped. Provide an absolute root path, enter individual document paths / URLs manually, or skip the fields you do not want to bind.
  • Multiple files match a single slot — the skill stops and prompts; pick or rename the file. Example: two Jetson-Linux-Developer-Guide*.pdf files → keep the active version, rename the stale one.
  • Profile comments lost after write — a non-round-tripping YAML writer was used; switch to ruamel.yaml and rerun against a fresh pristine copy.
  • Validation fails because a binding points outside documents.root_path — documents.* are relative paths only; move the file under the root and retry.

References

Keep looking

Skills are one crate of 328,083. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.