> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/docsfab/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/docsfab/_mcp/server. # Skill Behavior Contracts in DocsFab > Look up the eight documentation skills and their current runtime bindings. The canonical DocsFab [`packages/skills/`](../../../../packages/skills/README.md) catalog contains eight skill contracts. The runtime stages exact hashed copies and schedules them through fixed CLI workflows. Skills are playbooks; runtime is the job runner. Both pins must agree before a `docsfab` job exists. Refer to [skills ↔ runtime](../../../../packages/skills/README.md#relationship-to-the-runtime). For more information, refer to [DocsFab Quality Reference](/docsfab/reference/quality). ## Read a Skill Contract Each leaf reference identifies the skill's mode, runtime binding, primary evidence or mutation, stop conditions, forbidden actions, and authoritative source. A skill contract defines reusable behavior. The runtime manifest owns step order, activation, tool access, credentials, recovery, and write authorization. ## Choose a Skill Stage Plan an approved documentation change or produce a bounded patch. Refer to [Planning and Authoring](/docsfab/reference/quality/skills/plan-and-author). Inventory source facts, run deterministic checks, verify technical claims, or perform independent review. Refer to [Checking and Review](/docsfab/reference/quality/skills/check-and-review). Audit a documentation set, test audience understanding, or evaluate quality evidence. Refer to [Assessment and Evaluation](/docsfab/reference/quality/skills/assess-and-evaluate). ## Skill Catalog | Outcome | Canonical skill | Runtime exposure | | --- | --- | --- | | Audit focused Markdown or a documentation set | [`audit-documentation`](/docsfab/reference/quality/skills/assess-and-evaluate/audit-documentation) | Review and restructuring inventory steps | | Build an approved plan | [`plan-documentation`](/docsfab/reference/quality/skills/plan-and-author/plan-documentation) | Creation, update, and restructuring planning steps | | Create or edit documentation, including focused polish | [`write-documentation`](/docsfab/reference/quality/skills/plan-and-author/write-documentation) | Focused live-agent edits; mutation step in creation, update, and restructuring | | Run deterministic checks | [`lint-documentation`](/docsfab/reference/quality/skills/check-and-review/lint-documentation) | Review, creation, update, and restructuring validation steps | | Inventory source facts, then verify technical claims | [`verify-technical-accuracy`](/docsfab/reference/quality/skills/check-and-review/verify-technical-accuracy) | Inventory-only Pin A step on create and review; verification step in all four workflows | | Perform independent peer review | [`review-documentation`](/docsfab/reference/quality/skills/check-and-review/review-documentation) | Review, creation, update, and restructuring review steps | | Test audience understanding | [`test-audience-understanding`](/docsfab/reference/quality/skills/assess-and-evaluate/test-audience-understanding) | Review, creation, update, and restructuring audience steps | | Evaluate quality evidence | [`evaluate-document-quality`](/docsfab/reference/quality/skills/assess-and-evaluate/evaluate-document-quality) | Comprehensive review, creation, update, and restructuring evaluation steps | The [`create-documentation`](/docsfab/reference/workflow-contracts/create-documentation) and [`restructure-documentation`](/docsfab/reference/workflow-contracts/restructure-documentation) jobs compose these skills as separate catalog steps. Their mutation steps use `write-documentation` in different modes. Creation and comprehensive review schedule `verify-technical-accuracy` in inventory-only mode as Pin A step `derive-source-obligations` after planning or audit. That step is not a ninth catalog skill. Update can select a read-only [no-change assessment](/docsfab/reference/workflow-contracts/update-documentation) after planning; that branch skips writing and linting while retaining independent verification, review, audience, and quality evidence. Runtime owns execution state, scheduling, activation, tools, credentials, recovery, and SCM authority. `guided-demo` is a benchmark fixture name. Detailed quality methods remain available through references owned by the skills. ## Task-Specific Methods Skills load detailed methods when the task needs them. Planning selects research questions for the page purpose. Writing and review apply practical reader-path and presentation checks. Technical verification checks complete examples and changed interfaces, distinguishing static inspection from executed behavior. Correction reviews preserve originating findings and check adjacent content. A long or mixed-purpose page does not turn a bounded update into restructuring. The planner reports a separate structural recommendation or a material scope gap. Release-note work distinguishes reader impact from approved disclosure; neither drafting nor review grants publication authority. Audit findings use approved source authority to reconcile implementation, schema, and test conflicts. Agreement counts do not establish a probability of truth. Checks unavailable to the host remain explicit evidence gaps. ## Shared Boundaries - Use approved, version-identifiable sources and a clear request or brief. Workflow writing requires its supplied plan binding. - Require explicit runtime write authorization before a write step. - Keep Writer, source-obligation analyst, Technical Verifier, Peer Reviewer, audience participant, and Quality Evaluator evidence distinct. - Use clean non-author providers when independence is required. - Permit at most one automatic Writer revision by default. - Do not publish, commit, push, open a merge request, or merge. Disk presence does not activate a DocsFab workflow. Runtime use requires a locked, contained catalog pack. `docsfab run` selects a fixed workflow; `docsfab live-agent` can select available skills for a natural-language request. Both entry points preserve tool and write-authorization boundaries.