Conciseness

The body is ~1000 lines of dense, non-generic operational knowledge (exit codes, error-string semantics like `CONFIG_OPERATION_STORAGE_UNCERTAIN` and `attached read-only; retry input...`), so it rarely explains things Claude already knows. However, several deep edge-case sections — the provider refresh/Vault lifecycle, the `stable_placeholder` credential contract, and the 20-line stdin-attachment recovery paragraph — read like a troubleshooting manual dump that all users of the skill pay for on every invocation, even for a simple `sandbox create`. Anchor 3 ("mostly efficient but could be tightened") fits better than 4 given the volume of inline failure-mode detail.

Actionability

Nearly every section gives copy-paste-ready commands with concrete flags (e.g., `openshell sandbox create --from ... --provider claude-code -- claude`, `openshell policy set dev --policy current-policy.yaml --wait`), specific exit codes (0/1/124 for `--wait`), and exact error strings with named remediations. This is fully executable guidance covering the common cases.

Workflow Clarity

Ten workflows are clearly sequenced; Workflow 4 (policy iteration) has numbered steps, an explicit monitor-observe-modify-push-verify feedback loop with a diagram, validation via `--wait` plus `policy list` revision status checks, and a documented retry/repair path (e.g., `ConfigurationInvalid` repair window, `ProvisioningTimedOut` handling). Destructive operations like `sandbox delete` include validation guidance ("deletion accepted" means pending; inspect before assuming completion).

Progressive Disclosure

There are no bundle files at all (no references/, scripts/, or assets/), so everything — including advanced material that clearly belongs in separate files, like the provider-refresh lifecycle, BYOC UID/GID rules, and attachment-recovery semantics — is inlined in one ~1000-line SKILL.md. The body is well-sectioned with headers and does signal external authorities (`openshell --help` tree and published docs links), which lifts it above the anchor-2 'minimal structure' case, but the absence of any one-level-deep reference split for the deep-dive material keeps it at anchor 3.

Specificity

The description enumerates the full concrete capability surface of the CLI: "sandbox management, gateway registration, provider configuration and refresh, profile management, policy iteration, settings, service exposure, BYOC workflows, and attached-provider inference", which matches the workflows actually covered in the body with no visible gaps. It stays in a consistent third-person imperative voice ("Guide agents through using...") with no vague filler.

Completeness

It explicitly answers both questions: "what" in the opening sentence (guide agents through the OpenShell CLI for the listed management tasks) and "when" via an explicit "Trigger keywords - ..." clause listing concrete trigger phrases. The keyword list is the equivalent of a 'Use when...' clause with concrete triggers, so it meets the 5-anchor rather than being capped.

Trigger Term Quality

The keyword list covers natural phrasings users would actually say, including synonyms and expansions: "openshell", "sandbox create", "sandbox exec", "logs", "provider create", "policy set", "settings", "service expose", "forward", "port forward", "BYOC", "bring your own container", "use openshell", "run openshell", "CLI usage", "manage sandbox", "manage provider", "gateway add", "gateway select". Concrete command names plus colloquial variants give comprehensive coverage.

Distinctiveness Conflict Risk

Every trigger is namespaced to this specific tool ("openshell", "sandbox create", "provider refresh", "gateway add", "BYOC" in the OpenShell sense). The clear niche means minimal overlap risk with unrelated skills.