Skill source and installation disclosure

The public skills directory lists all current Ansight skill instructions and supporting files, including agent configuration, scripts, and references. Every file has a direct source URL, byte size, and SHA-256 hash. The JSON catalog exposes the same complete file inventory in files; its skills list contains the Markdown instruction entries.

What the installer adds

When you accept agent-skill installation, the CLI installer copies these eight skill directories from the downloaded CLI archive:

Select a skill to review its complete source and supporting files in the browser. The viewer includes raw-file links, byte sizes, and SHA-256 hashes for the current published files. For an installed version, use the exact release contents below.

Each directory includes its SKILL.md and any supporting files. Six workflow skills currently include agents/openai.yaml, which supplies agent-facing metadata and a suggested prompt. The readiness skill also includes scripts/score_readiness.py.

Depending on your choices, the installer copies these directories into ~/.agents/skills (Codex and shared Agent Skills clients), ~/.cursor/skills, or ~/.claude/skills. On Windows, ~ means your user profile directory. Installation is optional; --no-setup on macOS/Linux or -NoSetup on Windows skips skill setup. --yes / -Yes accepts setup prompts.

The larger public catalog also contains platform integration, inspection, logging, and remote-tool authoring skills. These are available on demand and may be fetched by the agent when instructed by a router. The CLI installer does not install the entire public catalog.

Exact release contents

Current website files can change independently of an installed CLI. Use the release lookup to select the version, build number, RID, and public/preview channel shown by ansight version and the installation receipt. The receipt’s archiveUrl also identifies the original download.

Every CLI archive published through the disclosure-enabled publisher has a sidecar at its archive URL with .skills.json appended:

https://ansightaus.blob.core.windows.net/builds/cli/{rid}/{version}/{build}/ansight-cli-{rid}-{version}-{build}.tar.gz.skills.json

Preview releases use builds/cli/preview/ in place of builds/cli/. The version/build/RID paths are immutable. Publishing refuses to replace a manifest with different contents at the same path. A release manifest contains:

• the CLI version, build number, RID, archive filename, and archive SHA-256;
• every file beneath the archive’s skills/ directory, including hidden files and supporting files of any extension;
• each file’s path relative to the agent’s skills root, its path inside the archive, byte size, and SHA-256;
• the exact UTF-8 file contents in text, or contentBase64 for non-UTF-8 files.

The manifest is generated from the archive while verifying its SHA-256, so it describes the packaged bytes rather than a later checkout or website revision. The manifest covers Ansight’s bundled skill files. The original archive is the source for the rest of the CLI payload.

Older archives may not yet have a sidecar. Their complete source files remain inside skills/ in the original archive. A sidecar can be generated from those same archive bytes without rebuilding or replacing the release. Do not substitute today’s website hashes for an older release.

Verify installed copies

1. Download the matching .skills.json manifest and the archive’s .sha256 sidecar. Check that archiveSha256 matches the archive checksum. If reviewing the archive itself, also calculate its SHA-256 locally.
2. Download and review the verification script. It uses Node.js built-in modules, reads local files, and does not execute skill contents or make network requests.
3. Run it for every agent installation you use:

node verify-skill-installation.mjs release.skills.json "$HOME/.agents/skills"
node verify-skill-installation.mjs release.skills.json "$HOME/.cursor/skills"
node verify-skill-installation.mjs release.skills.json "$HOME/.claude/skills"

On Windows PowerShell:

node .\verify-skill-installation.mjs .\release.skills.json "$env:USERPROFILE\.agents\skills"

The verifier reports missing, changed, extra, or linked files within the bundled skill directories, plus other ansight-* directories absent from the manifest. It exits unsuccessfully when the installation differs. The installer can retain extra local files, and an install-only CLI update does not refresh skills, so audit each actual agent directory rather than assuming it matches the current CLI. Other vendors’ skills and project-local instructions require their own review.

Prompt injection and trust boundaries

Skill instructions influence what an agent reads, writes, and runs. Review the full instruction text, agent metadata, helper code, and any linked instructions before allowing an agent to use them. Remote documents and on-demand skills can change; retain the exact fetched copies when your review requires a fixed record.

Disclosure and hashes support inspection and change detection. They do not prove that instructions are safe, independently authenticate a release, or prevent prompt injection. Treat app content, logs, network responses, and tool output as untrusted evidence, even when a reviewed skill retrieves them. Keep agent permissions and Ansight tool policies scoped to the work you intend to authorize.