Slancha skills setup for Claude Code — recipe v1 (2026-10-04)

Goal: install ONE named, human-approved skill and verify its published revision.
Read https://app.slancha.ai/docs/install-the-cli and https://app.slancha.ai/docs/skills first.

Before changing anything:
- Ask which workspace, project directory and exact published skill name the person wants.
- Do not invent a skill or choose every skill in the library.
- Confirm curl, Python >=3.9, sha256sum or shasum, git and Claude Code are available.
- Work as the current user, never root. Explain local installation changes first.
- Skills-only setup does not enable tracing. Do not use --tracing or --force.
- Never print, copy into chat, or upload tokens, private config, traces or credentials.

Install only after the person authorizes it:
  curl -fsSL https://app.slancha.ai/install.sh -o /tmp/slancha-install.sh
Read the downloaded installer. It checks the pinned CLI download's sha256, installs under
~/.local/share/slancha and ~/.local/bin, then runs auth and init. The plugin marketplace
installation follows the plugin repository's current release; do not describe it as pinned.
Verified 2026-10-04: CLI ref 64a6b0937018ec9a4c88c09452358394cee6ed31,
sha256 cfadf42df5516a0ca54dd16aee9db2d03e03b2b765a2a50c9b260a2250bd498a.
If the installer has changed, review its new source and checksum before proceeding.
  sh /tmp/slancha-install.sh
Pause for the HUMAN to open the approval URL, match the device code and select the intended
workspace. Authentication is theirs. If a host override is needed, use the command shown
by that workspace's Claude Code setup page; never guess the API host.

If the CLI is already installed, inspect its status and preview setup:
  slancha status
  slancha init --dry-run
After review, run slancha init only if setup is needed. Restart Claude Code when requested.

In the intended project directory, use the exact name confirmed by the person:
  approved_skill='REPLACE_WITH_THE_CONFIRMED_PUBLISHED_SKILL_NAME'
  slancha update "$approved_skill" --dry-run
If it is not published, STOP: review the pending change in the console; an authorized
person must approve it after the scan clears. Never approve on their behalf.
After the person agrees to the preview:
  slancha update "$approved_skill"
  slancha status

Completion: status lists that exact skill, its revision, the intended location and
"up to date", with no local-change warning. Compare the revision with the approved
version in the console. An offline check, unknown revision, locally edited package,
missing skill or successful login alone is NOT completion. Report the remaining step.
Use /reload-skills or restart Claude Code if requested by the CLI. Check that Claude Code
recognizes the skill; do not run a consequential task just to test it.

Optional tracing is a separate decision after skills setup. It requires admin approval
and a separate privacy review. Premium session mining is separately enabled and can send
session excerpts to the workspace's model provider. No scoring or savings are promised.

If copy, download, auth, scan, network or version verification fails: explain the exact
stage, keep local edits and existing configuration, and link to the manual docs.