CLI With Claude Skill — Authoring Discipline

Why this exists

Boss runs many projects where a CLI binary embeds and installs its own Claude Code skill into ~/.claude/skills/<tool>/SKILL.md. The same five mistakes recur across projects:

1. Hand-parsing os.Args / process.argv / sys.argv instead of using a real arg-parser.
2. The CLI ships no install-claude-skill command by default — Boss has to ask for it every time.
3. The skill file has no version: field, so the CLI cannot detect outdated installs and cannot auto-update.
4. CLI subcommands change but the embedded SKILL.md (the source-of-truth in the repo) is not updated in the same change. Shipped skill drifts from reality.
5. When Boss says "update the skill", the agent edits ~/.claude/skills/<tool>/SKILL.md (the installed copy) instead of the source file in the CLI repo. Next install-claude-skill overwrites the edits and the work is lost.

This skill enforces the rules that prevent each one. Violating the letter of the rules is violating the spirit of the rules.

Reference implementation

muveectl in the muvee repo is the canonical example. Mirror its structure for new CLIs:

• Embedded skill source: internal/skill/skill.md + internal/skill/skill.go (//go:embed).
• Install command + version-comparison auto-update: cmd/muveectl/main.go (functions claudeSkillPath, parseSkillVersion, shouldShowSkillOutdatedNotice, checkSkillNotice, cmdInstallClaudeSkill).
• Subcommand wiring with cobra: cmd/muveectl/login.go (look for the install-claude-skill cobra.Command).

Drop-in templates

The skill ships templates for Go, Rust, Node/TS, and Python under ~/.claude/skills/cli-with-claude-skill/template/. Pick the one matching your CLI's language.

Shared (language-agnostic):

Go + cobra:

Rust + clap (derive):

Node / TypeScript + commander:

Python + click:

The Rules

Rule 1 — Use a real arg-parser. Never hand-parse.

No exceptions:

• Don't write a switch os.Args[1] {…} dispatcher "just for now".
• Don't pull a "lighter" alternative without asking Boss first.
• One subcommand is enough to require the library. Do not wait until "it gets complicated".

Rule 2 — Every CLI ships install-claude-skill from day one.

The first commit of a new CLI MUST include all of:

1. An embedded skill source file at a stable repo path. Conventions:

• Go: internal/skill/skill.md (embedded via //go:embed).
• Rust: src/skill.md (via include_str!).
• Node: src/skill.md (bundled or read at install time).
• Python: <pkg>/skill.md (read via importlib.resources).

1. A <tool> install-claude-skill subcommand that writes the embedded content to ~/.claude/skills/<tool>/SKILL.md (creating the directory).
2. Startup notice logic in a PersistentPreRun hook (or equivalent):

• Installed file missing → print a one-line tip pointing at install-claude-skill.
• Installed version: older than embedded version: → auto-run install-claude-skill and print Notice: Claude skill is outdated. Auto-updating....

Do not invent your own naming. The subcommand name is exactly install-claude-skill. The install path is exactly ~/.claude/skills/<tool>/SKILL.md.

Rule 3 — The embedded SKILL.md MUST have a version: field.

---
name: <tool>
version: 1
description: Use when ...
---

• Integer counter (1, 2, 3, …) or semver — pick one and stick to it. Integer is simpler and matches Boss's existing tools.
• The CLI's update check parses this field. No version → no auto-update → Rule 2 is broken.
• Bump in the SAME commit that changes the skill body. Never bump-only commits, never body-only commits.

Rule 4 — CLI surface change ⇒ embedded SKILL.md update, in the same change.

When you add, rename, remove, or change the flags / behavior / defaults of a subcommand:

1. Edit the source code (cobra command, etc.).
2. Edit the embedded skill source file (internal/skill/skill.md or equivalent).
3. Bump version: in that file.
4. Build the CLI to confirm both compile.

Non-negotiable. A change that touches CLI surface but not the embedded SKILL.md is incomplete. Before reporting the task done, grep the embedded SKILL.md for the old command/flag name and confirm the doc is in sync. Run <tool> --help and skim — every subcommand listed there should also be documented in SKILL.md.

If unsure whether SKILL.md needs an update: read the diff. Any user-visible change (new flag, new subcommand, renamed argument, changed default, removed feature, changed output format) requires an update.

Rule 5 — Edit the source. NEVER the installed copy.

When Boss says "update the skill", "the skill is out of date", or "the skill is wrong", you will be tempted to Edit ~/.claude/skills/<tool>/SKILL.md directly. Do not. That file is rendered output. Editing it:

• Will be overwritten on the next <tool> install-claude-skill invocation (which the CLI may auto-run on every startup per Rule 2). Your edits vanish silently.
• Does not propagate to other machines / other developers / CI.
• Hides the real source from version control.

Procedure when asked to update an installed skill:

1. Identify the CLI behind the skill. The skill's name: is the binary name. which <name> if installed locally; otherwise the surrounding conversation will name the project.
2. Locate the embedded skill source in the project. Search order:

• internal/skill/skill.md
• cmd/<tool>/skill.md
• src/skill.md
• grep -rl "install-claude-skill\|claudeSkillPath\|embeddedSkill" . then follow the embed directive.

1. Edit THAT file. Bump version:.
2. Tell Boss to run <tool> install-claude-skill (or rebuild + invoke) to refresh the installed copy locally — do not run it yourself unless Boss approves, because rebuilding may not be free.

Red flags — STOP and re-evaluate if you catch yourself about to:

• Call Edit / Write on any path under ~/.claude/skills/<tool>/ while you are working inside that tool's source repo.
• Accept "update the skill" by jumping to the installed copy because it's "easier to find".
• Modify ~/.claude/skills/<tool>/SKILL.md and also the repo's internal/skill/skill.md "to keep them in sync" — only the repo file is authoritative; the installed copy is generated.

The installed copy at ~/.claude/skills/<tool>/SKILL.md is read-only from this skill's perspective. The only writer is the CLI's own install-claude-skill command.

Rationalization table

Quick checklists

Starting a new CLI from scratch

• [ ] Arg-parser library added as a dependency (cobra / clap / commander / click).
• [ ] Embedded skill source created at the conventional path with name, version: 1, and a CSO-compliant description (see skill-description-cso).
• [ ] install-claude-skill subcommand implemented.
• [ ] Startup notice (missing → tip; outdated → auto-update) implemented in PersistentPreRun or equivalent.
• [ ] --version flag implemented.
• [ ] First build passes.

Adding/renaming a subcommand or flag

• [ ] Subcommand registered via the arg-parser library — not hand-parsed.
• [ ] Embedded SKILL.md source updated (subcommand description, flags, one usage example).
• [ ] version: bumped in the embedded skill.
• [ ] Build passes.
• [ ] Optional sanity check: run <tool> install-claude-skill locally and grep <new-command> ~/.claude/skills/<tool>/SKILL.md to confirm the new section made it through.

When this skill does NOT apply

• Editing skills that are NOT installed by a CLI (e.g. hand-maintained skills like skill-tdd-and-persuasion, skill-description-cso). Those live at their installed path and are edited there directly.
• Pure library work with no binary entry point.
• Single-file scripts the user has explicitly said are throwaway.