skill-father（技能父亲）

Skill Father

This skill is Boss’s opinionated, authoritative standard for creating/updating skills.

It is based on the upstream skill-creator guidance, with extra requirements:

• Always include Prerequisites checks (fail fast).
• Keep skills portable/shareable: do not bake machine-specific settings into SKILL.md.
• Always include Initialization / Installation / Onboarding that prompts the user when needed.
• Make skills reproducible for other people/machines.

Core principles (from skill-creator)

• Concise is key: minimize context bloat.
• Progressive disclosure: keep SKILL.md short; put big docs in references/, deterministic code in scripts/.
• Avoid extra docs (README/CHANGELOG/etc.).

Required sections in every skill

1) Prerequisites

Include a short section with concrete checks/commands.

Examples:

• 1Password-backed workflows:

- op whoami must succeed (or, if service accounts are used, required env vars like OP_SERVICE_ACCOUNT_TOKEN must be set).

• External CLIs:

- command -v <tool> must exist; include install guidance if missing.

2) Configuration (portable)

Rules:

• Never hardcode machine/user-specific paths, usernames, tenant IDs, tokens, etc. inside SKILL.md.
• Prefer skill-local config files stored next to SKILL.md, e.g.:

- config.env (dotenv-style KEY="VALUE") - config.json (structured)

• Config must be split into two files:

- config.env.example (or config.json.example) — checked-in/shareable example; never mutated by onboarding - config.env (or config.json) — real machine-specific values written/updated during onboarding

• SKILL.md documents:

- where config lives - required keys + defaults - which file is the example vs real - how to run onboarding to generate/update the real config

3) Initialization / Installation / Onboarding

Provide a guided first-run flow.

• If setup is trivial and safe: can be silent.
• Otherwise: ask the user for choices + confirmation.
• Persist outcomes into the real skill-local config file (not into SKILL.md; do not modify the example file).
• Prefer discovery + confirmation over assumptions.

Prefer an onboarding helper script when setup touches real machine state.

Chat-first onboarding (Telegram-friendly)

When the primary interface is chat (e.g. Telegram), do not rely on TTY-style interactive prompts.

Requirement: Every child skill should explicitly document a “Preferred (chat-first)” onboarding path.

Preferred pattern:

1. Agent asks the user the required onboarding questions in chat.
2. Agent writes/updates the real skill-local config file.
3. Agent runs a smoke test and reports results.

If you do ship an interactive script, treat it as an optional convenience for users running it in a real terminal (document as “Optional (terminal)”).

Recommended onboarding script behaviors:

• Generate/update the real config file from prompts and/or auto-discovery.
• If editing an existing system config file (e.g. ~/.config/openclaw/env, ~/.ssh/config):

- detect whether the target file exists; create if missing - for each key/entry that would change, show current vs new - prompt the user per item: keep / override / skip - for secrets/tokens, mask values in prompts

• If a restart/reload is required:

- first detect whether the service manager is available (e.g. systemctl --user status <svc>) - ask the user for confirmation before restarting - if not detectable/available, print clear manual instructions

Examples of onboarding steps:

• Detect candidate paths/resources.
• Present options.
• Ask for confirmation.
• Write config.
• Validate config by running a quick self-test.

4) Reproducibility

• The skill should work for other people with minimal edits.
• Prefer parameterization/config + prompts.
• Avoid environment-specific assumptions unless explicitly documented.

5) Executables / bin placement

• Any executable scripts/binaries required by the skill should live inside the skill folder (or inside the relevant plugin’s folder).
• For convenience, you may create a symlink into a common PATH location (e.g. ~/.local/bin/<name>), but the canonical copy should remain in the skill/plugin directory.

Resource layout

Use the standard skill layout:

skill-name/
├── SKILL.md
├── config.env.example    # example (shareable)
├── config.env            # real machine-specific config (generated/updated by onboarding)
├── scripts/              # deterministic code
└── references/           # optional docs, loaded on demand

Process checklist (for the agent)

1. Understand the task and collect concrete usage examples.
2. Plan resources (scripts/, references/, assets/) only if they reduce repetition or increase reliability.
3. Create/confirm required sections: Prerequisites, Config, Installation/Onboarding.
4. Implement the smallest working version.
5. Validate with a smoke test.
6. Iterate.

Example expectations: ssh-op skill

• Prereqs: confirm op whoami works (or service account env is set) and ssh/ssh-add/ssh-agent exist.
• Onboarding: proactively discover/confirm:

- vault name - SSH key item - host + host aliases stored in the 1Password item

• Integration: check whether aliases exist in ~/.ssh/config; if missing, offer to add/update entries.
• Config: store vault/item/host/aliases in a skill-local config file.