You are creating a new skill for the user. Skills are short instruction sets that get injected into the system prompt when the user types /<skill-id> in chat. Treat the user's message as the full spec — don't run a question loop unless the request is too vague to act on (in that case ask exactly one clarifying question, then proceed).
Steps
Pick an id. Derive a short kebab-case id from the user's request. Rules: lowercase letters, digits, and hyphens only; ≤ 64 chars; descriptive but terse (e.g. summarize-url, daily-standup, recipe-from-photo).
Check for conflicts. Run execute_shell_command with ls /root/skills/ 2>/dev/null to see existing ids. If your chosen id is taken, append -2, -3, etc. until it's free.
Write a tight description. ≤ 1024 chars, non-empty, single line, written as a third-person summary of what the skill does. This shows up in the slash-command menu and Settings.
Draft the body. Markdown instructions addressed to a future model invocation — clear steps, any constraints, and how to interpret args after the slash command. Keep it focused; every byte ends up in the system prompt when the skill is active.
Install it. Run execute_shell_command once with this exact shape (substitute <id>, <description>, and the body):
mkdir -p /root/skills/<id> && cat > /root/skills/<id>/SKILL.md <<'SKILLEOF'
---
name: <id>
description: <description>
---
<body markdown here, verbatim>
SKILLEOF
Use the quoted heredoc delimiter (<<'SKILLEOF') so YAML/markdown special characters are preserved exactly.
Confirm. In one short message, tell the user the chosen id, the description, and that they can invoke it now as /<id>. Mention they can ask you in normal chat to tweak it later (you'll cat and rewrite the file).
Worked example
User: /create-skill summarize any URL the user pastes into 3 bullet points
You pick id summarize-url, check ls /root/skills/ (free), draft:
- description:
Fetches a URL and summarizes the page in 3 concise bullet points.
- body: a few lines telling the model to read the URL from the user's message, fetch it via
execute_shell_command (e.g. curl -sL), and reply with exactly three markdown bullets.
Then run the heredoc, confirm: "Installed /summarize-url — paste a URL after it to try."
Notes
- Don't propose bundled scripts/files unless the user specifically asked for them; SKILL.md alone covers most needs.
- If the user's request describes something better suited to existing tools (e.g. "summarize this PDF" → already covered by file uploads), say so and offer the skill anyway only if they confirm.
- The skill list auto-refreshes after this turn — no restart needed.
Source: SimonSchubert/Kai — distributed by TomeVault.
1---2name: create-skill-23description: Create a new Kai skill from a single natural-language request, then install it into the sandbox so it's invokable as /<id>. Use when this capability is needed.4---56You are creating a new skill for the user. Skills are short instruction sets that get injected into the system prompt when the user types `/<skill-id>` in chat. Treat the user's message as the full spec — don't run a question loop unless the request is too vague to act on (in that case ask exactly **one** clarifying question, then proceed).78## Steps9101. **Pick an id.** Derive a short kebab-case id from the user's request. Rules: lowercase letters, digits, and hyphens only; ≤ 64 chars; descriptive but terse (e.g. `summarize-url`, `daily-standup`, `recipe-from-photo`).11122. **Check for conflicts.** Run `execute_shell_command` with `ls /root/skills/ 2>/dev/null` to see existing ids. If your chosen id is taken, append `-2`, `-3`, etc. until it's free.13143. **Write a tight description.** ≤ 1024 chars, non-empty, single line, written as a third-person summary of what the skill does. This shows up in the slash-command menu and Settings.15164. **Draft the body.** Markdown instructions addressed to a future model invocation — clear steps, any constraints, and how to interpret args after the slash command. Keep it focused; every byte ends up in the system prompt when the skill is active.17185. **Install it.** Run `execute_shell_command` once with this exact shape (substitute `<id>`, `<description>`, and the body):1920 ```sh21 mkdir -p /root/skills/<id> && cat > /root/skills/<id>/SKILL.md <<'SKILLEOF'22 ---23 name: <id>24 description: <description>25 ---26 <body markdown here, verbatim>27 SKILLEOF28 ```2930 Use the quoted heredoc delimiter (`<<'SKILLEOF'`) so YAML/markdown special characters are preserved exactly.31326. **Confirm.** In one short message, tell the user the chosen id, the description, and that they can invoke it now as `/<id>`. Mention they can ask you in normal chat to tweak it later (you'll `cat` and rewrite the file).3334## Worked example3536User: `/create-skill summarize any URL the user pastes into 3 bullet points`3738You pick id `summarize-url`, check `ls /root/skills/` (free), draft:3940- description: `Fetches a URL and summarizes the page in 3 concise bullet points.`41- body: a few lines telling the model to read the URL from the user's message, fetch it via `execute_shell_command` (e.g. `curl -sL`), and reply with exactly three markdown bullets.4243Then run the heredoc, confirm: "Installed `/summarize-url` — paste a URL after it to try."4445## Notes4647- Don't propose bundled scripts/files unless the user specifically asked for them; SKILL.md alone covers most needs.48- If the user's request describes something better suited to existing tools (e.g. "summarize this PDF" → already covered by file uploads), say so and offer the skill anyway only if they confirm.49- The skill list auto-refreshes after this turn — no restart needed.5051---52> Source: [SimonSchubert/Kai](https://github.com/SimonSchubert/Kai) — distributed by [TomeVault](https://tomevault.io).53<!-- tomevault:4.0:skill_md:2026-06-29 -->
tomevault-io