kastalien-research Writing CLI Skills
How to write an agent skill for a command-line tool.
Quick Start
- Install the tool and play with it — don't just read docs. If the tool is unavailable, use WebFetch on official docs + man pages, but note this in your skill as "not hands-on verified"
- Run
--helpon every subcommand - Try the common operations yourself
- Note what surprises you or isn't obvious
- Copy
references/template.mdto your new skill directory - Fill in sections based on your hands-on experience
- Delete sections that don't apply
# First: actually use the tool
my-tool --help
my-tool subcommand --help
my-tool do-something # try it!
# Then: create the skill
# This will depend on the tool
# Claude Code
ln -s "$PWD/skills/my-tool" ~/.claude/skills/my-tool
# OpenCode
ln -s "$PWD/skills/my-tool" ~/.config/opencode/skills/my-tool
# Clawdbot
ln -s "$PWD/skills/my-tool" ~/clawd/skills/my-tool
Reading docs is no substitute for hands-on use. You'll discover defaults, gotchas, and real behavior that docs don't mention.
What NOT to Do
- Do not dump
--helpoutput verbatim — summarize the useful parts - Do not document every flag — focus on the 80% use cases
- Do not include commands you haven't tested
- Do not exceed 500 lines — this bloats agent context windows
Sections
Required
Every CLI skill needs at minimum:
| Section | Purpose |
|---|---|
| Frontmatter | name, description (with trigger phrases) |
| Installation | How to get the binary |
| Usage | The 80% use cases |
Recommended
| Section | When to Include |
|---|---|
| Requirements | Tool needs accounts, API keys, or dependencies |
| Quick Start | Tool has a simple "happy path" |
| Output Formats | Tool can output JSON or custom formats |
| Tips & Gotchas | Tool has some edge cases or things an agentic LLM should not use |
| Troubleshooting | Tool has debug modes or common failure modes |
| Configuration | Tool has config files or env vars |
| Uninstall | Tool leaves config/data behind |
| References | When there are useful docs or content that contains more details |
Writing Good Descriptions
Include trigger phrases so the agent knows when to load the skill:
# Good
description: Monitor RSS feeds for updates. Use when tracking blogs, checking for new posts, or building a feed reader workflow.
# Bad
description: RSS tool.
Organizing Commands
Group by task, not by command name:
## Usage
### View / List
### Create / Add
### Edit / Update
### Delete / Remove
### Search
Progressive Disclosure
Keep SKILL.md under ~500 lines. Move details to references/:
my-tool/
├── SKILL.md # Core usage
├── references/
│ ├── advanced-config.md # Deep config docs
│ └── api-reference.md # API details
└── scripts/
└── helper.sh # Helper scripts
Frontmatter Reference
---
name: tool-name # Required, matches directory
description: What + when # Required, include triggers
---
Checklist
Before publishing a CLI skill:
- Frontmatter has name + description with trigger phrases
- Installation covers target platforms
- Includes verification command (
tool --version) - Config file locations documented
- Required env vars listed
- Common operations in usage cover 80% of use cases
- Examples show realistic usage with sample output
- Output formats documented (if tool supports JSON/etc)
- Troubleshooting includes debug mode
- Uninstall cleans up config and data
- Under 500 lines (details in references/)
Template
See references/template.md for a complete starting point.