CLI Reference

Complete reference for Skillet command-line tools.

eval

Run evaluations against a baseline or with a skill loaded.

skillet eval <name> [skill] [options]

Arguments

Argument	Required	Description
name	Yes	Eval set name, directory path, or single .yaml file
skill	No	Path to skill directory (omit for baseline)

The name argument accepts:

• A name: looks in ~/.skillet/evals/<name>/
• A directory path: loads all .yaml files recursively
• A single .yaml file path

Options

Flag	Short	Type	Default	Description
--samples	-s	int	3	Number of iterations per eval
--max-evals	-m	int	all	Maximum evals to run (randomly sampled)
--tools		str	all	Comma-separated list of allowed tools
--parallel	-p	int	3	Number of parallel workers
--skip-cache		bool	false	Skip reading from cache (still writes)
--trust		bool	false	Skip confirmation for setup/teardown scripts
--no-summary		bool	false	Skip the failure summary LLM call

Examples

# Baseline eval (no skill)
skillet eval browser-fallback

# Eval with skill
skillet eval browser-fallback ~/.claude/skills/browser-fallback

# Single file
skillet eval ./evals/my-skill/001.yaml skill/

# More samples for statistical confidence
skillet eval my-skill -s 5

# Random sample of 5 evals
skillet eval my-skill -m 5

# More parallelism
skillet eval my-skill -p 5

# Force fresh runs (ignore cache)
skillet eval my-skill --skip-cache

# Skip script confirmation prompts
skillet eval my-skill --trust

# Skip failure summary (saves one LLM call)
skillet eval my-skill --no-summary

# Restrict available tools
skillet eval my-skill --tools "Read,Write,Bash"

tune

Iteratively improve a skill until target pass rate or max rounds.

skillet tune <name> <skill> [options]

Arguments

Argument	Required	Description
name	Yes	Eval set name or path
skill	Yes	Path to skill directory

Options

Flag	Short	Type	Default	Description
--rounds	-r	int	5	Maximum tuning rounds
--target	-t	float	100.0	Target pass rate percentage
--samples	-s	int	1	Samples per eval per round
--parallel	-p	int	3	Number of parallel workers
--output	-o	path	auto	Custom output path for results JSON

How It Works

1. Runs all evals against the current skill
2. Records pass rate and failure details
3. If pass rate < target, uses DSPy optimization to generate improved skill
4. Writes improved skill to disk
5. Repeats until target reached or max rounds

Results are saved to ~/.skillet/tunes/<name>/<timestamp>.json (or custom path with -o).

Examples

# Default tuning (5 rounds, 100% target)
skillet tune conventional-comments ~/.claude/skills/conventional-comments

# Aim for 80% pass rate
skillet tune conventional-comments skill/ -t 80

# More rounds
skillet tune conventional-comments skill/ -r 10

# Custom output file
skillet tune conventional-comments skill/ -o results.json

# Multiple samples for more reliable pass rates
skillet tune conventional-comments skill/ -s 3

create

Create a new skill from captured evals.

skillet create <name> [options]

Arguments

Argument	Required	Description
name	Yes	Eval set name (from ~/.skillet/evals/<name>/)

Options

Flag	Short	Type	Default	Description
--dir	-d	path	~	Base directory for output
--prompt		str	none	Extra prompt to customize generation

Output is written to <dir>/.claude/skills/<name>/SKILL.md.

Examples

# Create in home directory (default)
skillet create browser-fallback
# Creates ~/.claude/skills/browser-fallback/SKILL.md

# Create in current directory
skillet create browser-fallback -d .
# Creates ./.claude/skills/browser-fallback/SKILL.md

# Create in specific project
skillet create browser-fallback -d /path/to/project

# Add extra instructions
skillet create browser-fallback --prompt "Be concise, max 20 lines"

generate-evals

Generate candidate eval files from a SKILL.md.

skillet generate-evals <skill> [options]

Arguments

Argument	Required	Description
skill	Yes	Path to skill directory or SKILL.md file

Options

Flag	Short	Type	Default	Description
--output	-o	path	auto	Output directory for candidate files
--max		int	5	Max evals per category
--domain	-d	str	all	Filter to specific domain(s): triggering, functional, performance

Domains

Each generated eval is tagged with a domain indicating what aspect of the skill it tests:

Domain	Description
triggering	Does the skill activate when it should (and stay silent when it shouldn't)?
functional	Does the skill produce correct output once triggered?
performance	Does the skill meet quality, latency, or efficiency expectations?

By default all domains are generated. Use --domain to filter to specific ones. The flag can be repeated to select multiple domains.

Examples

# Generate from skill directory
skillet generate-evals ~/.claude/skills/browser-fallback

# Custom output location
skillet generate-evals skill/ -o ./my-evals/

# Limit to 3 per category
skillet generate-evals skill/ --max 3

# Only triggering evals
skillet generate-evals skill/ -d triggering

# Triggering and functional evals
skillet generate-evals skill/ -d triggering -d functional

lint

Lint a SKILL.md file for common issues.

skillet lint <path> [options]

Arguments

Argument	Required	Description
path	Yes*	Path to a SKILL.md file (*not required with --list-rules)

Options

Flag	Type	Default	Description
--no-llm	bool	false	Skip LLM-assisted lint rules
--list-rules	bool	false	List all available rules and exit

Rules

Skillet ships with 14 built-in rules across four categories:

Naming

Rule	Severity	Description
filename-case	warning	Skill file must be named exactly SKILL.md
folder-kebab-case	warning	Skill folder name must be kebab-case
name-kebab-case	warning	Name field must be kebab-case
name-matches-folder	warning	Name field must match the parent folder name
name-no-reserved	error	Name must not contain claude or anthropic

Structure

Rule	Severity	Description
frontmatter-valid	warning	Frontmatter must include name and description
frontmatter-delimiters	error	YAML frontmatter must have --- delimiters
frontmatter-no-xml	error	No XML angle brackets (< >) in frontmatter
description-length	warning	Description must be under 1,024 characters
body-word-count	warning	Skill body should be under 5,000 words
no-readme	warning	No README.md inside skill folder

Fields

Rule	Severity	Description
field-license	warning	Recommended license field should be present
field-compatibility	warning	Recommended compatibility field should be present
field-metadata	warning	Recommended metadata field should be present

Exit Codes

Code	Meaning
0	No issues found
1	One or more findings
2	Error (file not found, invalid input)

Examples

# Lint a skill
skillet lint ~/.claude/skills/my-skill/SKILL.md

# Lint without LLM-assisted checks
skillet lint --no-llm path/to/SKILL.md

# List all available rules
skillet lint --list-rules

compare

Compare baseline vs skill results from cache.

skillet compare <name> <skill>

Arguments

Argument	Required	Description
name	Yes	Eval set name or path
skill	Yes	Path to skill directory

Run skillet eval <name> and skillet eval <name> <skill> first to populate the cache.

Examples

skillet compare browser-fallback ~/.claude/skills/browser-fallback

show

Inspect cached eval results without re-running any evals.

skillet show <name> [options]

Arguments

Argument	Required	Description
name	Yes	Eval set name or path

Options

Flag	Short	Type	Default	Description
--eval	-e	str	none	Show detailed results for a specific eval file
--skill	-s	path	none	Show results with a skill loaded instead of baseline

Examples

# Show cached baseline results
skillet show browser-fallback

# Show results with a skill
skillet show browser-fallback --skill ~/.claude/skills/browser-fallback

# Show detailed results for a specific eval
skillet show browser-fallback --eval 001.yaml

# Combine: specific eval with skill
skillet show browser-fallback --skill path/to/skill --eval 001.yaml

Environment Variables

Variable	Default	Description
SKILLET_DIR	~/.skillet	Base directory for evals, cache, and tune results

Exit Codes

Code	Meaning
0	Success
1	Error (invalid arguments, missing files, etc.)