Capture with /skillet:add

The /skillet:add slash command is the primary way to capture evals during your Claude Code sessions.

When to Use It

Use /skillet:add immediately after Claude produces a response that doesn't match your expectations:

• Wrong format or structure
• Missing information
• Incorrect behavior
• Suboptimal approach

Capture failures as they happen—don't wait until later.

How It Works

1. Notice a Problem

> Write a commit message for: added user authentication

Claude: Added user authentication feature with login, logout,
and session management.

2. Run the Command

> /skillet:add

3. Describe What You Expected

Claude will ask what you expected:

Claude: What behavior or format did you expect?

> Should use conventional commits format like "feat(auth): add user authentication"

4. Choose a Category

Claude checks existing eval categories and presents options:

Claude: Where should I save this eval?

1. conventional-commits (existing, 3 evals)
2. commit-messages — New category for commit format
3. git-conventions — New category for git conventions
4. Other

5. Confirm

Claude: Eval saved to ~/.skillet/evals/conventional-commits/004.yaml

  ├─ Prompt: "Write a commit message for: added user authentication"
  ├─ Actual: "Added user authentication feature..."
  ├─ Expected: "Should use conventional commits format..."

You now have 4 evals for this skill.

What Gets Captured

The slash command creates a YAML file with:

timestamp: "2025-01-22T10:30:45.123Z"
name: conventional-commits
prompt: "Write a commit message for: added user authentication"
actual: "Added user authentication feature with login, logout, and session management."
expected: "Should use conventional commits format like feat(auth): add user authentication"

Field	Source
timestamp	Current time
name	Your category choice
prompt	From conversation history
actual	Claude's response (summarized if long)
expected	What you described

Tips for Good Evals

Be Specific

Instead of:

"Should be better formatted"

Say:

"Should use markdown headers for sections and code blocks for examples"

Focus on Observable Behavior

Instead of:

"Should understand the context better"

Say:

"Should check for existing files before creating new ones"

One Issue Per Eval

Don't combine multiple problems:

"Should use conventional commits AND add a scope AND be shorter"

Split into separate evals if needed.

Group Related Evals

Use the same category name for related issues. 5+ evals per category gives the skill generator more to work with.

Storage Location

Evals are saved to:

~/.skillet/evals/<category>/
  001.yaml
  002.yaml
  ...

This location can be changed with the SKILLET_DIR environment variable.

Next Steps

After capturing 3-5 evals:

1. Run baseline: skillet eval <category>
2. Create skill: skillet create <category>
3. Test skill: skillet eval <category> ~/.claude/skills/<category>
4. Tune: skillet tune <category> ~/.claude/skills/<category>