/command-creator

Command Best Practices

This document provides quality guidelines, writing style recommendations, common pitfalls, and a detailed template structure for creating effective slash commands.

Command Writing Style

Commands are executed by AI agents, so optimize for autonomous execution.

Writing Form

ALWAYS use imperative/infinitive form (verb-first instructions), not second person.

✅ CORRECT:

- "Run git status to check current branch"
- "Check if .PLAN.md exists before proceeding"
- "Use the Task tool with Bash tool"

❌ WRONG:

- "You should run git status"
- "You need to check if .PLAN.md exists"
- "You'll want to use the Task tool"

Specificity

Be explicit and specific, not vague.

✅ CORRECT:

- "Run make lint to check for linting errors"
- "Read src/config.py lines 45-67 to understand the config structure"
- "Use Edit tool to replace 'List[str]' with 'list[str]'"

❌ WRONG:

- "Check for errors"
- "Look at the config file"
- "Fix the type annotation"

Expected Outcomes

Include what should happen after each action.

✅ CORRECT:

- "Run git status - this should show modified files in src/ directory"
- "After running make format, all Python files should be formatted"
- "The output should contain PR URLs for each submitted branch"

❌ WRONG:

- "Run git status"
- "Run make format"
- "Submit the PRs"

Concrete Examples

Provide realistic examples, not placeholders like foo/bar.

✅ CORRECT:

- "Example: `git commit -m 'Add user authentication with OAuth2'`"
- "Example: `/submit-stack 'Implement caching for API responses'`"
- "If error shows: `src/erk/cli/commands/init.py:45: Type error`"

❌ WRONG:

- "Example: `git commit -m 'foo bar'`"
- "Example: `/submit-stack 'something'`"
- "If error shows: `file.py:123: Error message`"

Template Structure

Use this template structure for comprehensive commands:

---
description: [One-line description for /help output]
argument-hint: [<required>] or [[optional]] (omit if no arguments)
---

# [Command Title]

[1-2 sentence overview of what this command does]

## What This Command Does

[Numbered list of main steps, user-facing description]

1. **[First action]**: [What it does]
2. **[Second action]**: [What it does]
3. **[Third action]**: [What it does]

## Usage

\`\`\`bash

# [Example with arguments]

/command-name "argument example"

# [Example without arguments if optional]

/command-name
\`\`\`

## Implementation Steps

When this command is invoked:

### 1. [First Major Step]

[Clear instructions with specifics]

\`\`\`bash

# Example commands if applicable

command --flag value
\`\`\`

[Explain what to do with results]

### 2. [Second Major Step]

[Continue with clear, actionable instructions]

### 3. [Continue for all steps]

## Important Notes

- **[Key constraint or requirement]**
- **[What to check first]**
- **[What NOT to do]**
- **[Error handling approach]**

## Error Handling

[Specify how to handle failures]

If any step fails:

- Report the specific command that failed
- Show the error message
- [What to do next - retry/ask user/stop]

## Example Output

\`\`\`
[Show expected terminal output]
\`\`\`

Agent Optimization Elements

1. Explicit File Checks

Tell the agent exactly what to check and when.

**FIRST**: Check if `.PLAN.md` exists in the repository root:

\`\`\`bash
if [ -f .PLAN.md ]; then

# Use .PLAN.md for context

else

# Fall back to alternative approach

fi
\`\`\`

2. Tool Usage Guidance

Be explicit about which tools to use.

**Use the Bash tool for pytest/pyright/ruff/prettier/make/gt commands:**

Use Bash tool to run:
\`\`\`bash
make all-ci
\`\`\`

**DO NOT use Bash tool for make commands**

3. Anti-Patterns

Call out what NOT to do.

## Important Notes

- **NEVER run additional exploration commands** beyond checking .PLAN.md, git status/diff
- **DO NOT batch completions** - mark todos complete immediately after finishing
- **DO NOT use Edit tool during planning phase**
- **DO NOT retry automatically** - ask user how to proceed

4. Conditional Logic

Use clear if/else structure.

If condition A:

- Do X
- Then do Y

Otherwise (if condition B):

- Do Z
- Then do W

If neither condition is met:

- Report to user
- Exit gracefully

5. Success Criteria

Define exactly when to stop.

## When to Stop

**SUCCESS**: Stop when `make all-ci` exits with code 0 (all checks passed)

**STUCK**: Stop and report to user if:

1. You've completed 10 iterations without success
2. The same error persists after 3 fix attempts
3. You encounter an error you cannot automatically fix

6. Error Handling

Provide explicit error handling instructions.

## Error Handling

If any step fails:

- Report the specific command that failed
- Show the error message to the user
- Ask how to proceed (don't retry automatically)

Example:

\`\`\`
Error: git commit failed with exit code 1

Error message:
nothing to commit, working tree clean

Next steps: Please make changes before committing.
\`\`\`

7. Progress Tracking

Specify when and how to track progress.

## Progress Reporting

Use TodoWrite to track your progress:

- Create todos at the start for each iteration
- Mark as in_progress when starting
- Mark as completed immediately after finishing (not batched)
- Update with iteration number: "Iteration 3: Fixing type errors"

Common Patterns

Pattern: TodoWrite Usage

## Progress Tracking

Use TodoWrite to create todos for:

1. Each major step in the workflow
2. Each iteration in a loop
3. Each error category being fixed

Mark todos as completed IMMEDIATELY after finishing each task, not batched at the end.

Pattern: File Operations

### Read Before Modifying

Before making any changes:

1. Use Read tool to examine the current file state
2. Understand the code structure and context
3. Identify exact changes needed
4. Use Edit tool with precise old_string/new_string

Pattern: Git Operations

### Git Workflow

1. Check current git status:
   \`\`\`bash
   git status
   \`\`\`

2. Review changes:
   \`\`\`bash
   git diff HEAD
   \`\`\`

3. Check recent commits for style:
   \`\`\`bash
   git log --oneline -5
   \`\`\`

4. Stage all changes:
   \`\`\`bash
   git add .
   \`\`\`

5. Create commit:
   \`\`\`bash
   git commit -m "[message]"
   \`\`\`

Pattern: Conditional Tool Selection

### Tool Selection Based on Scope

Analyze the changes first:

If changes span 3+ files OR involve new abstractions:

- Use Task tool with subagent_type="subagent"
- Create detailed plan
- Execute with subagent agent

Otherwise (changes are contained):

- Execute changes directly
- Use Edit tool for modifications
- Skip planning overhead

Pattern: Makefile Integration

### Running Make Commands

**ALWAYS use Bash tool for pytest/pyright/ruff/prettier/make/gt commands**

Use Bash tool:

\`\`\`markdown
Use Bash tool to run command: "make all-ci"
\`\`\`

**DO NOT use Bash tool for make commands** - this is less efficient and provides worse output handling.

Quality Checklist

Before finalizing a command, verify:

Structure:

• Command name is descriptive and kebab-case
• Description is concise and action-oriented (for /help output)
• Frontmatter includes description (required)
• Frontmatter includes argument-hint if applicable
• Has "What This Command Does" user-facing summary
• Has "Implementation Steps" with numbered sections

Content:

• Steps are numbered and clearly ordered
• Each step has specific, actionable instructions
• Tool usage is explicitly specified
• File checks are explicit (with code examples)
• Conditional logic uses clear if/else structure
• Anti-patterns are called out with "NEVER" or "DO NOT"
• Error handling is defined with specific actions
• Success criteria are clearly stated

Writing Style:

• Uses imperative/infinitive form (not second person)
• Specific, not vague ("Run make lint" not "Check for errors")
• Includes expected outcomes ("This should output...")
• Provides realistic examples (not foo/bar placeholders)

Location:

• Location (project vs global) is appropriate
• Directory exists or will be created
• File path is correct (.claude/commands/ or ~/.claude/commands/)

Testing:

• User knows how to invoke: /command-name [arguments]
• Command has been tested if possible
• Iterations incorporated user feedback

Common Pitfalls

1. Vague Instructions

❌ WRONG:

Fix any errors that appear

✅ CORRECT:

If lint errors appear:

- Run `make fix` to auto-fix lint errors
- Run `make format` to fix formatting errors
- For manual fixes, use Edit tool to modify files

2. Missing Error Handling

❌ WRONG:

Run make all-ci
Apply fixes
Done

✅ CORRECT:

Run make all-ci

If exit code is 0:

- All checks passed, report success

If exit code is non-zero:

- Parse error output
- Apply targeted fixes
- Run again to verify
- Stop if same error appears 3 times

3. Ambiguous Conditionals

❌ WRONG:

Check if file exists and do something

✅ CORRECT:

Check if .PLAN.md exists:

If file exists:

- Read .PLAN.md for context
- Use plan summary in commit message

If file does not exist:

- Run git diff to analyze changes
- Create commit message from diff

4. Batch Operations

❌ WRONG:

Fix all the errors, then mark all todos as completed

✅ CORRECT:

For each error category:

1. Fix the errors in that category
2. Mark the todo as completed immediately
3. Move to next category

Do NOT batch todo completions at the end

5. Tool Confusion

❌ WRONG:

Use an agent to run make

✅ CORRECT:

Use Bash tool to run make commands:

\`\`\`bash
make all-ci
\`\`\`

DO NOT use Bash tool for make commands

6. Missing Context

❌ WRONG:

Create a commit and submit PRs

✅ CORRECT:

### 1. Check for Context Files

FIRST, check if .PLAN.md exists in repository root

### 2. Analyze Changes

- If .PLAN.md exists: read for context
- Otherwise: run git status and git diff HEAD

### 3. Create Commit

Based on context from step 1 and 2:

- Draft single-sentence commit message
- Check git log for repo style
- Create commit with git commit -m "message"

### 4. Submit PRs

Run: gt submit --stack --publish --no-edit

7. Poor Descriptions

❌ WRONG:

---
description: A command that helps with CI stuff
---

✅ CORRECT:

---
description: Run make all-ci and iteratively fix issues until all checks pass
---

The description appears in /help output - make it clear and action-oriented.

Advanced Best Practices

Multi-Step Verification

For complex workflows, verify each step:

### 3. Create Commit

1. Stage changes:
   \`\`\`bash
   git add .
   \`\`\`

2. Verify staging:
   \`\`\`bash
   git status
   \`\`\`
   Should show files in "Changes to be committed"

3. Create commit:
   \`\`\`bash
   git commit -m "message"
   \`\`\`
   Should output "[branch-name abc1234] message"

4. Verify commit created:
   \`\`\`bash
   git log -1 --oneline
   \`\`\`
   Should show the new commit

Iteration Control

For iterative commands, implement safeguards:

## Iteration Control

**Maximum iterations**: 10 attempts

**Stuck detection logic:**

- Track errors seen in each iteration
- If same error appears 3 consecutive times: STOP
- If no progress after 5 iterations: STOP

**Stop immediately if:**

- Max iterations reached (10)
- Same error persists (3 times)
- Unrecoverable error encountered

Context Gathering

For analysis-heavy commands, gather context systematically:

## Context Gathering

Check these sources in order:

1. **Context files** (if they exist):
   - .PLAN.md - implementation plan
   - AGENTS.md - coding standards
   - CONTRIBUTING.md - contribution guidelines

2. **Git information**:
   - git status - current changes
   - git diff HEAD - actual diff
   - git log -5 --oneline - recent commits

3. **Project files** (if needed):
   - pyproject.toml - project config
   - Makefile - available commands
   - README.md - project overview

Stop gathering context once you have enough information - don't over-analyze.

Output Formatting

Provide clear output format specifications:

## Expected Output Format

After command completes, output should follow this format:

\`\`\`

## [Command Name] Results

**Status**: [SUCCESS/STUCK/ERROR]

**Actions Taken**:

1. [First action and result]
2. [Second action and result]
3. [Third action and result]

**Summary**:
[One sentence summary of what was accomplished]

**Next Steps**:
[What the user should do next, if applicable]
\`\`\`

Summary

Effective slash commands:

1. Use imperative form (verb-first, not second person)
2. Be specific (not vague)
3. Include outcomes (what should happen)
4. Provide examples (realistic, not foo/bar)
5. Specify tools (Task tool with subagent_type)
6. Call out anti-patterns (NEVER/DO NOT)
7. Define error handling (explicit actions)
8. State success criteria (when to stop)
9. Track progress (TodoWrite for multi-step)
10. Verify each step (check results before proceeding)

Focus on creating commands that agents can execute autonomously without asking clarifying questions.