> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/coollabsio/jean/llms.txt
> Use this file to discover all available pages before exploring further.

# Custom Prompts

> Customize AI prompts and system instructions in Jean

## Overview

Jean includes customizable "magic prompts" that power AI features like issue investigation, code review, and PR generation. You can modify these prompts to match your workflow and coding standards.

## Available Magic Prompts

### Investigation Prompts

#### `investigate_issue`

Prompt for investigating GitHub issues.

**Default behavior**:

1. Read issue context files
2. Analyze problem (expected vs actual behavior)
3. Explore codebase for relevant code
4. Identify root cause
5. Check for regressions
6. Propose solution with specific files

**Variables available**:

* `{issueWord}` - "issue" or "issues" (plural)
* `{issueRefs}` - Issue numbers (e.g., "#123")

#### `investigate_pr`

Prompt for investigating GitHub pull requests.

**Default behavior**:

1. Read PR context and reviews
2. Understand changes and branches
3. Analyze approach
4. Security review (malicious code, backdoors, secrets)
5. Identify action items from reviewers
6. Propose next steps

**Variables available**:

* `{prWord}` - "PR" or "PRs" (plural)
* `{prRefs}` - PR numbers (e.g., "#456")

#### `investigate_workflow_run`

Prompt for investigating failed GitHub Actions workflows.

**Variables available**:

* `{workflowName}` - Workflow name
* `{branch}` - Branch name
* `{displayTitle}` - Commit/PR title
* `{runUrl}` - Workflow run URL
* `{runId}` - Workflow run ID

#### `investigate_security_alert`

Prompt for investigating Dependabot vulnerability alerts.

**Variables available**:

* `{alertWord}` - "alert" or "alerts"
* `{alertRefs}` - Alert identifiers

#### `investigate_advisory`

Prompt for investigating repository security advisories.

**Variables available**:

* `{advisoryWord}` - "advisory" or "advisories"
* `{advisoryRefs}` - Advisory identifiers

#### `investigate_linear_issue`

Prompt for investigating Linear issues.

**Variables available**:

* `{linearWord}` - "issue" or "issues"
* `{linearRefs}` - Linear issue IDs
* `{linearContext}` - Full issue context (description + comments)

### Generation Prompts

#### `pr_content`

Generates PR title and description from commits and diff.

**Variables available**:

* `{current_branch}` - Source branch
* `{target_branch}` - Target branch
* `{commit_count}` - Number of commits
* `{context}` - Related context (issue, etc.)
* `{commits}` - Commit history
* `{diff}` - Full diff

#### `commit_message`

Generates commit message from staged changes.

**Variables available**:

* `{status}` - Git status output
* `{diff}` - Staged diff
* `{recent_commits}` - Recent commit messages
* `{remote_info}` - Remote repository info

#### `release_notes`

Generates release notes from commits since last tag.

**Variables available**:

* `{tag}` - Previous release tag
* `{previous_release_name}` - Previous release name
* `{commits}` - Commits since tag

#### `session_naming`

Generates short session names from first user message.

**Variables available**:

* `{message}` - User's first message

**Rules**:

* Maximum 4-5 words
* Sentence case (only capitalize first word)
* No special characters
* No generic names
* No commit-style prefixes ("Add", "Fix", etc.)

### Analysis Prompts

#### `code_review`

Provides structured code review feedback.

**Variables available**:

* `{branch_info}` - Current branch info
* `{commits}` - Commit history
* `{diff}` - Branch diff
* `{uncommitted_section}` - Uncommitted changes section

**Focus areas**:

* Security & supply-chain risks
* Performance issues
* Code quality and maintainability
* Potential bugs
* Best practices violations

#### `context_summary`

Summarizes conversation for future context loading.

**Variables available**:

* `{project_name}` - Project name
* `{date}` - Current date
* `{conversation}` - Full conversation

**Output format**:

1. Main Goal
2. Key Decisions & Rationale
3. Trade-offs Considered
4. Problems Solved
5. Current State
6. Unresolved Questions
7. Key Files & Patterns
8. Next Steps

#### `session_recap`

Generates brief session digest when returning to unfocused sessions.

**Variables available**:

* `{conversation}` - Conversation transcript

**Output fields**:

* `chat_summary` - One sentence (max 100 chars)
* `last_action` - Last completed action (max 200 chars)

### Workflow Prompts

#### `resolve_conflicts`

Guides through resolving git merge conflicts.

**Appended to conflict resolution messages.**

#### `global_system_prompt`

Global system prompt appended to every chat session.

**Similar to** `~/.claude/CLAUDE.md` for Claude CLI.

**Default includes**:

* Plan mode defaults
* Subagent strategy
* Self-improvement loop
* Verification before done
* Demand elegance
* Autonomous bug fixing

#### `parallel_execution`

System prompt encouraging parallel sub-agent execution.

**Applied when** `parallel_execution_prompt_enabled` is true in preferences.

## Customizing Prompts

### Via Preferences UI

1. Open **Preferences** (`Cmd/Ctrl + ,`)
2. Navigate to **Magic Prompts** tab
3. Select prompt to customize
4. Click **Edit** button
5. Modify prompt text
6. Click **Save**

### Resetting to Defaults

To restore original prompt:

1. Open prompt editor
2. Click **Reset to Default**
3. Confirm reset

### Prompt Customization Tips

**Variables**: Use `{variable_name}` syntax for dynamic values

**Structure**: Follow XML-like tags for clarity:

```xml theme={null}
<task>Your main instruction</task>

<instructions>
1. Step one
2. Step two
</instructions>

<guidelines>
- Guideline one
- Guideline two
</guidelines>
```

**Output Format**: Specify desired output format:

```xml theme={null}
<output_format>
Respond with ONLY the raw JSON object:
{"field": "value"}
</output_format>
```

## Per-Prompt Model Overrides

You can configure which AI model handles each prompt type:

### Default Models (Claude)

```typescript theme={null}
{
  investigate_issue_model: 'opus',              // Heavy task
  investigate_pr_model: 'opus',                 // Heavy task
  investigate_workflow_run_model: 'opus',       // Heavy task
  pr_content_model: 'haiku',                    // Light task
  commit_message_model: 'haiku',                // Light task
  code_review_model: 'haiku',                   // Light task
  context_summary_model: 'opus',                // Heavy task
  resolve_conflicts_model: 'opus',              // Heavy task
  release_notes_model: 'haiku',                 // Light task
  session_naming_model: 'haiku',                // Light task
  session_recap_model: 'haiku',                 // Light task
  investigate_security_alert_model: 'opus',     // Heavy task
  investigate_advisory_model: 'opus',           // Heavy task
  investigate_linear_issue_model: 'opus'        // Heavy task
}
```

### Codex Preset

```typescript theme={null}
{
  investigate_issue_model: 'gpt-5.3-codex',
  pr_content_model: 'gpt-5.1-codex-mini',
  commit_message_model: 'gpt-5.1-codex-mini',
  // ... (top model for heavy tasks, mini for light tasks)
}
```

### Configuring Models

1. Preferences > Magic Prompts
2. Select prompt
3. Choose **Model Override**
4. Select model from dropdown
5. Save changes

## Per-Prompt Backend Overrides

You can also specify which backend (Claude/Codex/OpenCode) handles each prompt:

```typescript theme={null}
{
  investigate_issue_backend: null,  // null = use project/global default
  pr_content_backend: 'claude',     // Force Claude for PR content
  code_review_backend: 'codex',     // Force Codex for reviews
  // ...
}
```

## Per-Prompt Provider Overrides

For Claude CLI, you can override the provider (Anthropic/OpenRouter/etc.):

```typescript theme={null}
{
  investigate_issue_provider: null,        // null = use global default
  commit_message_provider: 'OpenRouter',   // Use OpenRouter profile
  // ...
}
```

## Storage Location

Custom prompts are stored in:

```
~/Library/Application Support/io.coollabs.jean/preferences.json
```

(Location varies by platform - see `~/.config/jean/` on Linux, `%APPDATA%\jean\` on Windows)

## Example: Custom Code Review Prompt

```xml theme={null}
<task>Review the following code changes and provide structured feedback</task>

<branch_info>{branch_info}</branch_info>

<commits>
{commits}
</commits>

<diff>
{diff}
</diff>

{uncommitted_section}

<instructions>
Focus on:
- **Security**: Check for vulnerabilities, hardcoded secrets, injection risks
- **Performance**: Identify bottlenecks, memory leaks, inefficient algorithms
- **Code Quality**: Look for code smells, duplication, unclear naming
- **Testing**: Assess test coverage and edge case handling
- **Documentation**: Verify inline comments and docstrings

Provide:
1. Overall summary (2-3 sentences)
2. Critical issues (must fix before merge)
3. Suggestions (nice-to-have improvements)
4. Praise (highlight good patterns)

Be constructive and specific. Include file paths and line numbers.
</instructions>

<output_format>
Use this structure:

## Summary
[Brief overview]

## Critical Issues
- [Issue 1 with file:line]
- [Issue 2 with file:line]

## Suggestions
- [Suggestion 1]
- [Suggestion 2]

## Praise
- [Good pattern 1]
- [Good pattern 2]
</output_format>
```

## Best Practices

* **Test incrementally**: Make small changes and test before committing
* **Keep structure**: Maintain XML-like structure for clarity
* **Use variables**: Leverage provided variables instead of hardcoding
* **Be specific**: Give clear, actionable instructions
* **Include examples**: Show expected output format
* **Version control**: Back up `preferences.json` before major changes