Prompts¶

Built-in prompts¶

Name	Focus	What it checks
`security`	Security vulnerabilities	Auth/authz bypass, privilege escalation, TOCTOU races, path traversal, hardcoded secrets/credentials, insecure defaults, weak crypto, business logic vulns
`logic`	Correctness	Wrong conditionals, missing edge cases (null, empty, boundary), missing error handling, silent failures, thread safety, unreachable code, off-by-one, resource leaks
`design`	Code quality	Misleading names, DRY/KISS/SRP violations, O(n^2) algorithms, N+1 queries, dev deps in production, dead config flags, contract violations, hardcoded OS paths
`docs`	Documentation gaps	New features/flags/APIs without docs, changed behavior not reflected in docs, missing docstrings, undocumented env vars
`common`	All categories in one pass	Combines security + logic + design in a single prompt. Uses 1 agent instead of 3. Good for quick reviews or tight token budgets

common vs security,logic,design: With pydantic backend, --prompts security,logic,design runs 3 parallel agents (one per prompt, results merged). --prompts common runs 1 agent covering everything. Three separate prompts are more thorough; common is cheaper and faster.

Custom prompts¶

Option 1: Add .md files to PROMPTS_DIR and reference by name:

PROMPTS_DIR=~/.junior/prompts junior --prompts security,my_team_rules

Option 2: Pass files directly with --prompt-file:

junior --prompt-file ./rules/api.md --prompt-file ./rules/naming.md

Prompt files use frontmatter format:

---
name: api-standards
description: API design rules for our team
---

You are an expert reviewing REST API code...

How prompts are used per backend¶

Backend	Behavior
`pydantic`	1 parallel AI agent per prompt, results merged. Each agent gets prompt body + project instructions as system prompt. After all agents finish, a summary agent writes a 2-3 sentence overview
`claudecode`	All prompts concatenated into system prompt. Claude reads files via built-in tools (no diff in user message)
`codex`	All prompts concatenated into system prompt. Codex reads files via sandbox
`deepagents`	1 subagent per prompt, LLM orchestrator coordinates

What the LLM sees¶

System prompt¶

Prompt body — from built-in prompts or --prompt-file
Base rules — shared instructions: focus on changed code, be constructive, use request_changes only for critical/multiple high issues
Project instructions — first found file from the repo root: AGENT.md, AGENTS.md, CLAUDE.md. Loaded as-is. Security note: read from the working tree, not target branch — a malicious MR can modify them. See Security

User message¶

Built by build_user_message() from collected context:

MR metadata — title, description, source→target branch, labels
Commit messages — list of commits in the MR
Changed files list — paths with status (added/modified/deleted)
Code diff — full unified diff (included for pydantic, deepagents; omitted for claudecode, codex — they read files via tools instead)
Extra context — from --context and --context-file flags

What is NOT sent in the user message¶

File contents (agents can read files via tools if needed)
Unchanged files (unless the backend explores them — claudecode, pydantic, deepagents all have file tools)
Git history beyond the diff
CI environment variables, API keys, or tokens