Preflight

Think before you build. A universal protocol that makes AI coding agents design before they code.

Website · Protocol Spec · v1.0.0

Preflight defines 5 mandatory phases every agent completes before writing code: Orient, Clarify, Approach, Confirm, Execute. It works with any AI coding agent — no runtime, no API key, one file.

Why

AI coding agents jump straight to implementation. They write code before reading the codebase, skip exploring alternatives, and don't confirm intent. This wastes tokens, produces wrong solutions, and creates rework.

Preflight fixes this with a lightweight prompt-based protocol that works across every major agent driver.

Quick Start

Auto-detect and install

bash <(curl -fsSL https://raw.githubusercontent.com/AgentGuardHQ/preflight/main/install.sh)

The installer detects which AI agent drivers you use and copies the appropriate config file.

Install for a specific driver

bash <(curl -fsSL https://raw.githubusercontent.com/AgentGuardHQ/preflight/main/install.sh) --driver claude-code

Manual install

Copy the driver file for your agent into your project:

Driver	Copy this file	To this location
Claude Code	drivers/claude-code/preflight.md	.claude/commands/preflight.md
Codex (OpenAI)	drivers/codex/AGENTS.md	Append to AGENTS.md
Gemini CLI	drivers/gemini/GEMINI.md	Append to GEMINI.md
Goose	drivers/goose/.goosehints	Append to .goosehints
GitHub Copilot	drivers/copilot/.github/copilot-instructions.md	Append to .github/copilot-instructions.md
Cursor	drivers/cursor/.cursor/rules/preflight.mdc	.cursor/rules/preflight.mdc

The Protocol

See PROTOCOL.md for the full specification.

The 5 Phases

Phase	What the agent does
1. Orient	Read relevant files, docs, recent changes. Understand what exists.
2. Clarify	Ask questions until purpose, constraints, and success criteria are clear.
3. Approach	Propose 2-3 approaches with trade-offs and a recommendation.
4. Confirm	Present the plan. No code until approved.
5. Execute	Write the code.

Core Rules

• Phases 1-4 produce zero code changes. Reading is encouraged, writing is not.
• Task-scoped. Each new task triggers a fresh pass. Continuing the same task doesn't re-trigger.
• Scales down, never off. Even a one-line fix runs all 5 phases — each can be one sentence.

Examples

• Simple task — bug fix with minimal Preflight (~30 seconds)
• Complex task — feature with full design flow
• Autonomous agent — Tier C agent self-confirming against a ticket

Adding a New Driver

See CONTRIBUTING.md for how to add support for additional AI coding agents.

Part of the Governed Swarm Platform

Repo	Role
AgentGuard	Governance — policy enforcement, telemetry, invariants
Octi Pulpo	Coordination — pipeline controller, model routing
ShellForge	Orchestration — multi-runtime agent execution
Preflight	Protocol — universal design-before-you-build standard
Extensions	Drivers, integrations, policies, example swarms

License

MIT — see LICENSE.

---

Created by the AgentGuard team.