PRE Engineering

Set up a multi-agent collaborative framework where three AI roles — Planner, Executor, Reviewer — coordinate through a shared collaboration log to continuously deliver project goals.

PRE is not a standalone project — it's a framework you add to an existing project to enable multi-agent collaboration. Think of it as giving your project a "team of AI workers" with clear role divisions and a shared communication channel.

Installation & Updating

Install: npx skills add zhangzhanluo/pre-engineering (recommended) or manually place skills/pre-engineering/ into your AI tool's skills directory.

Update: Re-run the install command to get the latest version. Existing .pre/{project_name}/ collaboration documents in your projects are unaffected — they're separate from the skill templates.

Step 0: Language Detection & Confirmation

Before collecting requirements, detect the user's language and confirm:

1. Detection rule: If user's input contains Chinese characters → default to zh; otherwise → default to en
2. Confirm with user — present an interactive prompt offering language options.

Set lang variable: zh or en. This determines template path and document file names:

Step 1: Collect Project Name + All Requirements

First collect a project name (used as subdirectory under .pre/{project_name}/ for multi-project support), then collect all requirements.

Collect project information via an interactive prompt: project name, overview, features, technical constraints, special notes.

Set project_name variable from user input. All documents will be placed in .pre/{project_name}/.

Information Inference Rules:

• Auto-scan README.md, design docs, package.json, requirements.txt, etc. to extract overview and tech stack — this saves the user from repeating what's already in their project
• If project directory has structured requirements, extract feature list directly
• Present inferred results to user for confirmation or modification

After collection, refine requirements: remove duplicates, normalize descriptions, supplement missing dimensions.

Step 2: Confirm Draft & Generate Documents

Synthesize collected information into a project goals document draft, present to user for final confirmation. Upon confirmation, generate all 5 core documents.

The user must be satisfied with the project goals document before proceeding — this document drives all subsequent agent work, so getting it right matters.

Project Goals Document

Write to .pre/{project_name}/{goals_file} upon confirmation.

For English projects:

# Project Goals

## Overview
{Project overview provided by the user}

## Feature Requirements
- {Feature requirements provided by the user, one per line}

## Technical Constraints
- {Technical constraints provided by the user, one per line}

## Notes
- {Special notes provided by the user, one per line}

For Chinese projects, adapt the section titles and content to Chinese while maintaining the same structure.

Important: This document is the driving core of the entire project. Agents can only read it — they must not modify it. Only the human user can change project direction by editing this file.

Collaboration Log Initial Entry

Write to .pre/{project_name}/{log_file} creating the initial entry with proper language (Chinese for lang=zh, English for lang=en).

Time format: [YYYY-MM-DD HH:MM], must execute date +"%Y-%m-%d %H:%M". Timezone: Confirm with user during initialization, record in project goals Notes.

The initial entry declares PLN_WAIT status, which triggers the Planner to start the first planning cycle — this is how the project begins.

Agent Guide Document Generation

Generate three agent guide documents from templates in references/{lang}/:

Generation steps:

1. Read the corresponding template file from references/{lang}/
2. Replace all placeholders with project-specific values:
   • {PROJECT_NAME} → project name collected in Step 1
   • {PROJECT_ROOT} → absolute path to the project root directory (e.g., /home/user/my-project)
   • If user specified a non-default code directory (e.g., src/), additionally replace {PROJECT_ROOT} in the Project Code Path section only with {PROJECT_ROOT}/{custom_dir} (e.g., {PROJECT_ROOT}/src)
3. Write to the corresponding guide document in .pre/{project_name}/
4. Check for existing files before overwriting

Startup Instructions

After all documents are generated, present startup instructions in the selected language (Chinese for lang=zh, English for lang=en).

For English Projects (lang=en):

PRE Engineering initialization complete! Files generated:

• .pre/{project_name}/{goals_file} — Project requirements (only humans can modify)
• .pre/{project_name}/{log_file} — Initial PLN_WAIT entry
• .pre/{project_name}/{planner_guide} — Planner guide
• .pre/{project_name}/{executor_guide} — Executor guide
• .pre/{project_name}/{reviewer_guide} — Reviewer guide

IMPORTANT: Git baseline setup

.pre/ is excluded via .gitignore by default. This ensures agents can always access collaboration documents on disk. Two options:

• Default: .pre/ excluded from git — collaboration documents are local-only, agents access them directly on disk. No baseline commit needed.
• Track in git: If user wants to version the collaboration documents, remove .pre/ from .gitignore and commit baseline:

cd {project_directory_path}
git add .pre/{project_name}/
git commit -m "PRE initialization: collaboration documents baseline"

Launch agents (3 separate terminals/sessions):

1. Planner (strong reasoning model): /loop 3m "Read {PROJECT_ROOT}/.pre/{project_name}/{planner_guide} and follow its instructions as the Planner role. Each cycle starts by reading {PROJECT_ROOT}/.pre/{project_name}/{log_file} to check the latest status."

2. Executor (fast coding model): /loop 3m "Read {PROJECT_ROOT}/.pre/{project_name}/{executor_guide} and follow its instructions as the Executor role. Each cycle starts by reading {PROJECT_ROOT}/.pre/{project_name}/{log_file} to check the latest status."

3. Reviewer (thorough review model): /loop 3m "Read {PROJECT_ROOT}/.pre/{project_name}/{reviewer_guide} and follow its instructions as the Reviewer role. Each cycle starts by reading {PROJECT_ROOT}/.pre/{project_name}/{log_file} to check the latest status."

For Chinese Projects (lang=zh):

Present equivalent instructions in Chinese, with file names and paths adapted accordingly. Refer to references/zh/ template files for localized content.

Record Loop Task IDs: Each /loop returns a job ID. Agents will automatically record their job ID and model name in the collaboration log when they start their first cycle. Users can verify job IDs by checking the collaboration log.

Project control: