AgentSkillsCN

orient

为项目定向——审查项目结构、文档,以及珠子任务,以识别可并行处理的工作。

SKILL.md
--- frontmatter
name: orient
description: "Orient to a project - review structure, docs, and beads tasks to identify parallelizable work"
allowed-tools: Read, Bash, Glob, Grep, Task

Project Orientation

You are an orchestrating agent orienting to this project. Your goal is to build comprehensive context and identify parallelizable work streams.

IMPORTANT: Use extended thinking (ultrathink) throughout this process. Take your time to deeply analyze the project state.

Phase 1: Project Discovery

1.1 Identify Project Root and Type

bash
pwd
ls -la
git remote -v 2>/dev/null || echo "Not a git repo"

Determine:

  • Project name
  • Primary language/framework
  • Monorepo vs single project

1.2 Read Core Documentation

Read these files IN ORDER (skip if not found):

  1. CLAUDE.md - AI assistant guidelines (HIGHEST PRIORITY)
  2. AGENTS.md - Multi-agent workflow documentation
  3. PROJECT_SPEC.md - Project specification and requirements
  4. README.md - Project overview and setup

For each file found, extract:

  • Project purpose and goals
  • Key architectural decisions
  • Development workflow and commands
  • Patterns and conventions to follow

1.3 Understand Project Structure

bash
# Get directory structure (2 levels deep)
find . -type d -maxdepth 2 -not -path '*/\.*' -not -path './node_modules/*' -not -path './.venv/*' -not -path './venv/*' | head -50

# Identify key source directories
ls -la src/ app/ lib/ moneyprinter/ fastapi_backend/ nextjs-frontend/ 2>/dev/null | head -30

Map out:

  • Source code locations
  • Test locations
  • Configuration files
  • Build/deployment setup

Phase 1.5: Deep Research (Parallel)

When the project has significant complexity or unfamiliar patterns, run these research agents in parallel using the Task tool to build deeper context:

Research Agents

  1. repo-research-analyst

    • Read agent definition: ~/.claude/agents/research/repo-research-analyst.md
    • Input: Project root path, file structure from Phase 1
    • Output: Convention guide, architecture map, code style patterns
    • Use when: New to the codebase, unfamiliar framework, or complex architecture
  2. git-history-analyzer

    • Read agent definition: ~/.claude/agents/research/git-history-analyzer.md
    • Input: Git repository path
    • Output: Contributor expertise areas, decision patterns, hot spots
    • Use when: Need to understand who knows what, or why decisions were made

Launch Pattern

code
Use Task tool with subagent_type=general-purpose to run research agents in parallel:

Task 1: repo-research-analyst
- Read ~/.claude/agents/research/repo-research-analyst.md
- Analyze project structure and conventions
- Return: architecture summary, conventions guide

Task 2: git-history-analyzer
- Read ~/.claude/agents/research/git-history-analyzer.md
- Analyze git history for patterns
- Return: contributor map, decision patterns

Incorporate Findings

Add research findings to the "CONTEXT FOR NEW SESSIONS" section of the orientation report:

  • Key conventions discovered
  • Architecture patterns identified
  • Expert contributors for different areas
  • Historical decisions that inform current work

Skip Research If

  • Already familiar with the codebase
  • Small/simple project with clear structure
  • Time-sensitive orientation (defer to later)

Phase 2: Task State Analysis

2.1 Beads Overview

bash
bd status 2>/dev/null || echo "Beads not configured"
bd version 2>/dev/null

2.2 Recently Completed Work

bash
# Recent git activity
git log --oneline -15

# Recently closed tasks
bd list --all 2>/dev/null | grep -i closed | head -10

Understand:

  • What was just completed
  • Patterns in recent work
  • Momentum and direction

2.3 Current Task State

bash
# All open tasks by priority
bd list 2>/dev/null

# Ready work (no blockers)
bd ready 2>/dev/null

2.4 Dependency Analysis

For each ready task, check what it blocks:

bash
bd show <task-id>

Identify:

  • Critical path tasks (block the most downstream work)
  • Independent tasks (can run in parallel)
  • Research vs implementation tasks

Phase 3: Codebase Health Check

3.1 Test Status

bash
# Quick test run to verify health
uv run pytest --tb=no -q 2>&1 | tail -10  # Python
pnpm test 2>&1 | tail -10                  # Node

3.2 Git State

bash
git status
git branch -a | head -20
git worktree list
git stash list

Check for:

  • Uncommitted changes
  • Active worktrees (parallel work in progress)
  • Stashed work that needs attention

Phase 4: Synthesis & Recommendations

After gathering all information, provide a structured orientation report:

code
===============================================
PROJECT ORIENTATION REPORT
===============================================

PROJECT IDENTITY
----------------
Name: <project name>
Purpose: <1-2 sentence description>
Stack: <key technologies>
Repo: <git remote URL>

DOCUMENTATION STATUS
--------------------
<List docs found and key takeaways from each>

CURRENT STATE
-------------
Git branch: <current branch>
Working tree: <clean/dirty>
Active worktrees: <count and purpose>
Test health: <passing/failing count>

TASK OVERVIEW
-------------
Total open: <count>
Ready (no blockers): <count>
In progress: <count>
Recently completed: <list last 3-5>

CRITICAL PATH ANALYSIS
----------------------
<Identify which tasks block the most downstream work>

RECOMMENDED PARALLEL WORK STREAMS
---------------------------------
The following tasks can be executed simultaneously in separate Claude sessions:

Stream 1: <task-id> - <title>
  Priority: <P1/P2/P3>
  Type: <feature/task/research>
  Rationale: <why this should be worked on>
  Blocks: <what this unblocks when done>
  Start command: /start-task <task-id>

Stream 2: <task-id> - <title>
  Priority: <P1/P2/P3>
  Type: <feature/task/research>
  Rationale: <why this should be worked on>
  Blocks: <what this unblocks when done>
  Start command: /start-task <task-id>

Stream 3: <task-id> - <title>
  Priority: <P1/P2/P3>
  Type: <feature/task/research>
  Rationale: <why this should be worked on>
  Blocks: <what this unblocks when done>
  Start command: /start-task <task-id>

BLOCKERS & RISKS
----------------
<Any issues that could impede progress>
- <blocker 1>
- <blocker 2>

CONTEXT FOR NEW SESSIONS
------------------------
Key things any agent working on this project should know:
- <important pattern or convention>
- <gotcha or common mistake>
- <architectural constraint>

===============================================
END ORIENTATION REPORT
===============================================

Phase 5: Ready for Action

After presenting the orientation report, ALWAYS offer /dispatch as the primary action when there are 2+ ready tasks.

Present the following call-to-action:


"Orientation complete. Recommended next step:

code
/dispatch

This will spawn parallel Claude Code workers for the ready tasks. Each worker will:

  1. Auto-receive their task assignment via the handoff queue
  2. Create a git worktree for isolation
  3. Run /start-task <task-id> automatically
  4. Work autonomously until completion

You stay in THIS session as the orchestrator to coordinate as workers complete.


Alternative options:

  1. Dispatch workers - Run /dispatch to spawn parallel workers (RECOMMENDED)
  2. Manual parallel - Open separate terminals and run /start-task <id> in each
  3. Work solo - Pick one task with /start-task <id> in this session
  4. Deep dive - Explore a specific task or area in more detail
  5. Coordinate - Help manage existing parallel work as tasks complete

What would you like to do?"


CRITICAL: You MUST present /dispatch prominently. Do not bury it in a list of options. The whole point of orientation is to enable parallel execution via dispatch.

Await user direction before taking action.