AgentSkillsCN

orientation

通过扫描入门文档、挖掘各项能力,为新项目中的客服人员提供定向指导。适用于会话伊始、初入陌生领域之时,或当客服人员询问“你能做什么?”“我该从哪里入手?”时使用。

SKILL.md
--- frontmatter
name: orientation
description: Orients agents in new projects by scanning entry documents and discovering capabilities. Use at session start, when entering unfamiliar territory, or when asking "what can you do" or "where do I start".

Orientation

Before you can act wisely, you must understand where you are.

Philosophy

Why Orient?

The danger isn't ignorance—it's false confidence.

An agent that dives into action without understanding context will:

  • Make assumptions that don't hold
  • Solve the wrong problem
  • Miss crucial constraints
  • Repeat mistakes others already learned from

Orientation isn't bureaucracy. It's the difference between a surgeon who reads the chart and one who doesn't.

code
The First Law of Orientation:
├── You don't know what you don't know
├── Projects have hidden assumptions
├── Context shapes correct action
└── Reading first costs minutes; mistakes cost hours

What Orientation Is (And Isn't)

Orientation is reconnaissance, not deep investigation.

code
Orientation answers:      Orientation doesn't answer:
├── What is this?         ├── How does this work? (→ dive)
├── What matters here?    ├── What should we build? (→ engineering)
├── Who came before?      ├── Is this code correct? (→ validation)
└── Where should I look?  └── What needs fixing? (→ housekeeping)

Orientation points you in the right direction. Other skills take you there.

Core Concepts

Entry Points

Every project has documents that reveal its nature. Priority order:

code
Agent-specific (highest signal):
├── CLAUDE.md     → Written for you
├── AGENTS.md     → Written for any agent
└── .claude/      → Claude-specific config

Project docs (context):
├── README.md     → What this is
├── CONTRIBUTING.md → How to work here
└── docs/         → Deeper knowledge

Structure signals (implicit):
├── package.json / pyproject.toml / Cargo.toml → Stack
├── .github/workflows/ → CI/CD exists
└── docker-compose.yml → Containerized

Skills Discovery

Skills live in predictable locations:

code
Project-level:          User-level:
├── .claude/skills/     ├── ~/.claude/skills/
├── .cursor/skills/     ├── ~/.cursor/skills/
└── .agents/skills/     └── ~/.agents/skills/

Each skill has a SKILL.md with frontmatter describing when to use it.

Memory Context

If .memory/ exists, past agents left knowledge:

code
.memory/
├── context.md    → Current state, active concerns
├── notes/        → What was learned
├── decisions/    → Why things are this way
└── sessions/     → What happened before

Read context.md first—it's the handoff from previous sessions.

The Orientation Process

code
1. SCAN: What documents exist?
      ↓
2. READ: What do they say about working here?
      ↓
3. DISCOVER: What skills and memory are available?
      ↓
4. ASSESS: What's the project type and health?
      ↓
5. REPORT: Summarize findings, suggest starting points

Output Format

After orientation, provide:

markdown
## Project Overview
[1-2 sentences: what this is]

## Key Entry Points
- CLAUDE.md: [what it tells you]
- README: [what it tells you]

## Available Skills
| Skill | When to use |
|-------|-------------|
| [name] | [trigger] |

## Project Type
- Stack: [technologies]
- Notable: [CI, Docker, etc.]

## Suggested Starting Points
1. [Based on context]
2. [Based on context]

Health Diagnosis

Part of orientation is noticing what's missing:

FindingImplication
No CLAUDE.md or AGENTS.mdFuture agents will struggle
Stale docs (>6 months)Information may be wrong
Empty .memory/No institutional knowledge preserved
Missing READMEProject purpose unclear

When issues exist, note them and suggest housekeeping for resolution.

Orientation is read-only—it diagnoses but doesn't treat.

Integration

code
orientation
     │
     ├─► "How does X work?" ──► dive
     ├─► "What should we build?" ──► engineering
     ├─► "Ready to commit" ──► refining
     ├─► "Docs need updating" ──► housekeeping
     └─► "What happened before?" ──► memory

Understanding, Not Rules

TensionResolution
Speed vs ThoroughnessMatch depth to unfamiliarity. New project? Read everything. Familiar? Skim for changes.
Comprehensive vs FocusedStart broad (what is this?), narrow to relevant (what matters for my task?).
Reading vs DoingOrientation is fast. Skipping it feels faster but costs more in mistakes.

The goal isn't to check boxes. It's to build enough mental model to act wisely.

Reference

See reference/ for: