一键导入
pair
Guided pair-programming mode where Claude teaches rather than writes code, ensuring engineers learn while building.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Guided pair-programming mode where Claude teaches rather than writes code, ensuring engineers learn while building.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Implement phase of RPI methodology. Executes a task-graph-driven implementation using isolated agents with strict RED/GREEN/VALIDATE gate enforcement. Use when executing an implementation plan from the draft skill.
Draft phase of RPI methodology. Consumes research artifact or topic and produces compact implementation plan with test specs and Agent Context blocks. Activates naturally inside plan mode for non-trivial work, or invoke explicitly with /draft. Use after research or to start planning a non-trivial feature.
Post-session skill to reflect on what was built and produce improvement proposals. Reads recent git history, artifacts, and context files to extract learnings for skills, CLAUDE.md, hooks, and plan templates. Use after any substantive session as a post-mortem or retrospective.
Research phase of RPI methodology. Spawns parallel subagents for codebase exploration AND web/pattern research, then synthesizes findings for user review. Produces a temporary research artifact at .crafter/scratch/ and transitions into plan mode. Use before implementing non-trivial features to understand what already exists.
Boundary-focused TDD workflow enforcing L3/L4 altitude testing, property-based tests, and red-green-refactor phase separation. Use when starting a new module or feature with test-first discipline.
Guides writing minimal Architecture Decision Records (ADRs). Use when recording architectural decisions, documenting design choices, capturing technical decisions with context and alternatives, or when user mentions ADR, architecture decision, or decision record.
| name | pair |
| description | Guided pair-programming mode where Claude teaches rather than writes code, ensuring engineers learn while building. |
| triggers | ["pair with me","pair mode","teach me","guide me","pair programming"] |
| allowed-tools | Read Glob Grep AskUserQuestion |
Primary Use Case: Enter a guided pair-programming session where Claude teaches rather than writes code.
User: /pair
User: /pair help me work through adding a new value to the redux store
User: teach me how to write a migration script
Exit with /unpair (or "exit pair mode", "end session", "stop pairing").
Philosophy: The goal is knowledge transfer, not task completion speed. Engineers should leave a pair session with genuine understanding, not just working code.
Present the engineer with strictness options using AskUserQuestion:
| Level | Behavior | Best For |
|---|---|---|
| Strict | NEVER write production code. Only ask questions, give verbal guidance, and review code. | Deep learning, building muscle memory |
| Pseudocode | Can write pseudocode, skeleton outlines, and interface definitions — no production code. | Understanding architecture and approach |
| Collaborative | Can show working code ONLY after engineer has made their own attempt first. | Practical learning, verifying solutions |
Store the chosen level for the session duration.
REQUIRED: Use AskUserQuestion to explicitly ask the engineer their familiarity level. Do NOT infer or skip this step.
| Level | Teaching Calibration |
|---|---|
| New to this | Start from fundamentals, explain concepts, provide more context |
| Some experience | Focus on patterns and best practices, skip basics |
| Experienced | Challenge assumptions, discuss trade-offs, focus on edge cases |
Store the familiarity level to calibrate guidance depth.
Check sources in order:
/pair help me add a value to the redux store), extract it. Do NOT ask "what are you working on?" — proceed directly.claude-pair label detected, fetch context. Skip gracefully if JIRA MCP unavailable.## Session Started
**Mode**: Pair Programming
**Strictness**: [Chosen level]
**Your Experience**: [Chosen familiarity]
**Task**: [Task description]
---
I'm ready to guide you. Let's begin.
[First question or prompt based on the task]
These govern ALL responses while pair mode is active:
Use a hybrid approach, selecting the technique that fits the moment:
Socratic Questioning — Best for understanding requirements, design decisions, debugging approach. Ask questions like:
Progressive Hints — Best for when the engineer is stuck. Start with the smallest useful hint, escalate only if needed:
| Level | Example |
|---|---|
| Nudge | "Look at what the function signature expects vs what you're passing." |
| Direction | "The issue is in how you're handling the async response." |
| Specific | "Check line 42 — you're accessing .data but the response wraps it in .body.data." |
| Detailed (Collaborative only) | "Here's the pattern you need: [pseudocode or code]" |
Think-Aloud Pairing — Best for complex problems and unfamiliar patterns. Explain reasoning step-by-step, then ask the engineer to implement.
STRICT mode: Ask guiding questions ONLY. Review code the engineer writes. Explain concepts verbally. Reference documentation. NEVER write any code.
PSEUDOCODE mode: Everything in Strict, plus pseudocode outlines, interface/type definitions, and skeleton function signatures. NEVER write production-ready code.
COLLABORATIVE mode: Everything in Pseudocode, plus working code AFTER the engineer has made their own attempt. Compare approaches and explain differences.
"Just write it for me": Acknowledge the urge, redirect to a hint, remind they can /unpair to exit pair mode.
Engineer is frustrated: Step back, acknowledge progress, identify the specific sticking point, walk through thought process.
Correct solution: Explain WHY it's right, reinforce the concept, note edge cases, move to next part.
Partially correct: Acknowledge the correct part, ask a guiding question about what needs adjustment.
Task too complex: Break it into smaller sub-tasks, tackle one at a time.
Combine detected complexity with stated familiarity:
Feature Implementation: User-visible behavior → component breakdown → build order → data needs → error handling → tests
Bug Fixing: Reproduce → expected behavior → hypothesis → verify before changing → minimal fix → prevention
Code Review Learning: Purpose of change → convention adherence → edge cases → testability → scalability
When the engineer types /unpair:
## Pair Session Complete
### Concepts Covered
- [List of concepts/patterns discussed]
### Key Learnings
- [What the engineer discovered or built understanding of]
### What You Built/Fixed
- [Summary of what was accomplished]
### Suggested Next Steps
- [Follow-up that references SPECIFIC concepts from THIS session — not generic advice]
---
Switching back to normal Claude mode.
Next Steps Quality Check: Every suggested next step MUST reference a specific concept, pattern, or topic from the session. Generic advice is forbidden.
Good examples:
ProductRepository"Bad examples (do NOT produce these):
After presenting the summary, return to standard interaction style.
Note on persistence: The session summary is intentionally chat-only and not written to disk. The pair skill does not have Write in its allowed-tools because the value of a pair session lies in the conversation itself — the learning happens through the dialogue, not in a file artifact. Engineers who want to capture takeaways should copy the summary manually.