show-draft

展示 /r2r:draft 的产出——一份经过结构化处理的 ADR 草稿，在研究正式启动之前便已成型。

SKILL.md

--- frontmatter

name: show-draft
description: Show what /r2r:draft produces - a sample ADR draft that structures research before it begins.

Show Draft Example

When triggered, display this sample draft output:

/r2r:draft Output Example

Here's what /r2r:draft "AI-powered code review for PR automation" produces:

RFC DRAFT: AI-Powered Code Review for PR Automation

Status: WIP Author: JD Fiscus Created: 2025-01-20

Problem

Manual code reviews are a bottleneck. Senior engineers spend 4-6 hours/week reviewing PRs. We want to explore whether AI can handle first-pass reviews to catch common issues before human review.

Research Questions

• Which AI models perform best on code review tasks? (GPT-4, Claude, CodeLlama)
• What's the false positive rate for AI-flagged issues?
• How do developers perceive AI review comments?
• What's the cost per PR at scale (1000+ PRs/month)?
• Can we integrate with existing GitHub Actions workflow?

Options to Evaluate

•OpenAI GPT-4: Highest capability, highest cost, API-based
•Anthropic Claude: Strong reasoning, good at explanations, API-based
•Self-hosted CodeLlama: Lower cost at scale, requires infrastructure
•GitHub Copilot for PRs: Native integration, limited customization

Solution

[To be filled after research]

Rationale

[To be filled after research]

Scope

In Scope

•Automated first-pass review on new PRs
•Common issue detection (security, performance, style)
•Integration with existing CI/CD

Out of Scope

•Replacing human reviewers entirely
•Architectural review
•Business logic validation

Success Criteria

•Reduce senior engineer review time by 30%
•< 10% false positive rate on flagged issues
•Developer satisfaction score > 7/10
•Cost < $0.50 per PR reviewed

Stakeholders

Stakeholder	Interest	Engagement
Engineering	Time savings	Decides
Security	Vulnerability detection	Consulted
Finance	Cost implications	Informed

Risks

•Developers may ignore AI comments (alert fatigue)
•False positives erode trust
•Cost could exceed manual review at scale
•Security/privacy concerns with code leaving org

Next Steps

• Research model options (1 week)
• Build POC with top 2 models (2 weeks)
• Run pilot on internal repos (2 weeks)
• Run /r2r:assess on findings
• Make go/no-go decision

This creates structure BEFORE research begins. After completing the research plan, run /r2r:assess to evaluate findings and continue through the pipeline.

The full /r2r:draft command:

•Asks clarifying questions about your research topic
•Lets you pick a template style (Google/Uber/Sourcegraph/Experiment)
•Generates a structured draft with research plan
•Saves to .r2r/00-draft-adr.md