关于
This Claude Skill runs structured after-action reviews for incidents, launches, or projects to capture timelines, root causes, and lessons learned. It triggers on terms like postmortem, retrospective, or RCA and is designed for blameless analysis after an event concludes. The output focuses on actionable improvements rather than just documentation.
快速安装
Claude Code
推荐npx skills add rampstackco/claude-skills -a claude-code/plugin add https://github.com/rampstackco/claude-skillsgit clone https://github.com/rampstackco/claude-skills.git ~/.claude/skills/after-action-report在 Claude Code 中复制并粘贴此命令以安装该技能
技能文档
After-Action Report
Run a structured retrospective on a launch, incident, or completed project. Produce actionable lessons, not just a document.
This skill is for after-the-fact analysis. For active incident response, use incident-response. For planning launches, use launch-runbook.
When to use
- After any incident (any severity)
- After every major launch
- At the end of a project (sprint retro, quarterly retro, project closeout)
- When a recurring issue has happened enough times to demand investigation
- When a decision didn't work out and the team wants to learn
When NOT to use
- During an active incident (use
incident-response) - For pre-launch planning (use
launch-runbook) - For one-off bug fixes that don't merit broad analysis
Required inputs
- The event being analyzed (incident, launch, project)
- A timeline reconstructed from logs, chat, tickets
- Participant accounts of what they observed and did
- Outcomes and impact (what actually happened to users, the business)
The framework: blameless analysis
The most important principle: blameless. Without it, retrospectives produce hidden information and theatrical lessons rather than real ones.
What blameless means
- Focus on systems, not individuals
- Assume everyone made reasonable decisions given what they knew at the time
- The question is "why was this decision reasonable to make?" not "who screwed up?"
- Fixing the system means the next person in that situation succeeds where this person didn't
What blameless does not mean
- No accountability (action items still have owners)
- No hard truths (sometimes the system is broken in obvious ways)
- No standards (some patterns of failure are individual, not systemic)
- No discomfort (real reflection is uncomfortable)
The framework: 6 sections
A complete AAR covers six sections.
1. Summary
A 2 to 3 paragraph overview. Captures:
- What happened
- Impact (users, business, time)
- Root cause (in plain language)
- Top action items
This is what executives read. Anyone who reads only this section should leave with the most important information.
2. Timeline
A reconstructed timeline of events.
For incidents:
- T-0: Detection
- T+X: Acknowledgment
- T+Y: Severity assessed, IC assigned
- T+Z: Investigation began
- ... mitigation, communication, resolution events
- T+N: Resolution declared
For launches:
- Pre-launch decisions and milestones
- Launch day events
- Post-launch monitoring observations
For projects:
- Major milestones, decisions, pivots
- Both planned and emergent
The timeline is the source of truth. Disagreements about what happened get resolved here.
3. Root cause analysis
What caused this, in plain language.
Use one or both of:
Five whys. Start with the surface symptom. Ask "why?" Repeat 5 times (or until you reach a true root). Each "why" should yield a substantive answer, not a tautology.
Example:
- Why did the site go down? Database connection pool exhausted.
- Why was the pool exhausted? Background job opened too many connections.
- Why did the background job open too many connections? Connection cleanup code didn't run on errors.
- Why didn't cleanup run on errors? Original code review didn't cover error paths.
- Why didn't the review cover error paths? No checklist for error handling in our review process.
The fifth why often reveals the system fix. In this case: improve the review process.
Causal chain. Multiple contributing factors that combined.
- Factor 1: Background job opened too many connections (technical)
- Factor 2: Connection limit was set too low for actual traffic (configuration)
- Factor 3: No alert on connection pool saturation (monitoring)
- Factor 4: Recent traffic doubled without infra capacity review (process)
No single fix addresses the incident. Multiple gaps need attention.
4. Contributing factors
Factors that didn't cause the event but made it worse, or removed safety nets that would have caught it.
- Monitoring gaps
- Documentation gaps
- Process gaps
- Tooling gaps
- Knowledge gaps
A "would have been caught earlier if..." factor.
5. What went well
Real lessons require capturing successes, not just failures.
- What detection worked?
- What response worked?
- What decisions were good?
- What tools or processes performed as expected?
This is not consolation. It's calibration. Things that worked here should be reinforced and replicated.
6. Action items
Specific, owned, dated.
| Action | Owner | Due | Type |
|---|---|---|---|
| Add alert on connection pool saturation | [name] | [date] | Monitoring |
| Add error handling checklist to PR template | [name] | [date] | Process |
| Audit other background jobs for similar issue | [name] | [date] | Code |
Action item criteria:
- Specific. "Improve monitoring" is not actionable. "Add alert on connection pool saturation, threshold 80%, page on-call" is.
- Owned. A name. Not "the team."
- Dated. A real date. Not "soon."
- Sized. Roughly hours, days, or weeks of effort.
- Closeable. Definition of done is clear.
Action items that don't close in their committed timeframe should re-surface in the next AAR. Patterns of unclosed actions point to deeper organizational issues.
Workflow
1. Schedule the AAR
Within 1 to 2 weeks of the event. Long enough that emotions cooled and facts gathered. Short enough that memories are fresh.
For incidents: pre-decided in the response procedure. For launches: schedule on the runbook. For projects: schedule at project closeout.
2. Gather inputs
Before the meeting:
- Reconstructed timeline (often the scribe's notes if there was one)
- Logs, chat transcripts, tickets, incident updates
- Individual accounts from each participant (written, before the meeting)
- Impact data (users affected, duration, revenue impact, etc.)
3. Run the meeting
Typical agenda (60 to 90 minutes):
- Read the summary as drafted (5 min)
- Walk the timeline together. Add corrections. Resolve disagreements. (20 to 30 min)
- Discuss root cause. Use five whys or causal chain. (15 to 20 min)
- Discuss contributing factors. (10 min)
- Discuss what went well. (10 min)
- Identify action items. Owners and dates. (10 min)
A facilitator runs the meeting. Often the IC for an incident, or a project lead for a project. The facilitator is not the scribe.
4. Write the document
Within a few days of the meeting. The full AAR includes all 6 sections.
5. Distribute
Internal: post in a known location. Make searchable. Reference in onboarding.
For high-severity incidents: external summary may be appropriate (status page, customer email, public blog).
6. Track action items
Every action item should be tracked to closure. The next AAR re-surfaces unclosed ones.
Failure patterns
- Skipping the AAR for "small" incidents. Patterns get missed.
- Naming and shaming. Real lessons get hidden when people fear blame.
- Generic action items. "Improve testing" instead of specific testing change.
- Action items that never close. Filed, forgotten. Same incident recurs.
- Theater retrospectives. Going through the motions without genuine reflection.
- Skipping "what went well." Misses calibration on what's working.
- Blame externalized. "Our vendor failed." OK, what's our system for vendor risk?
- Single-person AAR. One person writes the whole thing. Misses other perspectives.
- AAR only for failures. Successful launches deserve AARs too. Lessons from success are valuable.
- Long delays. Memories fade. Conversations cool. Get it done within 2 weeks.
Output format
A markdown document at aar-[date]-[event-name].md.
Structure:
# AAR: [Event name]
**Date of event:** [YYYY-MM-DD]
**AAR date:** [YYYY-MM-DD]
**Severity / scope:** [SEV-1 / Major launch / Project closeout]
**Facilitator:** [Name]
**Participants:** [Names]
## Summary
[2 to 3 paragraphs]
## Impact
- Users affected: [number, segment]
- Duration: [time]
- Revenue / business impact: [if applicable]
## Timeline
[Timestamped events]
## Root cause analysis
[Five whys or causal chain]
## Contributing factors
[List]
## What went well
[List]
## Action items
| Action | Owner | Due | Type | Status |
|---|---|---|---|---|
| | | | | |
## Lessons
[Reflections that don't fit elsewhere. Often the most quotable section.]
If required data is unavailable
This skill's output depends on data, measurements, or tool results it cannot generate on its own. When a required input, tool, or data source is unavailable or unverifiable, the sanctioned output is the deliverable with the gap stated: what was needed, what was actually obtained or verified, and which parts of the output are affected. Fabricating, estimating, or interpolating a required number to complete the deliverable is never sanctioned. A stated gap is a complete answer.
Reference files
references/aar-template.md- Fillable AAR template covering incidents, launches, and projects.
GitHub 仓库
常见问题
什么是 after-action-report Skill?
after-action-report 是一个 Claude Skill,作者为 rampstackco。Skill 将 Claude 按需加载的说明和资源打包,让 Claude 无需额外提示即可执行与 after-action-report 相关的任务。
如何安装 after-action-report?
使用本页的安装命令:将 after-action-report 作为插件添加到 Claude Code,或将其仓库克隆到 skills 目录,然后重启 Claude 以加载该 Skill。
after-action-report 属于哪个分类?
after-action-report 属于元分类。
after-action-report 可以免费使用吗?
可以。after-action-report 已收录在 AIMCP,可免费安装。
相关推荐技能
Content Collections 是一个 TypeScript 优先的构建工具,可将本地 Markdown/MDX 文件转换为类型安全的数据集合。它专为构建博客、文档站和内容密集型 Vite+React 应用而设计,提供基于 Zod 的自动模式验证。该工具涵盖从 Vite 插件配置、MDX 编译到生产环境部署的完整工作流。
这个Claude Skill为开发者提供完整的Polymarket预测市场开发支持,涵盖API调用、交易执行和市场数据分析。关键特性包括实时WebSocket数据流,可监控实时交易、订单和市场动态。开发者可用它构建预测市场应用、实施交易策略并集成实时市场预测功能。
该Skill帮助开发者创建OpenCode插件,用于接入命令、文件、LSP等25+种事件。它提供了插件结构、事件API规范和JavaScript/TypeScript实现模式,适合需要拦截操作、扩展功能或自定义事件处理的场景。开发者可通过它快速构建响应式模块来增强OpenCode AI助手的能力。
SGLang是一个专为LLM设计的高性能推理框架,特别适用于需要结构化输出的场景。它通过RadixAttention前缀缓存技术,在处理JSON、正则表达式、工具调用等具有重复前缀的复杂工作流时,能实现极速生成。如果你正在构建智能体或多轮对话系统,并追求远超vLLM的推理性能,SGLang是理想选择。
