Aftermatter — AI skill for Claude Code
Evidence-based review of your AI coding agent's workflow.
How to install Aftermatter
This entry records only its repository, not the path inside it, so there is no
exact command to give. Open Ember452/aftermatter and copy the folder into
~/.claude/skills/, or the file into ~/.claude/agents/.
What Aftermatter does
Evidence-based review of your AI coding agent's workflow. Analyzes Claude Code, Codex & Cursor session logs plus your repo, turns gaps into prioritized findings, and verifies fixes actually stick.
Alternatives in AI
- Uisight — Your AI can already see the screen — it just can't measure it 103 ★
- Artemis — ARTEMIS turns natural-language instructions into reliable Android automation 101 ★
- AI Agent Session Center — Real-time dashboard that turns AI coding agent sessions (Claude Code, Gemini CLI, Codex) into animated 3D robo 82 ★
README
中文 · English
Harness engineering, with evidence.
AfterMatter 读你的 AI 编码 Agent(Claude Code / Codex / Cursor / Qoder)留在本机的会话与仓库证据, 按五维 Agent Work Loop 产出改进发现,**每条断言都能指回原始字节**。它不止给报告:还会起草有界修复, 并用纵向统计验证改进是否真的发生了。
**状态:pre-alpha。** M0 骨架已完成(tag `v0.1.0-m0-skeleton`),**目前没有可运行命令**。 M2 出第一份不依赖 LLM 的报告,M6 起才是 `pip install aftermatter`。
为什么
Agent 改代码很快,薄弱环节往往在围绕它的工作流,而只看最终 diff 会系统性漏掉这些:目标被解决错了、 执行路径无人可复现、"Agent 说测试过了"无从考证、审查与交付检查被绕过、同样的摩擦在下一个任务重演。
面向单次调用的 Agent 可观测(tracing、token 大盘)已经是红海;**面向日常工作流的证据驱动体检 + 统计纵向验证**仍是空白带。AfterMatter 只做后者——它不是 coding agent(不执行开发任务), 不是 tracing 工具,也不做模型能力榜单。
核心机制
**双大脑,同一契约。** 同一个确定性证据内核(零 LLM)向上供两推理前端:大脑①寄生宿主,通过 MCP 只读工具 + SKILL.md 由宿主 Agent 承担推理,零 API 成本;大脑②自带多 Provider 引擎,可在 cron / CI 无人值守运行。两者消费同一份 `EvidenceBundle`、产出同一份 `Finding` schema——推理前端可插拔,契约不可分叉。
**修复 + 纵向验证闭环。** 有界修复起草 → 白名单执行(dry-run 先行、`apply` 需显式确认、写入前有内容 哈希快照可回滚)→ 干预台账 → 在后续**可比较**的 Episode 上做 bootstrap 置信区间与 CUSUM 变点检验。 显著才升 `verified`;不显著就保持 `fixed` 并写明还缺多少样本——"证据不足"是合法输出,不是失败。
**执行级证据,不信转述。** 被声称"已通过"的验证命令会在断网沙箱里复跑(`claimed_ok → replay_ok`); 再用静态影响面(blast radius)× 实际测试覆盖面找**验证空洞**——改核心只跑边缘测试会被抓出来。
**分数不能靠话术膨胀。** 证据状态决定该维度上限:缺失 / 未观测 / 不适用 ≤ 59,Present ≤ 74, Wired ≤ 84,Exercised ≤ 94,Outcome-supported ≤ 100。超过天花板的分数在写盘前被校验器拒绝。
**Local-first 隐私。** 原始会话、prompt 全文、真实路径永不出
Related Skills
Dresscode
Scores the design system across 20 areas with evidence, tests whether AI applies it without guessing, and turn
Marketread
Free market research skill for Claude, Codex, and any AI assistant. It interviews you, mines what your buyers
Autonomous Debugger
Evidence-driven debugging skill for Claude Code and AI coding agents. Investigates codebases, reproduces failu
Deep Ass Research
IDE-agnostic deep-research framework for AI agents: maps a topic wide, commits to depth on purpose, adversaria
Gh Review Loop
Fixes the AI reviewer findings on your PR, including the sibling instances the bot missed. Runs your tests bef
Craftsman Marketplace
Engineering-readiness audits for AI-built apps. A Claude Code and Codex plugin that finds production risks acr
Related Agents
Monk Installer
Install, upgrade, and troubleshoot monk-agent plus Monk CLI/daemon for Claude, Codex, and Cursor users.
Career Gap Analyst
Analyzes a CV or career history against a target role to identify skill gaps, narrative gaps, and positioning
GPT Image Prompter
Prompt engineer for the Codex image generator (gptimage skill, ChatGPT subscription). Turns image ideas — ofte