Auto Experiment
自动化超参数/架构搜索循环。给定模型代码、数据集和优化目标，自动生成 train.py，经双 agent 审计门禁通过后，循环调用 DeepSeek 提出改进建议、训练、用 git 追踪每轮变更，保留最优配置和 checkpoint。
Install
npx skills add zxl696/smallmodel-auto-experiment-skill --skill auto-experimentgit clone --depth 1 https://github.com/zxl696/smallmodel-auto-experiment-skill.git /tmp/smallmodel-auto-experiment-skill
mkdir -p ~/.claude/skills
cp -r /tmp/smallmodel-auto-experiment-skill/. ~/.claude/skills/auto-experimentPersonal skills live in ~/.claude/skills. Use .claude/skills in a project folder to share a skill with your team.
From the SKILL.md
# auto-experiment Skill
You are running the **auto-experiment** skill. Your job is to orchestrate an automated, git-tracked hyperparameter / architecture search loop using DeepSeek as the analysis engine.
The skill directory is located at: `~/.claude/skills/auto-experiment/`
(Expand `~` to the actual home path when constructing bash commands.)
---
## PHASE 1 — Collect Experiment Configuration
Ask the user for the following. Store all answers in memory for this session.
**AskUserQuestion 使用规范（重要）：**
每次需要用户提供具体内容（路径、代码、数值、列表）时，选项中**必须包含一个"请在 Other 中输入…"的选项**，让用户知道在哪里填写。格式固定为：
```
{ label: "请在 Other 中输入…", description: "（点击 Other 输入具体内容）" }
```
这个选项本身不会被选中——它的作用是引导用户点击 Other 并输入文本。对于纯选择题（无需自由输入）则不需要此选项。
---
### 1.0 任务描述（最先询问，最重要）
**这是整个实验的上下文基础，必须第一个问，后续所有决策都依赖此信息。**
用 AskUserQuestion 询问：
```
question: "请描述本次实验的目标任务"
options:
- label: "请在 Other 中输入任务描述", description: "请说明：① 任务目标是什么（预测/分类/检测/生成…）② 模型是什么（名称+简要结构）③ 数据是什么（格式/来源/规模）④ 希望优化什么（loss/acc/F1…）"
```
用户需要描述清楚以下四点，如有遗漏则追问：
1. **任务目标**：做什么？（如：预测未来24步的电力负荷 / 对心电图做异常检测 / 图像分类）
2. **模型**：用什么模型？（如：DLinear / ResNet / 自定义Transformer）
3. **数据**：数据是什么？（如：ETTh1.csv，7个通道，8640行 / ImageNet子集，10类）
4. **优化目标**：希望最小化/最大化什么指标？（如：val MSE越低越好 / test F1越高越好）
将用户的完整描述存为 `TASK_DESCRIPTION`，在后续 DeepSeek prompt 的开头始终附上，让 DeepSeek 清楚实验背景。
---
### 1.1 Model code
用 AskUserQuestion 询问：Details
- Skill name
auto-experiment- Repository
- zxl696/smallmodel-auto-experiment-skill
- Repo updated
- 21 March 2026
FAQ
What does the Auto Experiment skill do?
自动化超参数/架构搜索循环。给定模型代码、数据集和优化目标，自动生成 train.py，经双 agent 审计门禁通过后，循环调用 DeepSeek 提出改进建议、训练、用 git 追踪每轮变更，保留最优配置和 checkpoint。
How do I install the Auto Experiment skill?
skills CLI (Claude Code, Cursor, Codex and more): npx skills add zxl696/smallmodel-auto-experiment-skill --skill auto-experiment. Skills go in ~/.claude/skills (personal) or .claude/skills (project), and Claude loads them automatically when a task matches the description.
Is Auto Experiment popular?
It lives in zxl696/smallmodel-auto-experiment-skill, which has 5 GitHub stars and 1 forks.
Related skills
Gitlab Skill
5ekohe/gitlab-skill
GitLab integration for fetching issues/MRs, generating executive summaries, performing comprehensive code reviews, and automating issue resolution workflows. Use when working with GitLab issues, merge requests, or when asked to summarize, review, comment on, or resolve GitLab work items.
Blitz
6ADM-NKH/claude-blitz
A continuously-firing background work runner for Claude Code. Queue real tasks ("/blitz add"), track long-term goals ("/blitz goal add"), and enable per-project audit rotations ("/blitz audit enable"). Blitz fires on a configurable cadence with a per-project idle gate, decomposes goals into increments on a long-lived branch, and persists every fire's output to ~/blitz/runs/<timestamp>/. Includes commands to manage the backlog, goals, audits, schedules, and kill switches (off/on/skip).
Shipreel
3theBstar/shipreel
Turn a pull request into a short narrated walkthrough video (about two and a half minutes) with animated diagrams, code panels, optional screen recordings of the real app, captions, crossfade cuts and an ambient music bed. Runs locally with no paid services by default. Use when asked for a PR video, PR walkthrough, demo video of a change, or "make a shipreel".
Worktree Closeout
6leonsong09/worktree-closeout
Use when the user asks for worktree closeout, branch closeout, parallel closeout, worktree janitor, 工作树收口, 分支收口, or 并行收口, especially when work spans multiple Codex sessions and you need a date-based scan of open worktrees, closeout status, suggested merge/prune order, and prompt-ready handoff instructions.
Dialog Tree
3ikotelkin/claude-skills
Maintain a visual "dialog tree" — a data file plus a self-contained HTML viewer in the project repo that map the branches of a long conversation (user questions → the steps/variants of the answers) as an interactive tree with resolve/delete marks and a "built up to" marker for incremental updates. Use whenever the user asks to "create/set up a dialog tree", "update the tree", "add this to the tree", "show open branches", or mentions dialog-tree.html — and proactively offer it when a deep-dive discussion branches enough that the user risks losing track of unexplored side-threads.
Limit Up Dao
6laozdao/limit-up-dao
老子道涨停股批量归因分析Skill。基于开源大模型，输入日期自动获取当日所有涨停股票数据，逐一输出多维度归因分析报告。GitHub https://github.com/laozdao/limit-up-dao。触发场景：用户要求分析某日涨停股、批量涨停归因、今日涨停复盘、涨停板原因分析、连板股分析等。
Browse the most popular Claude skills, all git & code review skills, or write your own with the skills generator.