experiment-forge SKILL.md skillA

experiment-forge from yichen-zju/claude-for-research is an agent skill: a SKILL.md folder an AI agent loads when a task calls for it.

Install it, pinned
mdr add yichen-zju/claude-for-research/experiment-forge@git:20261012.70e1643

Pins git:20261012.70e1643 in mdr.lock. Pin the fingerprint instead.

Audit grade
A, 17 of 17 checks passed
Kind
SKILL.md skill
Repository
yichen-zju/claude-for-research
Path
skills/experiment-forge/SKILL.md
Versions recorded
5
Latest version
git:20261012.70e1643, committed Oct 12, 2026
First version recorded
Oct 3, 2026
Size
5,270 bytes
Repository stars
10
License
none declared

Forge self-contained experiment packages with code, configs and runbooks from a task description; use when launching or packaging structured autonomous experiments.

Collected from public GitHub. Every version is saved exactly as it was and never changes. Install it at an exact version with the mdr tool, and every file is checked before it reaches your agent. The main file is checked against the SHA-256 fingerprint recorded here, and the others against the fingerprints GitHub keeps for its source commit. The automatic checks below grade the latest version, and the same checks always give the same file the same grade.

What the file says

# Experiment Forge

## 任务包结构

```
<task-dir>/
├── prepare.py        # 🔒 锁定:数据准备、tokenizer/dataloader、评估函数。agent 禁止修改
├── train.py          # 🔓 开放:agent 唯一可编辑的文件(模型、优化器、训练循环、超参)
├── program.md        # 🧭 人类可迭代:给 agent 的研究组织指令(用 templates/program.md 生成)
├── pyproject.toml    # 依赖清单 —— agent 禁止新增依赖
└── results.tsv       # 实验台账(表头由 forge 生成,agent 追加行,不提交 git)
```

名字按任务实际调整(比如 `evaluate.py`/`model.py`),但 **锁定/开放/指令 三件套不可缺**。

## 任务契约(Forge 时一次确认;被 orchestrator 调用则视为已授权)

逐项确认**一次完成**(一条消息列全),不逐条追问;被 research-orchestrator 调用时视为上游已授权(BOOTSTRAP 双闸门),直接采用 state 里的 scope/预算/环境,只对**新增实质性选择**(预算扩张、权限变化、评估口径改变)提问。

1. **固定 wall-clock 预算**:每次实验跑固定时长(如 5 分钟/30 分钟),不按 step/epoch —— 否则 agent 换大模型就"作弊"赢了。
2. **架构无关指标**:指标必须对架构改动公平(karpathy 用 val_bpb 而非 loss,因为 loss 随 vocab size 变)。问用户:"换个架构/分词器/特征,这个指标还可比吗?"不可比就换指标。
3. **评估代码锁定**:评估函数、数据加载、随机种子控制放在锁定文件里,是 ground truth。
4. **超时即失败**:超过 ~2× 预算的运行 kill 掉记 `discard`。
5. **VRAM/资源软约束**:允许合理增长,不允许爆炸。
6. **首跑必基线**:第一次运行必须是不改任何代码的 baseline。
7. **简洁性准则**:同等效果更简为赢;0.001 提升加 20 行 hack 不值得。
8. **Guard 防回归**:配一条每轮必须通过的安全命令(如冒烟测试 `pytest -x tests/smoke`)。指标涨但 Guard 挂 → 照样 revert。
9. **min_delta 确认跑**:改进幅度小于阈值时视为噪声 —— 重跑 1-2 次取中位数再定 keep/discard。
…

Read the whole file at its exact version.

How to install

The exact contents, by fingerprint
mdr add yichen-zju/claude-for-research/experiment-forge@sha256:d857857c53dc4001

The line at the top of this page pins the label git:20261012.70e1643, which is what most people want. This one pins the sha256 fingerprint of the contents instead, so it survives a label being moved. Either way the exact version is written to mdr.lock, and mdr install downloads and checks those contents again, so it installs them exactly or stops with an error.

Badge

mdr badge

[![mdr](https://markdownregistry.com/badge/art_zigwroporth77z4n.svg)](https://markdownregistry.com/a/art_zigwroporth77z4n)

2 badge image fetches in 30 days. What the badge shows.

Versions

The registry has recorded 5 versions of this file, between Oct 3, 2026 and Oct 12, 2026, about one every 2 days. Its first recorded version is graded B and its latest A.

versioncommittedcommitsizeaudit
git:20261012.70e1643 latest2026-10-12 70e1643 5,270 BA view · diff
git:20261008.13c7b1e2026-10-08 13c7b1e 5,033 BA view · diff
git:20261007.f635cde2026-10-07 f635cde 4,856 BB view · diff
git:20261003.6bb902c2026-10-03 6bb902c 4,526 BB view · diff
git:20261003.0767d992026-10-03 0767d99 4,279 BB view

Audit of the latest version

A  17 of 17 checks passed. Fixed automatic checks with no AI involved, so the same checks give the same answer every run.
  • pass: Frontmatter block present
  • pass: Frontmatter declares a name
  • pass: Frontmatter declares a description
  • pass: Size between 200 bytes and 200 KB (5270 bytes)
  • pass: No zero-width or bidi control characters
  • pass: No instruction hidden inside an HTML comment
  • pass: No link to an exfiltration or paste host
  • pass: No credential-shaped string
  • pass: No instruction to send local credentials anywhere
  • pass: No text hidden with inline styles
  • pass: No prompt-injection phrasing
  • pass: No curl or wget piped into a shell
  • pass: No recursive delete of root, home or parent
  • pass: No instruction to read or print local credentials
  • pass: No base64 blob over 200 characters
  • pass: No link to a raw IP address
  • pass: No script tag

Source

GitHub

yichen-zju/claude-for-research · 10 stars · license none declared · pushed 2026-10-21 · branch main

API

GET https://markdownregistry.com/api/v1/artifacts/art_zigwroporth77z4n
GET https://markdownregistry.com/api/v1/resolve?ref=yichen-zju/claude-for-research/experiment-forge
GET https://markdownregistry.com/api/v1/blob/d857857c53dc400193176f9c530f02c52ac7b5847006e5c17ba965aaf93300c5

Your AI assistant does the legwork. You hear about the deals worth your word. Hand yours the standing instructions at modelranch.com and it joins the network that reads files like this one.

Guides: the SKILL.md format and its required fields · what to check before you install a skill · how to pin a skill to an exact version · where to find agent skills · keeping a team on the same skills · which agent file does what.

More from yichen-zju/claude-for-research

CLAUDE.md CLAUDE.md file
yichen-zju/claude-for-research · CLAUDE.md
git:20260930.9b33ccf · audit A · 10 stars
CLAUDE.md@claude CLAUDE.md file
yichen-zju/claude-for-research · claude/CLAUDE.md
git:20261014.9f2a5e0 · audit A · 10 stars
evolving-ai-agents SKILL.md skill
yichen-zju/claude-for-research · claude/skills/a-evolve/SKILL.md · Provides guidance for automatically evolving and optimizing AI agents across any domain using LLM-driven evolution…
v1.0.0 · audit A · 10 stars
academic-plotting SKILL.md skill
yichen-zju/claude-for-research · claude/skills/academic-plotting/SKILL.md · Generates publication-quality figures for ML papers from research context. Given a paper section or description…
v1.0.0 · audit A · 10 stars
academic-research-suite SKILL.md skill
yichen-zju/claude-for-research · claude/skills/academic-research-suite/SKILL.md · Claude-native Academic Research Skills suite (ported from the Codex adapter) for deep research, academic paper writing…
v0.1.14 · audit A · 10 stars
huggingface-accelerate SKILL.md skill
yichen-zju/claude-for-research · claude/skills/accelerate/SKILL.md · Simplest distributed training API. 4 lines to add distributed support to any PyTorch script. Unified API for…
v1.0.0 · audit A · 10 stars
ai-use-disclosure SKILL.md skill
yichen-zju/claude-for-research · claude/skills/ai-use-disclosure/SKILL.md · Generate venue-specific AI-use disclosure statements for manuscript submission (ICLR, NeurIPS, ICML, Nature, Science…
git:20261014.9f2a5e0 · audit A · 10 stars
alpha-research SKILL.md skill
yichen-zju/claude-for-research · claude/skills/alpha-research/SKILL.md · Search, read, and query research papers via the alphaxiv MCP server (2.5M+ arXiv papers). Use when the user asks about…
git:20261014.9f2a5e0 · audit A · 10 stars
alphafold2 SKILL.md skill
yichen-zju/claude-for-research · claude/skills/alphafold2/SKILL.md · Predict or audit protein structures with AlphaFold2-style workflows. Use when a research task needs monomer/multimer…
git:20261014.9f2a5e0 · audit A · 10 stars
anti-defensive-writing-en SKILL.md skill
yichen-zju/claude-for-research · claude/skills/anti-defensive-writing-en/SKILL.md · Stops defensive writing across the entire paper lifecycle — writing, revising, cutting, and organizing experiments…
git:20261014.9f2a5e0 · audit A · 10 stars
anti-defensive-writing SKILL.md skill
yichen-zju/claude-for-research · claude/skills/anti-defensive-writing/SKILL.md · 论文写作、修改、压缩和实验组织全流程中阻止【防御性写作】。把论文视为一场学术发布会, 而不是项目总结、实验日志或自我审查报告:识别这项工作最值得发表的价值,围绕它建立…
git:20261014.9f2a5e0 · audit A · 10 stars
autoresearch SKILL.md skill
yichen-zju/claude-for-research · claude/skills/autoresearch/SKILL.md · Autonomous experiment loop with evidence-driven move selection — scales from parameter tuning (directional search) to…
git:20261014.9f2a5e0 · audit A · 10 stars

Every file in yichen-zju/claude-for-research

Other files named experiment-forge

experiment-forge SKILL.md skill
yichen-zju/claude-for-research · claude/skills/experiment-forge/SKILL.md · Forge self-contained experiment packages with code, configs and runbooks from a task description; use when launching or…
git:20261014.9f2a5e0 · audit A · 10 stars

Browse by kind, by grade A, or by owner.