` link |
| h2 | Installation | "Install requirements:", one five-line `bash` block, then a paragraph on the API key file with two inline-code paths |
| h2 | AgentStudio Overall Benchmark Tasks | the observation/action space figure, then one paragraph (205 tasks) with a link to the task README and the project page |
| h2 | AgentStudio Datasets Decompose Agent Abilities | one paragraph naming GroundUI, IDMBench and CriticBench with links |
| h2 | AgentStudio Tools | a lead sentence, a three-item bullet list, a paragraph of three doc links |
| h2 | Contributing | one sentence linking CONTRIBUTING.md |
| h2 | Acknowledgement | one sentence listing seven linked projects |
| h2 | Citation | one sentence and a `bibtex` block |
Absent from the source: author, affiliation and venue lines (the venue appears only as the `booktitle` in BibTeX), a teaser GIF or video, a results table in Markdown, a model zoo, a quick-start run command, a license section (the license is only a badge), a table of contents, emoji in headings.
Present but unusual: the first screen holds two images before any heading, and the second is a feature-comparison table delivered as a PNG rather than a Markdown table; both images use `![]()` with empty alt text; a disabled badge stays in the source as an HTML comment; three section headings repeat the project name ("AgentStudio Overall Benchmark Tasks", "AgentStudio Datasets ...", "AgentStudio Tools").
## Components
**Centred title and badge paragraph.** Only two header elements are centred: `` and the badge `
`. Everything after them is ordinary left-aligned Markdown, and the two images look centred only because they span the 856px column.
**Badge row.** Six flat static badges, label grey #555555, messages #b31b1b, #44bb00, #007ec6, #000000, #007ec6, #44bb00. Paper links come first, tooling badges after. Quotes are mixed in the source (single quotes on the first two `href`/`src`, double on the rest); `alt` is set on four of six.
**Desiderata-to-component diagram.** A four-column grid image: a filled left column of three colored desiderata (#dae7d5, #dde8fc, #f5e5cb) and three outlined columns (Environments, Tools, Benchmarks) whose rows inherit the color of the desideratum they serve; a merged "Online benchmark tasks" cell spans three rows. Column titles are bold black two-line headings inside the image.
**Bold-lead numbered list.** Four `1.` items under the intro, each opening with a bold phrase and adding bold key terms mid-sentence (**generic observation and action spaces**, **auto-evaluation**).
**Comparison table as image.** Columns Environment, Domain, Obs. Space, Action Space, then five yes/no feature columns (Interactive, Data & Task Tools, Language Feedback, Lightweight, Decompose Abilities) marked with ticks #40651a and crosses #bb271a. Rows are grouped by rules (web, mobile, desktop) and the project's own row is last, in bold, with ticks in every feature column.
**News list.** `- **Mon D, YYYY**: Sentence.` with the month abbreviated and no emoji; three entries, newest first.
**Code blocks.** Two fenced blocks: `bash` for install (with a trailing `# If using Ubuntu 22.04` comment on one line) and `bibtex` for the citation. Paths and file names elsewhere are inline code.
**Citation.** "If you find AgentStudio useful, please cite our paper:" then `@inproceedings{zheng2024agentstudio, ...}` with `Agent{S}tudio` brace-protected and the ICLR proceedings written out in `booktitle`.
**Acknowledgement and contributing.** One-sentence prose sections with inline links, no logos and no badges.
## Imagery
| Image | Type | Source size | Markdown | Rendered at 920px | Placement |
|---|---|---|---|---|---|
| overview.png | PNG, pastel grid diagram on white | 1680×488 | `` | 856×249 | header, directly under the badge row |
| comparison.png | PNG, paper-style table with ticks and crosses | 1924×578 | `` | 856×257 | header, after "Comparisons with existing work:" |
| agent_space.jpg | JPEG, two-panel icon diagram (web browser vs AgentStudio environments) split by a dashed line | 6236×2120 | `` | 856×291 | first line of "AgentStudio Overall Benchmark Tasks" |
All three are wide and short (aspect ratios 3.4 to 2.9), white-backed and free of captions; none uses ``, a width attribute or a link wrapper written by the author (GitHub adds its own link to the full image). There is no GIF and no video.
## Adapting to Other Content
- **A library or toolkit.** This is the native fit: keep the centred h1 and one flat badge row of label/value pairs (paper, docs, Python version, code style, license), and replace the desiderata diagram with a single wide feature-map image about 1700px wide (856px rendered, roughly 250px tall).
- **A benchmark.** Keep the comparison image but consider a Markdown table instead, with the same columns (domain, observation space, action space, then yes/no features) and the benchmark's own row last in bold; a Markdown table stays searchable and readable in dark mode.
- **A paper repo with authors.** Insert a centred author line and a bold venue line between the h1 and the badge row; AgentStudio leaves both out, so this is an extension.
- **A repo without a figure.** Let the badge row close the header and move the bold-lead numbered list directly under it as the summary.
## Do's and Don'ts
| Do | Don't |
|---|---|
| Put one wide summary diagram directly under the badges | Open with a long paragraph before any picture |
| Key the diagram's rows by color so the reader can trace a need to a component | Use colors in the diagram that mean nothing |
| Show positioning against prior work as a tick/cross table with your row last and bold | Claim superiority in prose without a comparison |
| Keep badges in one flat family on the default #555555 label | Leave dead badges commented out in the header source |
| Give every image alt text | Ship `![]()` images with empty alt as AgentStudio does |
| Keep sections short and link out to sub-READMEs | Duplicate the sub-READMEs' instructions in the main page |
## Agent Prompt Guide
1. "Write a README header in GitHub Markdown: `ProjectName
`, then one `` with six flat shields.io static badges on the default #555555 label: arXiv number #b31b1b, Project Page #44bb00, python 3.11 #007ec6, code style black #000000, License #007ec6, pre-commit enabled #44bb00 with `logo=pre-commit&logoColor=white`. No author or venue lines."
2. "Under the badges put a full-width summary diagram (``, about 1680×488 source, 856×249 rendered): a left column of three desiderata filled #dae7d5, #dde8fc and #f5e5cb, and three outlined columns for environments, tools and benchmarks colored to match. Follow with one intro sentence with a bold core phrase and a four-item numbered list whose items open in bold."
3. "Add 'Comparisons with existing work:' and a wide comparison table (about 1924×578) listing prior environments by domain, observation space and action space, with five feature columns marked by ticks #40651a and crosses #bb271a and your row last in bold. Then plain `##` sections: News (bold `Mon D, YYYY` dates), Installation (`bash`), one section per component, Contributing, Acknowledgement, Citation (`bibtex`)."
## Similar Pages
- readme-nepfaff-scenesmith: also a centred h1 over one centred row of flat static badges (arXiv #b31b1b first) before any figure, though SceneSmith adds an icon in the title and an author block.
- readme-drexubery-viewcrafter: uses the same two badge URLs in the same form (`arXiv--b31b1b.svg`, `Project-Page-Green`), so its first two tags look the same as AgentStudio's.
- readme-seohongpark-ogbench: another agent benchmark README with a centred title and a uniform badge row; it uses `for-the-badge` badges in one brand color rather than the flat mixed palette here.
- readme-eku127-satnav: a benchmark README that explains itself with one wide diagram, but it adds a paper-style author header and a runnable Quick Start that AgentStudio lacks.
AgentStudio's distinguishing choices are the absence of any author block, a color-keyed desiderata map as the header figure and a positioning table delivered as an image before the first heading.
## Quick Start
````markdown
ProjectName
ProjectName is **a one-phrase summary of what it is** for some audience. It provides:
1. **First component** with **a key property**
2. **Second component** for a task
3. **Third component** that evaluates something
4. **Datasets**: A, B and C, for three abilities
Comparisons with existing work:
## News
- **Mon D, YYYY**: Released the full version.
- **Mon D, YYYY**: Released the beta version.
## Installation
Install requirements:
```bash
git clone https://github.com/org/name.git
conda create --name name python=3.11 -y
conda activate name
pip install -e .
```
## Component One
One paragraph with a link to the component README.
## Component Two
## Contributing
## Acknowledgement
## Citation
If you find ProjectName useful, please cite our paper:
```bibtex
@inproceedings{key,
title={...},
}
```
````
## Provenance
- Repo: https://github.com/ltzheng/agent-studio (`README.md`, branch `main`, last README commit 64b839d on 2025-06-16).
- Paper: "AgentStudio: A Toolkit for Building General Virtual Agents", ICLR 2025, arXiv 2403.17918; first author Longtao Zheng. Project page https://ltzheng.github.io/agent-studio.
- Selection: a colored desiderata-to-component diagram serves as the header summary, and a feature-comparison table against prior environments adds a positioning style the other picks lack.
- Evidence: raw Markdown and GitHub-rendered HTML (saved 2026-10-28), 1440×900 and 2700px captures at a 920px README column, shields.io SVGs for all six badge URLs, PNG/JPEG headers and pixel color counts for `overview.png`, `comparison.png` and `agent_space.jpg`.
- Written 2026-10-28.
` lines and one HTML comment (a disabled mypy badge, ``) that GitHub drops. Each badge is 20px tall; the source newlines render as single spaces and the row spans about 665px. Details in Badges.
3. **Summary diagram** (y≈210–459): `` with an empty alt and no width. The 1680×488 PNG scales to 856×249 under GitHub's `max-width: 100%`; it is centred only because it fills the column. No caption.
4. **Intro paragraph** (y≈485): plain Markdown, left-aligned, "AgentStudio is **a trinity of environments, tools, and benchmarks** for ..." followed directly by a numbered list (`1.` to `4.`) whose items open with bold phrases (**A lightweight, interactive environment**, **Online benchmark tasks**, **Three benchmark datasets**).
5. **Comparison lead-in** (y≈745): the plain line `Comparisons with existing work:`.
6. **Comparison table image** (y≈778 to about 1035, below the fold): ``, 1924×578, rendered 856×257. This ends the header block; `## News` follows.
## Badges
All six use the static path form `https://img.shields.io/badge/