survey-generator

Survey Generator (Fireworks AI)

A Claude Code plugin that generates polished, single-file HTML survey papers on any AI/ML topic. The invoking agent curates a research bundle from a public anchor resource; Kimi K2.6 on Fireworks AI then produces an academic-style HTML artifact with inline SVG figures, numbered sections, and a References list in a single API call.

How It Works

The skill splits the work cleanly between the agent (research curation) and the model (one-shot artifact generation):

1. Anchor resource read - The agent fetches a public index you provide (an arXiv survey, a curated GitHub list, a canonical blog post, etc.) and extracts subtopics and referenced works.
2. Taxonomy and sections - The agent drafts a 4 to 8 branch taxonomy and 6 to 10 numbered sections covering introduction, foundations, methods, evaluation, and open problems.
3. Bibliography curation - The agent assembles 20 to 100 real papers (default 20 for a quick survey, 40 to 50 for a comprehensive one, 80 to 100 for an exhaustive one). Every paper needs key, authors, year, title, venue, and a 1 to 2 sentence summary.
4. Artifact generation - build_artifact.py sends research_bundle.json and style_spec.json to Kimi K2.6 on Fireworks, which returns a single self-contained HTML document with inline CSS, inline SVG figures, and parenthetical citations.

The style spec is topic-agnostic - figure geometries (taxonomy tree, paradigm panels, system stack) are parametrized off the research bundle and scale automatically as the taxonomy grows.

Setup

1. Get a Fireworks AI API Key

Create a Fireworks AI account at fireworks.ai and grab your API key from the dashboard.

2. Export the API Key

Add this to your shell profile (~/.zshrc or ~/.bashrc):

export FIREWORKS_API_KEY="your_api_key_here"

Then restart your terminal or run source ~/.zshrc.

3. Install the Plugin

/plugin marketplace add dair-ai/dair-academy-plugins
/plugin install survey-generator@dair-academy-plugins

Usage

Once installed, ask Claude Code to run the survey generator:

Run the survey generator on the topic "Reasoning Models"
using https://github.com/dair-ai/AI-Papers-of-the-Week
as the anchor source. Target 50 papers.

DAIR.AI's AI Papers of the Week is a good default anchor for broad AI/ML topics - it is an open-source, continuously updated index of notable papers with short summaries. Any public curated list, arXiv survey, or canonical blog post works too.

Claude will:

1. Read the anchor source and extract the landscape of relevant work.
2. Draft a taxonomy, section list, and bibliography into research_bundle.json.
3. Run python3 build_artifact.py from the skill directory.
4. Write the generated survey to output/survey_kimi-k2p6_v{N}.html.

Open the HTML file in your browser to read the finished artifact.

Inputs

Input	Required	Description
topic	yes	Concise survey topic, e.g. "Agentic Engineering", "Diffusion Language Models"
source_url	yes	Public anchor resource (arXiv survey, curated list, canonical blog post)
bibliography_size	no	Default 20. Use 40 to 50 for comprehensive, 80 to 100 for exhaustive
section_count	no	Default 6 to 10 numbered sections

If you do not supply these, the agent will ask.

Budget and Scaling

The skill has been validated from 20 to 100 papers on a single Kimi K2.6 call with max_tokens=81920.

Bibliography	Completion tokens (typical)	Elapsed (typical)
20 papers	~12k to 16k	~60 to 90 s
50 papers	~25k to 32k	~120 to 180 s
100 papers	~45k to 55k	~240 to 360 s

Prompt structure stays stable across runs, so Fireworks prompt caching keeps repeat runs on the same topic cheap.

Iterating

If a figure looks weak, sharpen style_spec.json (the required_figures and figure_quality_note keys) and rerun. If prose is thin or sections are missing, tighten the guidance fields in research_bundle.json.

Do not edit the generated HTML directly. Iterate on inputs and rerun. Each run writes a new versioned output file so you can compare versions.

To compare models side by side:

FIREWORKS_MODEL=accounts/fireworks/models/kimi-k2p5 python3 build_artifact.py

Output filenames are slugged by model (survey_kimi-k2p5_v1.html, survey_kimi-k2p6_v1.html, etc.).

Reference Example

A worked 100-paper example on Agentic Engineering lives under skills/survey-generator/examples/agentic-engineering/:

• research_bundle.json - the curated input with 8 taxonomy branches, 10 sections, 100 real papers.
• survey.html - the generated single-file artifact (viewBox 820x880 auto-scaled for 30 leaves).

Use it as a structural starting point for your own topic.

Why Fireworks AI?

Generating a 5000 to 6500 word survey with inline SVG figures in a single call needs a model that can handle a long structured prompt and emit a long structured output. Fireworks makes this practical:

• Context and completion headroom - Kimi K2.6 handles the full research bundle plus style spec plus an 80k-token completion budget.
• Speed - Optimized inference keeps even 100-paper runs under six minutes.
• Cost - Open-weight pricing keeps iteration on style and prose affordable.

Credits

• Built by DAIR.AI
• Powered by Fireworks AI and Kimi K2.6