Generate academic survey papers with inline figures from research

Survey Generator produces an academic-style survey paper as a single self-contained HTML file with inline SVG figures, numbered sections, and a References list. The agent's job is research curation only - reading a public anchor resource (a GitHub awesome-list, arXiv survey, blog post, or index page), extracting the topic's landscape, and building a structured research_bundle.json (title, taxonomy, sections, a bibliography of real papers). That bundle plus a fixed, topic-agnostic style_spec.json are sent in one API call to Kimi K2.6 via Fireworks, which writes all prose, figures, and HTML - the agent never authors the survey text itself.

Use it given a topic and a source_url anchor resource, with optional bibliography_size (default 20 for a quick survey, 40-50 for comprehensive, 80-100 for exhaustive, tested up to 100 entries) and section_count (default 6-10) - collected via AskUserQuestion if not supplied. Its hard rules are non-negotiable: never invent a bibliography entry (every citation must be a real paper with a real venue), every section's papers array must reference keys that actually exist in the bibliography, the generated HTML is never edited directly - only the inputs (research_bundle.json, style_spec.json) are iterated and the build rerun - and no em dashes or arrow symbols appear in the research bundle's prose fields.

Requires FIREWORKS_API_KEY in the environment and Python 3 with only the stdlib (urllib) - no external dependencies. The taxonomy is drafted with 4-8 non-overlapping branches of 2-4 children each, mapped to 6-10 sections following an introduction-foundations-methods-evaluation-open-problems progression; Figure 1's viewport height auto-scales with the taxonomy's total leaf count via a geometry contract baked into style_spec.json. Running python3 build_artifact.py reads both JSON files, calls Kimi K2.6, and writes a versioned output/survey_kimi-k2p6_v{N}.html; swapping FIREWORKS_MODEL (e.g. to kimi-k2p5) produces a comparably-slugged output for side-by-side model comparison.

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

When an agent has paper-search tools available (a Papers-of-the-Week MCP, arXiv search, Semantic Scholar, Google Scholar), it uses them to expand the candidate bibliography beyond what the anchor resource cites directly - DAIR.AI's AI Papers of the Week is called out as a particularly rich anchor for broad AI/ML topics given its years of weekly, summarized issues. The skill also documents specific SVG figure failure modes and their style_spec.json fixes - panel-local coordinate offsets, minimum rect widths, deterministic node-spacing formulas for sibling nodes, and enforced figure-caption ordering - for anyone extending the figure generation.

Researchers or writers who want a properly-cited, visually-structured survey paper generated from a starting anchor resource, without hand-writing the prose or fabricating citations - iterating on the structured research bundle and style spec rather than editing generated text directly.