Overview
FractiSkills treats an entire website as a corpus of agent skills: every reachable page is normalized into a SKILL.md-style artifact so that a whole site — not a single document — becomes loadable context for an agent. The work is organized around a four-stage pipeline — discover, render, publish, research — with Skillarum serving as the engine that turns raw crawl output into citable, receipt-bearing skill documents. Discovery unions the declared sitemap with a bounded, robots-respecting breadth-first crawl (reaching a maximum depth of 5), resolves redirect and canonical aliases, and partitions every page into site-derived sections; the render stage emits one receipt-bearing skill per page; publish and research then validate, organize, and measure the corpus.
The resulting corpus covers 405 pages from 53 sitemap URLs, organized into 12 sections and 405 skills totaling 497896 words (3887524 characters). Augmentation — a declared, same-origin document retrieval for client-rendered pages — was attempted 173 times and succeeded 169 times (1988848 document characters retrieved), with 4 failures recorded as receipts. The size effect is stark: augmented skills have a median of 1824 words versus 486 for static pages, a ratio of 3.8$\times$. The run issued 417 network requests, all evidence origins marked live, under pipeline version 0.8 and cache version 7.
Reproducibility is enforced rather than promised: every statistic in this manuscript is emitted as a token that must resolve against a machine-generated data contract, and every page carries a fetch receipt — URL, timestamp, SHA-256 content hash, and augmentation status — so any number can be traced to a specific observation within the 2026-09-10T23:51:05.959201+00:00 through 2026-09-10T22:19:21.394219+00:00 (UTC) window. Failed augmentations are reported, not silently dropped. Keywords: agent skills, SKILL.md, web scraping, Skillarum, provenance, reproducible research.
---
Associated artifacts
GitHub release: v0.1.0 (https://github.com/docxology/FractiSkills/releases/tag/v0.1.0)
DOI: https://doi.org/10.5281/zenodo.22712651
Zenodo: https://zenodo.org/records/22712651
PDF SHA-256: af96db3ef73bb648ddbfed92c2c4f9dff099e772ca9e9ff0041101eeb6135b4b
Overview source: Curated paper metadata.
Methods and contributions
Read the source for the full argument, qualifications, and evidence.
Findings and contributions
- The run produced a corpus of 405 pages from 53 sitemap URLs, in 12 sections and 405 skills totaling 497896 words.
- Sitemap-only ingestion would have missed 88.9% of the corpus: 360 of 405 pages were found only by the crawl, including all Ship-Blog and Voyage content.
- Augmentation succeeded in 169 of 173 attempts, with the 4 failures recorded as receipts rather than dropped.
- Augmented skills had a median of 1824 words versus 486 for static pages, a 3.8x ratio.
- The modal BFS depth was 4, holding 53% of pages, with the deepest page at depth 5.
Methods
- Four-stage discover-render-publish-research pipeline built on Skillarum v0.2.0 — Normalizes every reachable page of a website into one SKILL.md-style skill, with each stage a separately testable module.
- Sitemap-union-crawl discovery with canonical alias resolution — Unions the declared sitemap with a bounded, robots-respecting BFS crawl, deduplicating URLs by canonical or final post-redirect identity.
- Declared same-origin JSON augmentation for client-rendered pages — An AugmentedGenerator retrieves documents behind client-rendered shells from reviewed endpoints and stamps per-fetch receipts.
- Case study: crawl of the SS Vibelandia Omniversal Canvas website — Applies the pipeline to the SS Vibelandia site, which mixes static pages, client-rendered readers, JSON catalogs and redirect stubs.
- Token-bound manuscript statistics with per-page fetch receipts — Every reported statistic resolves against a generated data contract; each page carries URL, timestamp and SHA-256 hash.
Key concepts
- agent skills
- SKILL.md
- web scraping
- Skillarum
- reproducible research
- digital art documentation
- provenance
Summary sources: Paper metadata and evidence · Learning guide · Extracted source text.
PDF downloads
Archived files available directly from this site.
- FractiSkills_combined.pdf PDF · 1.78 MiB · extracted-text source
Citation
Citation metadata follows the unified bibliography.
Catalog details and resources
Related in Computational
Other catalogued works in the same domain.