SkillCenter: Autonomous AI Skill Library

Updated 9 July 2026

SkillCenter is an open, large-scale skill library that aggregates over 216,938 structured skills across 24 domain bundles for autonomous AI operations.
It packages skills as actionable, source-grounded guidance with traceability and detailed metadata to enhance correctness, security, and maintainability.
It employs a five-stage processing pipeline and offline SQLite FTS5 databases to deliver reproducible and operational knowledge to autonomous agents.

Searching arXiv for the primary "SkillCenter" paper and closely related skill-library / skill-evaluation work to ground the article. Searching "SkillCenter autonomous AI agents" SkillCenter is a large-scale, open, source-grounded skill library for autonomous AI agents, defined as a collection of “structured, retrievable, source-grounded unit[s] of operational knowledge” packaged for direct retrieval and use during task execution. It is presented, to the authors’ knowledge, as the largest open skill library for agents by total count, with 216,938 structured skills across 24 domain bundles; 114,565 are pipeline-produced and source-grounded, and 102,373 are integrated community skills. Its central aim is to supply agents with grounded operational knowledge that is not merely executable but also more correct, secure, maintainable, and traceable than raw document retrieval alone (Sha et al., 8 Jul 2026).

1. Conceptual basis and scope

SkillCenter treats a skill as a packaged unit of actionable guidance rather than as a raw document chunk, a bare tool wrapper, or a transient prompt. In the paper’s formulation, a skill can encode best practices, pitfalls, validated methods, configuration recommendations, or reproducible procedures, together with provenance, applicability, and quality metadata. This framing is motivated by a specific failure mode in autonomous agents: execution alone does not guarantee that an output is correct, secure, maintainable, or scientifically sound. Code may run while relying on deprecated APIs, mishandling edge cases, leaking resources, or applying methods that are statistically inappropriate. SkillCenter is intended to narrow that gap by distilling heterogeneous sources into operational guidance that is directly actionable during agent execution (Sha et al., 8 Jul 2026).

A defining concept is source grounding. The system defines source grounding as a traceability guarantee: each retained claim maps to an exact quotation in its source. During iterative refinement, every claimed improvement must map to an exact quote of 20 words or fewer, found by deterministic substring matching against the original source. The paper is explicit that this is not factual verification. A grounded skill may still inherit an outdated or incorrect source. This distinction is central to the platform’s epistemic model: SkillCenter guarantees provenance and traceability, not truth in the stronger sense (Sha et al., 8 Jul 2026).

The platform is also designed as an offline deployment substrate. Its skills are published as standalone bundles intended for direct use by downstream agents without vector databases, embedding models, or network access at query time. This design places SkillCenter closer to a curated operational-knowledge substrate than to a conventional retrieval corpus (Sha et al., 8 Jul 2026).

2. Corpus composition, source populations, and bundle structure

The released library combines a source-grounded pipeline subset with externally integrated community bundles.

Subset	Count	Notes
Source-grounded pipeline subset	114,565	Research 90,084; Technical 24,481
Community subset	102,373	GitHub 90,984; ClawHub 11,389
Total library	216,938	24 domain bundles

The source-grounded subset is produced from 55,199 distinct source documents. The paper reports 86,594 skills from journals, 20,071 from GitHub, 3,130 from web pages, 3,608 from ArXiv, and 1,162 from forums, with average skills per source of 3.0 for journals, 2.1 for ArXiv, and 1.0 for GitHub, web, and forum sources. Upstream source types are web pages, GitHub repositories, Stack Overflow forums, ArXiv preprints, and journals accessed via APIs, including PLOS journals, eLife, and Nature-family journals. Public GitHub SKILL.md files and a mirrored ClawHub marketplace bundle form the community portion (Sha et al., 8 Jul 2026).

The organizational model distinguishes 13 domain configurations from 24 published bundles. A domain configuration controls acquisition and routing—such as host allowlists, GitHub query constraints, and forum tags—whereas a published bundle is the SQLite database shipped to users. The 24 bundles comprise 21 pipeline-produced bundles and 3 integrated community bundles. Technical bundles are grouped by topic domain; research bundles are grouped primarily by publication venue. The coarse bundle count and the finer per-venue decomposition are not perfectly one-to-one in the text, but each published skill belongs to exactly one bundle, and cross-bundle overlap is intentionally low (Sha et al., 8 Jul 2026).

This heterogeneous composition is analytically important. The headline total is not a uniform corpus: nearly half of the full library comes from community bundles that bypass SkillGate, generation, source-grounding, and the publish gate. The paper therefore advises that users requiring a uniformly processed corpus should focus on the 114,565-skill pipeline subset (Sha et al., 8 Jul 2026).

3. Construction pipeline, SkillGate, and publication controls

The source-grounded subset is built through a five-stage pipeline: Source Acquisition, Quality Gate, Skill Generation, Iterative Improvement, and Publishing. Web pages are fetched by HTTP GET and processed with readability extraction; GitHub content is collected through the GitHub REST API using domain-specific search queries, minimum star thresholds, and recency filters such as pushed_after; Stack Overflow content is collected through the Stack Exchange API v2.3; ArXiv content uses the ArXiv API plus PDF-to-text extraction; and journals use XML APIs for section-level parsing. All sources then undergo common postprocessing, including text extraction, truncation to 12,000 characters, URL canonicalization, and storage as capture artifacts (Sha et al., 8 Jul 2026).

The pre-generation gate is SkillGate, an LLM-based filter operating on source excerpts truncated to 4,000 characters. It emits one of three verdicts—pass, maybe, or fail—together with a 0–10 suitability score plus good_signals and bad_signals. Sources under 200 characters fail automatically. SkillGate evaluates actionability, reproducibility, and content density, and the default policy is permissive: maybe is treated as pass. The paper gives a representative distinction between a good GitHub README that passes with score 8 and a bare code repository that fails because it lacks context and operational guidance. Corpus-wide acceptance and rejection totals are not yet tracked (Sha et al., 8 Jul 2026).

Generation is template-driven. The paper lists four template types: Default (tech), paper_writing, paper_writeup, and experiment_design. Technical templates use sections such as Background, Use Cases, Inputs, Outputs, Steps, Verification, and Evidence. Research-oriented templates differ by source type; for example, journal papers may yield multiple skill kinds such as paper.idea_intro, paper.experiment, paper.method, and paper.picture, whereas ArXiv uses a different taxonomy that includes paper_writing, paper_writeup, experiment_design, and results_analysis (Sha et al., 8 Jul 2026).

Iterative improvement runs for up to 3 passes by default, configurable from 1 to 5, and is driven by the score

$\text{score} = \text{lint\_count} \times 100 + \text{missing\_count}$

where lint_count measures structural problems and missing_count counts unresolved improvement suggestions. Each pass lints the markdown, extracts unresolved suggestions, rewrites the skill under those constraints, runs deterministic source-grounding, and retains the better version. The loop stops when the score reaches zero or the pass budget is exhausted. Most skills reportedly reach zero within two passes, although no full histogram is given (Sha et al., 8 Jul 2026).

Publication applies a stricter gate. A skill must satisfy four conditions: overall quality score $\ge 3.0$ on the 1–5 rubric, license whitelist/blacklist rules, plagiarism ratio < 0.35, and “Not provided” count < 15. The license policy allows, for repositories, MIT, Apache-2.0, BSD-*, ISC, MPL-2.0, and Unlicense, while denying GPL-3.0, AGPL-3.0, and LGPL-*; forums allow CC-BY-SA-4.0 and CC-BY-SA-3.0. The paper emphasizes that this licensing gate is a technical default rather than a legal determination (Sha et al., 8 Jul 2026).

4. Skill representation, local search, and agent-facing use

Conceptually, a SkillCenter skill contains title, description or background, applicability, evidence, and score. In the canonical markdown example, the structure is:

# [Skill Title]
## Background
## Use Cases
## Inputs
## Outputs
## Steps
## Verification
## Evidence
## Sources

Operationally, skills are stored in SQLite bundles with a four-table schema: skills_index, skills_content, skills_fts, and bundle_meta. skills_index stores metadata such as skill_id, domain, profile, source_type, source_url, title, overall_score, skill_kind, language, source_id, and primary_source_id; skills_content stores metadata_yaml, skill_md, and library_md; skills_fts is an FTS5 virtual table; and bundle_meta stores key-value metadata (Sha et al., 8 Jul 2026).

Distribution is through offline-searchable SQLite FTS5 databases. Technical bundles average roughly 120 MB and research bundles about 1 GB. Search uses FTS5/BM25 over indexed fields, with optional filters by domain, skill kind, or minimum quality score, and merge-sorting across multiple installed bundles. The current FTS index covers title and domain, not full body content. The agent-facing pattern is therefore metadata-first discovery, followed by on-demand loading of full markdown. Source URLs are stripped from the injected prompt to reduce stale or hallucinated link generation, although provenance remains in the stored bundle (Sha et al., 8 Jul 2026).

This limited indexing strategy is not incidental; it becomes a central bottleneck in evaluation. The system’s downstream pattern—retrieve candidate skills, inject selected content, and execute—works only insofar as retrieval can surface the needed skill. The paper’s results show that the corpus contains useful knowledge, but current title-only lexical retrieval often fails to expose it (Sha et al., 8 Jul 2026).

5. Evaluation, audits, and principal limitations

SkillCenter’s downstream evaluation uses roughly 2,000 offline, objectively graded tasks across 51 algorithmic and data-processing domains with four solver models: gemini-3.5-flash, claude-haiku-4.5, claude-sonnet-4.6, and gpt-5-mini. The comparison includes baseline, +Keyword using top-3 FTS5-retrieved skills, +Placebo with three irrelevant but length-matched skills, and, in a separate probe, +Oracle with the genuinely needed skill. On the general task pool, keyword retrieval does not help and is sometimes mildly harmful. For example, claude-haiku-4.5 drops from 87% baseline to 83% with +Keyword, while gpt-5-mini drops from 93% to **91%`; placebo behaves similarly. The paper interprets this as added context cost on tasks already within model competence (Sha et al., 8 Jul 2026).

The picture changes on knowledge-gap probes. In a synthetic “agent-gap” benchmark, baseline, placebo, and keyword all score 0%, whereas injecting the oracle skill raises success to 72%–100% depending on model, with exact McNemar $p$ -values from $5\times 10^{-13}$ to $1\times 10^{-63}$ . On 18 tasks built from real SkillCenter research skills whose answers require exact study-specific parameters, baseline, placebo, and keyword again score 0%, while the relevant real skill yields 61%–78% across models, with pooled 69% over $n=72$ , $\Delta=+50$ , and $p \approx 2\times 10^{-15}$ . The paper’s conclusion is that the corpus does contain knowledge the models do not have, but current retrieval fails to surface it reliably (Sha et al., 8 Jul 2026).

The library also undergoes corpus audit. Identity duplication is zero at the skill_id level, but 9,831 skills (4.53%) are verbatim duplicates after normalization, and 24,988 skills (11.52%) are near-duplicates at Jaccard $\ge 0.8$ via 128-permutation MinHash over word 5-shingles with LSH. Redundancy is concentrated almost entirely in the community bundles; every pipeline domain is effectively duplicate-free at 0.01% or below. A licensing audit reports, within the pipeline subset, 61,575 skills (53.7%) as CC-BY-4.0 journal-derived, 11,771 (10.3%) MIT, 5,353 (4.7%) Apache-2.0, and 32,655 (28.5%) with no SPDX tag, mostly web-derived skills admitted through plagiarism filtering. A baseline safety scan finds prompt-injection phrases in 318 pipeline skills versus 2,342 community skills, again indicating substantially higher risk concentration in externally harvested bundles (Sha et al., 8 Jul 2026).

The paper is explicit about its limitations. Source grounding is traceability only, not verification. Deduplication is lexical rather than semantic. The quality signal is compressed, with about 82.1% of pipeline-scored skills receiving exactly 4. Nearly half of the headline corpus bypasses the source-grounded pipeline. A small legacy set of 302 skills, or 0.26% of the released pipeline subset, remains below the nominal 3.0 publication threshold. The system therefore demonstrates a large operational-knowledge substrate, but not a solved end-to-end retrieval-and-deployment stack. This suggests that repository scale, by itself, is insufficient without stronger retrieval, calibration, and governance (Sha et al., 8 Jul 2026).

6. Position within the skill-systems literature

SkillCenter belongs to a rapidly forming literature on skill repositories, skill production, skill evaluation, and skill organization. Closely related work differs in emphasis. SkillNet defines skills as file-based packages centered on SKILL.md, organizes them through a three-layer ontology comprising taxonomy, relation graph, and package library, and evaluates them along Safety, Completeness, Executability, Maintainability, and Cost-awareness; it reports over 200,000 skills initially aggregated and 150k+ curated after filtering (Liang et al., 26 Feb 2026). SkillFab addresses the production side of the lifecycle by treating missing capability as a first-class issue, coupling reuse-first registry search with demand-first issue capture, Git-backed evidence ingestion, maintainer review, and versioned publication (Xu et al., 4 Jul 2026). SkillChain studies production-scale skill evolution for image-commerce assistants, separating skill Description for routing from skill Body for behavior and updating them through Route Optimizer and Body Refiner under monotone deployment gates (Hu et al., 11 Jun 2026).

Evaluation-centric systems expose the pressures acting on a library such as SkillCenter. OpenSkillEval automatically constructs realistic task instances from evolving real-world artifacts across five downstream application categories and shows that skill availability does not guarantee effective skill use, that the benefit of skill augmentation depends strongly on both model and framework, and that many popular skills do not consistently outperform no-skill baselines (Ying et al., 22 May 2026). SkillAudit shifts evaluation from fixed task suites to skill-centered assessment, automatically generating capability-aligned utility and safety probes from arbitrary skill packages and reporting that 17 out of 226 skills, or 7.5%, were flagged as risky under its thresholding rule (Yu et al., 21 Jun 2026). SkillJuror isolates organization from content, showing that Progressive Disclosure raises distinct Skill resources touched per trajectory from 1.18 to 3.85 and effective uptake events from 1.33 to 3.92, while yielding 17 additional verifier-passing trials out of 410 over a normalized flat baseline (Chen et al., 10 Jun 2026). SkCC adds a compiler perspective by introducing a strongly typed intermediate representation, compile-time security analysis, and per-framework emission, reducing adaptation complexity from $O(m \times n)$ to $\ge 3.0$ 0 and improving pass rates on SkillsBench from 21.1% to 33.3% on Claude Code and 35.1% to 48.7% on Kimi CLI (Ouyang et al., 5 May 2026). SkillAxe addresses refinement of text-authored skills, improving pass rates by 28% relative over unimproved LLM skills on SkillsBench and, in a library setting, raising SpreadsheetBench pass rate from 16.0% to 52.0% using only 22 skills (Gautam et al., 9 Jun 2026).

Taken together, these systems place SkillCenter in a broader transition from prompt snippets and ad hoc agent memory toward governed skill artifacts with provenance, packaging, retrieval, auditing, and revision. Within that landscape, SkillCenter’s distinctive contribution is the combination of large-scale source grounding, offline SQLite FTS5 distribution, and an explicit demonstration that real skills can close real knowledge gaps even though current retrieval remains the limiting factor (Sha et al., 8 Jul 2026).