ASB Skill Collections Evidence-grounded skills and software-tool records for scientific AI agents, distilled from published method papers and their code by the AgenticScienceBuilder pipeline. A skill is a short procedure that an AI agent can read and follow. The AgenticScienceBuilder (ASB) pipeline writes each one from a published method paper (journal article or preprint) and, where one exists, its public code…
Evidence-grounded skills and software-tool records for scientific AI agents, distilled from published method papers and their code by the AgenticScienceBuilder pipeline.
A skill is a short procedure that an AI agent can read and follow. The
AgenticScienceBuilder (ASB) pipeline writes each one from a published method paper
(journal article or preprint) and, where one exists, its public code repository, anchors it to
verbatim evidence from those sources (for repo-oa papers, the repository’s own
documentation; CONTENT_POLICY.md), annotates it with
EDAM terms, and passes it through a licence and PII gate before release. This
repository is the public registry of those collections. It holds the install and
search tooling, the governance of the community that curates them, and the
release registry. Released content lives in per-domain repositories. The project started at the
Dagstuhl Seminar Computational Metabolomics
(26181,
26 to 30 April 2026).
Releases
The registry collections.yaml records the released
collections of ten domains. Each release is deposited on Zenodo and pinned to its
domain repository asb-skills-<domain>, commit and tree digest. The domain
repositories are public, and the Zenodo record of each release carries the full
collection. This repository holds the tooling and the registry.
The concept DOI of each domain is in collections.yaml. The
rest of this file uses metabolomics v3 as its worked example; asbb fetch and
asbb install take any registered collection, for example
--collection genomics/v1.
Metabolomics v3
Release metabolomics-v3.1.0: 11,830 skills across 1,989 tools, distilled
from 477 published method papers (journal articles and preprints), plus 20
composite workflow super-skills that chain the atomic skills into end-to-end
pipelines. The release is deposited as
10.5281/zenodo.24224655. Coverage is mostly
LC-MS/MS, with LC-MS, GC-MS, mass-spectrometry imaging, ion mobility and
lipidomics, some NMR, and multi-omics, statistics and pathway tools.
The collection is released from the domain repository
asb-skills-metabolomics. The marketplace of this repository serves the v3
collection and the v2-derived technique packs from that domain repository.
Layer
Contents
State
ASB-Skills
evidence-grounded procedural skills
released
ASB-Workflows
composite end-to-end pipeline super-skills
released
ASB-Tools
deduplicated software-tool records with EDAM terms and DOIs
released
ASB-Evaluation definitions
definitions for per-paper tasks and claim-retrieval tests
not published; definitions only, with no validated benchmark, reference results or agent-performance leaderboard
ASB-Capsules
raw per-paper pipeline outputs, for full traceability
later
Install
Claude Code (plugin marketplace):
/plugin marketplace add HolobiomicsLab/asb-skill-collections
/plugin install metabolomics@asb-skill-collections
# Or one of the eight technique packs, for example:
/plugin install metabolomics-lc-ms@asb-skill-collections
The other released domains install the same way under their domain name, for
example /plugin install genomics@asb-skill-collections. Each marketplace entry
points at its domain repository, which is public, so installation needs no read
access. The nine domains released on 2026-11-09 were installed this
way on 2026-11-10, with Claude Code and with Codex CLI; the metabolomics
entry was last installed before its source moved to the domain repository
(docs/STATUS.md).
The same holds for Codex: no read access is needed. The Zenodo record of the
release (version DOIs in the Releases table) is the archive route;
read it as described in AGENTS.md.
Install asbb from this public registry checkout (it is not published on
PyPI):
fetch reads collections.yaml and downloads the selected
leaves from the domain repository at the pinned commit. GITHUB_TOKEN is
optional; when set, it is sent with every request.
A derived selection from a domain checkout (asbb install):
The human guide is <domain checkout>/collections/metabolomics/v3/USAGE.md in
plain text. The complete release is also available from the
Zenodo archive. Install routes,
measured v2 examples and refusals are in
docs/partial-install.md; the agent procedure is in
AGENTS.md.
Choose what to install
Install the whole collection, a technique pack, a selection of skills
(asbb install --skills, --skills-file or --query), or fetch chosen skills
without a checkout (asbb fetch). Commands, outputs and refusals:
docs/partial-install.md.
The full metabolomics plugin serves v3: 11,830 skills over 1,989 tools. The
eight technique packs are the v2-derived packs shipped by the domain repository.
Their measured v2 sizes remain useful for choosing a pack; no v3 full-install
size is claimed here.
a fetch, asbb fetch a,b --collection metabolomics/v2 --dest DIR
chosen slugs
10 in the example
112 KiB
The eight packs install by name with /plugin install <name>@asb-skill-collections.
Their sizes are du -sk of the v2 units measured before the domain split. The
selection and fetch rows are the 10-skill example of
docs/partial-install.md, and a fetch also
downloaded the 5.2 MiB (5,430,046 bytes) v2 skills_index.json at measurement
time.
Packs overlap: a skill tagged with several techniques is in several packs, so
install the full collection or a few packs, not both. The 1,520 technique-agnostic
skills (statistics, pathway and generic data-handling procedures) carry no
technique tag and are in no pack; they come with the full collection, a selection
or a fetch.
Use: search, apply, ground
Search a domain checkout offline, with no API key and the tooling installed
from the registry checkout as above. Pass the domain checkout explicitly:
<collection root> is the directory that holds skills_index.json; in an
installed plugin it is two directories above the router’s SKILL.md. Written as
an absolute path, the command runs from any directory. The routers of the
releases of 2026-11-09 print python bin/search_skills.py, relative to the
collection root: it fails from any other directory and on a system with only
python3, so run the form above instead. The CLI, the router, the MCP server and the collection-level search
scripts share one versioned selector; each hit reports its score, matched fields
and qualified slug. How offline skill selection works
documents the rule, its tie handling and its 22-query measurement.
Apply. Read the skill’s SKILL.md (the read: path of the hit) and follow
the procedure. Tool requirements are in its frontmatter and in
<domain checkout>/collections/metabolomics/v3/USAGE.md, section 0.
Ground. To check an exact parameter, threshold or claim (the difference
between “the skill says about 5 ppm” and “the paper specifies 5 ppm”), ground the
skill against its source paper. Grounding is optional. Every released collection
ships the grounding map kb_bundle.json; the binder is
scripts/perspicacite_kb_bind.py in this
repository, and a selection built with asbb install carries a copy. It uses
Perspicacité,
Holobiomics Lab’s local-first literature RAG engine, with two backends:
kb: RAG over the paper’s full text and supplementary information. The
per-paper KB (asb-paper-<doi>) is created on first use through the Perspicacité
HTTP API; no MCP server is required. Needs Perspicacité at PERSPICACITE_BASE
(default http://127.0.0.1:8000).
local: no server. Clones the skill’s source repository and fetches the
open-access paper where possible. Needs git and network.
Run python3 scripts/perspicacite_kb_bind.py (prepare, query, local,
resolve) with --collection "<collection root>" --skill <slug>; in a selection
built with asbb install, its bin/ copy and the /ground command do the same.
Full guide: <domain checkout>/collections/metabolomics/v3/USAGE.md, section 5.
What is in the collection
The LC-MS view of the collection (the metabolomics-lc-ms pack). The full release covers all techniques.
File
Contents
leaves/<slug>/SKILL.md
one evidence-grounded skill each (frontmatter: EDAM IRIs, provenance.source_papers DOIs and titles, tools with their roles, metadata.license_tier; verbatim evidence lines in the body). Shipped as data and read on demand, so installing does not load the corpus into the session’s context.
skills/_router/SKILL.md
the one advertised entry, the router <domain>-router
workflows/<slug>/SKILL.md
composite workflows, their index workflows/workflows_index.json and their router workflows/_workflow_router/SKILL.md (not advertised)
tools/<slug>.yaml
deduplicated software-tool records with EDAM terms and source DOIs
the offline search over skills, workflows and tools (standard library only)
kb_bundle.json
skill to source-paper KB slugs and repo_urls (the grounding map)
collection.yaml, corpus.yaml, mapping.json
collection record; per-paper access basis (repo-oa); each leaf mapped to its source build records and DOIs
CITATION.cff
citation
These files are relative to <domain checkout>/collections/metabolomics/v3/;
the other released collections have the same layout. No released collection
ships the grounding binder or a commands/ directory. The release-gate report
and the other receipts are in the domain repository under receipts/<cut-id>/.
Local build-capsule archives and record metadata are packed and screened by
scripts/pack_capsules.py, without deposit or upload.
What their evaluation definitions contain, counted per domain, and what does
not exist is in docs/evaluation-definitions.md.
How it was generated
Skills come from method papers anchored to the computational metabolomics review
series (SOURCES.md: sources and inclusion criteria;
CONTENT_POLICY.md: legal and open-access policy).
The receipts under <domain checkout>/receipts/<cut-id>/ and the Zenodo record
carry the release provenance. The method records the exact build command
and model routing: Claude Opus 4.8 for outline and card revision, Claude Haiku
4.5 for the other stages, OpenAI text-embedding-3-large and -small for
embeddings, each build recorded in build_manifest.json.
Repository family
Repository
Role
Access
asb-skill-collections (this one)
asb_skill_collections tooling, scripts/, tests, governance, registry, catalogue and docs-site
public
AgenticScienceBuilder
the pipeline that generates the collections
private
asb-skills-metabolomics
<domain checkout>/collections/metabolomics/<version>/, <domain checkout>/packs/metabolomics/<technique>/, <domain checkout>/staged-collections/metabolomics/ and receipts; v3 deposited on Zenodo
public
asb-skills-<domain>
one repository per other domain, with <domain checkout>/collections/<domain>/<version>/, <domain checkout>/packs/<domain>/<technique>/, <domain checkout>/staged-collections/<domain>/ and receipts; each release deposited on Zenodo
public
indicium
public standard for a typed scientific claim bound to its evidence; PyPI package indicium-standard 2.0.0
public
indicium-adapters
public core API for domain adapters; PyPI package 0.6.0
public
indicium-adapters-<domain>
per-domain adapters that enrich indicium claims with ontology terms; one repository for each of the ten domains
public
asb-schema
public schemas for ASB task cards, capsules and collections; PyPI package 0.4.0
multi-agent framework for autonomous scientific research
public
collections.yaml pins each published domain release to its
repository, commit, collection path and tree digest. Domain and adapter
repositories exist for metabolomics and nine further domains
(agentic AI, astronomy, epigenomics, genomics, metabarcoding, metagenomics, multi-omics,
proteomics, transcriptomics). Each domain has a release recorded in the
registry and deposited on Zenodo, and is released from its own domain
repository, which is public. The legacy epigenomics,
transcriptomics and metabolomics v1 trees are recorded in the registry and are
not part of any release.
Cite the collection and the original paper behind each skill you use
(the DOIs in provenance.source_papers in its SKILL.md).
Other domains. The concept DOI of each domain and the version DOI of each
release are recorded in collections.yaml; the version
DOIs are also listed under Releases. The entries below are those
of metabolomics.
Previous version DOI, v2 (22 June 2026):
10.5281/zenodo.20794027. The
registry records the domain commit and tree digest for each version.
Version citation files. The v2 and v3 CITATION.cff files are
<domain checkout>/collections/metabolomics/v2/CITATION.cff and
<domain checkout>/collections/metabolomics/v3/CITATION.cff.
Roles (roles in collection.yaml, attribution: in each workflow SKILL.md): generator (the ASB
pipeline), curators (none yet), promoter (Louis-Félix Nothias), sponsor
(CNRS and Université Côte d’Azur, who paid the API cost).
Governance and licensing
Every skill carries a license_tier (in skills_index.json and in its SKILL.md
frontmatter, metadata.license_tier) that answers what may I do with the underlying tool?
Tier
Meaning
open
commercial use allowed (MIT, Apache-2.0, GPL, CC-BY, …)
noncommercial
academic or noncommercial use only; confirm permitted use before applying the skill
restricted
licence established and it constrains reuse; check it before commercial use or redistribution
unknown
no licence for the tool was established; describing and using are unaffected; treat it as all rights reserved and never redistribute the tool’s code
The search tools do not filter or rank by tier; each skill hit prints its tier.
The router asks the agent to get confirmation of a non-commercial purpose before
applying a noncommercial skill
(LICENSE_TIERS.md). Code is
under Apache-2.0, collection content under CC-BY-4.0; verbatim quotations
(evidence_spans) are minimal and attributed, and a rights holder may request removal
(LICENSING.md). Other policies are in governance/, starting
with MAINTAINERS.md and COI_POLICY.md.
asbb report <route> drafts any of the four with the collection and the skill
filled in, redacts paths and credentials, and prints the text and a prefilled
address; it never posts. Contributions go through issues here, and a maintainer
carries them into the domain repository.
Contributing
The four routes above are described, with what happens next, in
CONTRIBUTING.md.
Propose or annotate a paper, a collection or a workflow:CONTRIBUTING.md.
Add papers to a collection. See the contributor proposal rail and the
maintainer build, view, cut and candidate rail in
Adding papers to a collection. Released collections
are listed in collections.yaml; each lives in a
domain repository asb-skills-<domain>. Domain content and candidates
belong in the corresponding asb-skills-<domain> checkout; this repository
holds tooling, governance and the registry.
Credit. Merged contributions count on the contributor leaderboard and can lead
to co-authorship of the next release DOI (CONTRIBUTION_LOOP.md).
Before packaging, run python -m scripts.unit_closure from the repository root:
it compares each shipped unit’s indexes with its leaves and checks declared
metadata.helper files.
Status
Step 2 of the per-domain split is complete: this master is the registry and tooling repository; released content is in the domain repositories.
The managed-install and Claude Code measurements predate the new marketplace sources; see docs/STATUS.md for their scope and dates.
Pending or experimental: the other seven packs, the web UI, runtime adapters other than agents, the MCP server. Not yet published: PyPI, the dataset mirror or a capsule record. The evaluation material is definitions only: no validated benchmark, reference results or agent-performance leaderboard exists. The documentation site is live.
The w3id.org/holobiomicslab/ and w3id.org/asb identifiers resolve. Dates, measurements and detail per surface: docs/STATUS.md.
Acknowledgements
We thank the participants and organisers of Dagstuhl Seminar 26181, Computational
Metabolomics, where this collaboration began. API generation costs for this
collection were sponsored by CNRS and Université Côte d’Azur. Maintained by
Holobiomics Lab (CNRS, Université Côte d’Azur).
![dependabot[bot]](https://avatars.githubusercontent.com/in/29110?v=4)
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?v=4)
评论