Skills

English · 简体中文

Skills

A skill is a packaged workflow — instructions, conventions and often helper scripts — that the agent loads when your request matches it. Wisp ships 42 of them.

You do not invoke skills by name. You describe the task ("fold this sequence", "find papers that contradict this", "make this figure publication-ready") and the matching skill loads itself. Each skill's description exists to make that matching reliable.

Skills live in skills/ in the repo and are bundled into the installer, so a release works without the source tree.

---

Structure prediction and protein design

GPU-backed models for folding, docking and inverse folding. All are in the biomodels category; all but the MPNN family declare requirements: [gpu].

Skill	What it does	GPU
alphafold2	Monomer and multimer structure prediction via the ColabFold runner. Also validates designed sequences by self-consistency pLDDT, ipTM and RMSD.	✅
openfold3	Structure prediction with OpenFold3, an open-weights PyTorch reproduction of AlphaFold3 (AlQuraishi Lab).	✅
boltz	Protein / nucleic-acid / small-molecule complexes with Boltz-2.	✅
chai1	Complexes with the Chai-1 foundation model.	✅
esmfold2	Biohub ESMFold2 all-atom co-folding — single-sequence and MSA modes; protein, DNA, RNA, ligand.	✅
diffdock	Blind diffusion docking of a ligand into a protein with DiffDock-L.	✅
proteinmpnn	Inverse-fold a backbone into sequence with ProteinMPNN.
ligandmpnn	Inverse folding with ligand, nucleic-acid and metal context.
solublempnn	Inverse folding biased toward cytosolic solubility.

⚠️ alphafold2 sends your sequence to a third party. It uses the public MMseqs2 server at api.colabfold.com for MSA generation. That host has no published terms of service or privacy policy. Every skill declares its third-party weights and services in its own frontmatter — check there before running anything on sequences you cannot share.

Genomics and single-cell

Skill	What it does	GPU
evo2	Score, embed and generate DNA with Evo 2, a long-context genomic foundation model.	✅
borzoi	Predict genome-wide functional tracks (RNA-seq, CAGE, DNase, ChIP) from sequence; score regulatory effects of variants.	✅
fair-esm2	Per-residue and per-sequence protein embeddings with Meta AI's ESM-2.	✅
scgpt	Cell embeddings and annotation with scGPT.	✅
scvi-tools	Probabilistic scRNA-seq — scVI latent space, scANVI label transfer, Bayesian differential expression.	✅

Literature — the bear-* family

Nine skills that all share one rule: they search for real papers and never invent a citation. If nothing is found, they say so. Retrieval goes through scimaster-cli (installed by local-env-setup).

Skill	Question it answers
bear-support	"What real literature backs this claim?"
bear-counter	"What would a reviewer hit this with?" — opposing results, boundary conditions, failed replications, ranked by threat
bear-review	Both sides at once: strongest support vs. most dangerous challenge
bear-scoop	"Has someone already done my idea?" — collision map before you commit
bear-propose	Pre-project assessment: collision risk + support in the quiet zone + challenges to think through
bear-map	Concept map around a term, every node anchored to a real paper. Outputs Mermaid + standalone HTML
bear-trace	How a problem evolved — backward through predecessors, forward to current work
bear-onboard	New field, fast: concept map (space) and evolution (time) in one report
literature-review	Full multi-source reviews, from "what's the seminal paper" to synthesis

bear-support and bear-counter pair well — run both on the same paragraph to see the evidence from both directions.

Documents and figures

Skill	What it does
figure-composer	Build or improve a publication-grade multi-panel figure from a claim, data paths, or an existing image
figure-style	Correctness and legibility rules; loads before any plot and applies a role-mapped font-size and style system
journal-club-ppt	A paper PDF → group-meeting slides (文献汇报 PPT / 组会 PPT)
paper-narrative	Judge and reshape a manuscript's story — is Figure 1 a hook, is the figure order right
pdf-explore	Answer questions that need the actual contents of an attached PDF
indication-dossier	Therapeutic indication dossier: population, epidemiology, disease biology, standard of care, regulatory precedent

Environment and compute

Skill	What it does
local-env-setup	Bootstrap the local runtime — uv/Python, Node + scimaster-cli for the bear-* skills, pixi for multi-env bioinformatics
compute-env-setup	Build and validate a reproducible Python or R environment on a chosen context
probe-compute-environment	Inspect a registered server and persist its hardware, scheduler, runtime and privilege profile
remote-compute-ssh	Submit recoverable SSH-direct Runs with live progress and model-free monitoring
remote-compute-modal	Explains Wisp's Modal boundary and migrates a Modal workload to a supported SSH Run
managed-model-endpoints	Explains the managed-endpoint boundary and plans a safe integration
using-model-endpoint	Call an already-configured model endpoint and capture the inference as a Run

Extending Wisp

Skill	What it does
customize	Create a specialist or a project-local skill, using the tools Wisp actually exposes
skill-creator	Author, validate and evaluate skills — including tuning the trigger description
self-awareness	Wisp's real tool surface and runtime boundaries; consulted when deciding which tool can do a job

Other

Skill	What it does
browser-use	Drive your real, persistent Chrome session — open, read, click, fill forms, switch tabs
agent-infini	Multi-turn data analysis via the InfiniSynapse CLI, with database/RAG context
product-self-knowledge	Reference for facts about Anthropic products, so the model does not guess

---

Writing your own

Ask for it: "make me a skill that does X" triggers skill-creator. Skills are plain directories with a SKILL.md, so a project-local skill is just a folder in your project. customize covers the specialist-vs-skill choice — a specialist is a saved agent persona, a skill is a reusable workflow.