Back to Liquid Agent DOCUMENTATION

Skill Catalog and Selection Guide

Skills provide scientific judgement, not executable engines or a mandatory pipeline. All 34 maintained packages live in this repository's skills/ tree. The LLM chooses relevant guidance from the question, conversation and observed data. The situations below describe selection cues, not hard-coded triggers. An unrelated question should not start an analysis merely because a skill matches.

Three Levels of Guidance

The root orchestrator preserves intent and user control. The assay router helps identify the molecular material and representation. Specialists supply assay constraints, while shared-method skills can accompany any compatible specialist.

Skill ID	Why it exists and what it contributes	Typical automatic selection cue	Example user request
liquid-biopsy-analysis	Prevents generic advice and fixed end-to-end scripts; connects evidence, user intent, capabilities and review.	Scientific planning, interpretation or authorized execution. Also loaded as an ancestor.	"Suggest two useful next steps, but do not run anything."
reflective-learning	Reviews reusable conversational feedback and plans private skill revisions, including its own guidance.	Durable user preferences, recurring corrections or explicit retrospectives.	"In future, show uncertainty before conclusions; suggest skill changes for review."
memory-curation	Quietly separates user, task and linked-task memory; updates changed preferences and ages obsolete task-derived habits.	An explicit durable preference/background update, conflict, forgetting request or memory recovery.	"I am now leading the study, so use a more technical level from now on."
assay-routing	An extension is not an assay. Separates molecular material, measurement and file representation.	Mixed folders, ambiguous tables, a newly encountered assay; ancestor of specialists.	"Which of these files are RNA counts and which are methylation measurements?"

Shared Methods

Skill ID	Motivation and role	Selection cue	Example user request
local-model-setup	Guide hardware fit, private inference profiles and official local runtime setup; bootstrap does not require an LLM.	Local model installation, switching or troubleshooting.	"Which local models fit this computer, and how can I verify tool calling?"
task-memory	Maintains task context and explicit durable preferences without bypassing skill review.	Meaningful decisions, task continuation, or an ongoing preference.	"Remember that I prefer limitations first in future reports."
task-handoff	Preserves source-specific context, artifacts and unfinished work.	Explicitly selected conversations to link.	"Summarize these tasks separately before combining them."
linked-task-synthesis	Integrates evidence while checking contradictions and cohort overlap.	A linked conversation or cross-task question.	"Are these two results independent evidence? Do not rerun anything."
data-intake	Protect originals, inspect archives and establish sample/provenance identity before computation.	Attachment, extraction, mixed sources, incomplete downloads or malformed files.	"Inspect the folder and tell me what is missing; do not extract yet."
local-data-privacy	Keep local records separate from the remote model while providing useful aggregate evidence for decisions.	All data inspection, interpretation and reporting.	"Use local tools to summarize QC; keep individual records and complete tables local."
cohort-design	Avoid leakage and invalid group comparisons; preserve patient, time, replicate and batch identity.	Metadata choices, longitudinal samples, integration, model validation.	"These are paired visits. Is a random sample split appropriate?"
feature-encoding	Choose a compatible representation and encoder; distinguish reference sequence from patient sequence.	Embeddings, encoding, feature stores, cross-assay integration.	"Can I encode these intervals without a reference genome?"
result-region-followup	Preserve exact figure-region provenance and separate visual explanation from numerical subset follow-up.	A result crop or reference to selected marks.	"Analyze only the measurements selected in this heatmap."
scientific-visualization	Choose informative axes, units, missingness displays and representative views; verify real figures.	Plot requests, result interpretation, report assembly.	"Show the distributions and a representative locus; state any sampling."
scientific-reporting	Turn measured outputs into an English scientific narrative, not a log transcript.	Report writing, a conclusion, summary of several completed steps.	"Write a short report with the actual tables, figures and limitations."
literature-review	Date and critically appraise primary evidence; prevent invented citations and overclaiming abstract access.	New methods, publication comparison or current literature requests.	"Search recent plasma cfRNA papers and distinguish abstracts from full text."
genetics-dna-analysis	Coordinate variant interpretation, annotation and germline/CHIP limitations across tasks.	Variant features, VAF explanations, DNA-specific follow-ups.	"Could these plasma variants originate from blood cells?"
genomics-epigenomics	Connect genome-wide and region-level questions while preserving assay-specific semantics.	Joint coverage, chromatin, methylation or genomic feature questions.	"Can we compare enrichment and coverage without treating both as methylation percentages?"
cancer-research	Frame biomarkers and monitoring as research evidence with appropriate uncertainty.	Cancer labels, response monitoring, subtype or clinical-sounding conclusions.	"What can this exploratory separation tell us, and what can it not establish?"

Assay Specialists

Skill ID	Motivation and role	Selection cue	Example user request
raw-sequencing	Assay-aware read QC, reference/alignment and UMI prerequisites; no arbitrary FASTQ-to-result shortcut.	cfDNA FASTQ, BAM or CRAM preparation.	"What must be checked before these paired-end reads can support fragment analysis?"
genomic-tracks	Distinguish BED/bigBed intervals from quantitative bigWig/bedGraph signals and nucleotide sequence.	Genome browser tracks, representative loci or interval distributions.	"Plot these bigBed tracks, but do not claim the interval widths are fragment lengths."
fragmentomics	Interpret true fragment lengths, ends and nucleosome-related signals with library provenance.	Fragment histograms, paired-end alignment, end motifs or nucleosome profiles.	"Compare fragment length distributions and flag library-related confounding."
ctdna-variants	Review VAF, depth, error suppression and CHIP/germline evidence.	Plasma SNV/indel tables, VCF/MAF, serial VAF.	"Summarize the variants and explain what a missing matched normal prevents us from concluding."
copy-number	Separate depth variation from copy number and tumour fraction; require normalization evidence.	Low-pass WGS, bins, segments or CNV matrices.	"Could this apparent copy-number difference instead be GC or coverage bias?"
methylation-bisulfite	Keep methylated/total counts, coverage and conversion QC distinct from enrichment signals.	Bisulfite/enzymatic base-level calls.	"Which CpGs have sufficient coverage for a comparison?"
methylation-enrichment	Interpret capture/enrichment counts with controls rather than as methylation percentages.	cfMeDIP-seq, MeDIP or MBD regions.	"Review enrichment QC and propose a count-based comparison."
methylation-arrays	Review probe annotation, beta values, detection QC and batch; do not substitute proxy ratios for normalization.	IDAT files, beta/M-value or paired intensity matrices.	"Inspect missing probes and beta distributions before differential analysis."
cell-free-rna	Apply RNA count models, library strategy and contamination constraints to plasma RNA.	Gene/transcript counts, TPM or RNA reads.	"Run count QC first. After review, compare the declared independent groups."
small-rna	Respect adapter/length, isomiR, haemolysis and normalization specifics.	miRNA/small-RNA libraries or processed counts.	"Can these normalized miRNA abundances be used as raw counts?"
plasma-proteomics	Keep identification, platform, abundance scale and missingness explicit.	Protein abundance or affinity-platform tables.	"Plot protein distributions without replacing non-detections with zero."
plasma-metabolomics	Separate feature signals from confident metabolite identity; consider blanks and drift.	LC/GC-MS or NMR feature tables.	"Which QC metadata do we need before interpreting these metabolite features?"
extracellular-vesicles	Apply isolation/characterization and cargo principles; molecular data alone do not establish EV origin.	Blood EV counts, cargo RNA/protein tables.	"Can this cargo profile establish tumour-derived vesicles?"
circulating-tumour-cells	Maintain enumeration definition, volume denominators and enrichment bias in numeric/molecular measurements.	CTC counts or molecular profiles.	"Convert these supplied counts and volumes to cells/mL with uncertainty."
digital-pcr	Distinguish occupancy, concentration, uncertainty, saturation and detection limits.	Accepted/positive partition counts, ddPCR concentrations or mutant/WT assays.	"Quantify these partitions and retain uncertainty for the zero-positive wells."

Manual Loading, Precisely

Every ID above can be inspected explicitly:

liquid-agent skills load default:digital-pcr
liquid-agent skills load default:digital-pcr references/evidence.md
liquid-agent skills tree methylation --json

In the interactive shell use /skills load default:digital-pcr. These commands display the instructions; they do not run an assay and do not permanently pin the package into every future LLM turn. In Web, expand Skills, select a package and read its instructions or source notes. Viewing is also not execution.

To ask the agent to apply guidance to the current conversation, say:

Load default:digital-pcr and use it to assess this assay. Explain the required controls before proposing any analysis. Do not run anything yet.

The controller can then call load_skill; parent guidance is included automatically. A full-load receipt records the instruction hashes. A reference request loads one listed source note, not arbitrary external files. Repeating a view or load does not create a result report or delete a previous plan.

Knowledge Versus Engines

For supported explicit tables, the assay-table engine adds dPCR/CTC quantification, processed-matrix QC and an optional PyDESeq2 count contrast. The specialist still checks whether those operations answer the user's question. A beta-table QC engine is not an IDAT preprocessing engine; a protein abundance plot is not raw mass-spectrometry identification.

For broader capabilities see the capability matrix. For inheritance, storage, learned notes and safety see the professional skill guide.