MAGIC Agent Skills is now open source! Star on GitHub
MAGIC Agent SkillsMAGIC Agent Skills
Skills

Skills Reference

The Linguistic Agent Skills suite contains 18 skills organized across a 5-phase pipeline. The linguistic-orchestrator coordinates routing; 14 specialist skills own phase-specific content; 3 optional Mindset stubs cover Phase 4 decisions.

Pipeline Overview

Phase 0: Scope    → scope, scripts, tokenize, ethics
Phase 1: Acquire  → corpus, bitext, transfer
Phase 2: Analyze  → morph, syntax, annotate, semantics, discourse, speech
Phase 3: Evaluate → eval
Phase 4: Optional → codeswitch, historical, lexicon
         + orchestrator (coordinates all phases)

Skills by Phase

Phase 0 — Scope

Identify the target language precisely and set strategic direction before touching any data.

SkillScorePurpose
linguistic-scopeA− (105)ISO 639-3 + Glottolog resolution, Joshi classification, URIEL typological profiling, transfer-source selection
linguistic-scriptsA− (104)Unicode normalization policy (NFC/NFKC), confusable folding, diacritic preservation for tone languages
linguistic-tokenizeA− (104)Fertility audit, SentencePiece config, vocab-extension method (FOCUS/OFA/HyperOfa)
linguistic-ethicsA− (106)CARE/FPIC, license audit, sacred-text gating, attribution registry — runs at Scope AND Release

Phase 1 — Acquire

Gather monolingual and parallel data ethically and reproducibly.

SkillScorePurpose
linguistic-corpusA− (103)Catalog (OLDI/CulturaX/MADLAD-400/Glot500), paragraph LID, MinHash dedup, contamination audit
linguistic-bitextA− (102)LASER3/SONAR mining, Vecalign alignment, margin threshold tuning, synthetic bitext
linguistic-transferA− (105)LoRA rank by URIEL distance, MAD-X adapters, forgetting mitigation, tool selection

Phase 2 — Analyze

Run linguistic analysis layers needed for evaluation, augmentation, or training.

SkillScorePurpose
linguistic-morphA− (102)UniMorph, SIGMORPHON segmenters, FST/HFST, paradigm-completion augmentation
linguistic-syntaxA− (102)UD treebanks, cross-lingual parser transfer, agreement-probe construction
linguistic-annotateA− (103)IAA metric selection (κ/α/γ), guideline authoring, adjudication, active learning
linguistic-semanticsA− (102)WordNet/OMW coverage, FrameNet/PropBank SRL, MWE/PARSEME, semantic-equivalence eval
linguistic-discourseA− (102)RST/PDTB/GUM frameworks, coreference (zero-anaphora for pro-drop), coherence eval
linguistic-speechA− (101)ELAN/Praat/FLEx → Lhotse CutSet, G2P/IPA, MMS/Whisper ASR, VITS TTS

Phase 3 — Evaluate

Honestly measure performance with metrics fit for the target language.

SkillScorePurpose
linguistic-evalA− (104)chrF++/COMET/GEMBA-MQM, BLiMP-style probes, contamination-aware reporting, per-dialect breakdowns

Orchestrator

SkillScorePurpose
linguistic-orchestratorA− (102)Entry point; phase routing; workspace state management

Phase 4 — Optional (Mindset Stubs)

Activate when the specific scenario applies.

SkillScorePurpose
linguistic-codeswitchB+ (97)Code-switching awareness for Hinglish/Spanglish/Singlish/MSA+dialect communities
linguistic-historicalB+ (97)Cognate sets, Swadesh lists, sound correspondences for Class 0–1 bootstrap
linguistic-lexiconB+ (98)Dictionary-building, sense splitting/lumping, MWE inventories for RAG/MT post-edit

Quality Scores

All scores from skill-judge (8-dimension, 120-point rubric), snapshot 2026-04-23. Entry-point skills required A− (≥102/120); specialist skills required A−; Mindset stubs required B+ (≥96/120).

Shared Utilities

The _linguistic_shared/ directory contains interaction_utils.py and findings_presenter.py — shared utilities used across all skills. See Shared Utilities for details.

Was this page helpful?
Edit on GitHub

Last updated on

On this page