hchouTW
@hchouTW
Public Skills
ams-analysis
by hchouTW
Use for the Alpha Magnetic Spectrometer (AMS-02) on the ISS - its Tracker/permanent magnet, TOF, TRD, ECAL, RICH and ACC; rigidity/charge/velocity/mass observables; and AMS-style charged cosmic-ray analyses (proton/He/nuclei fluxes, B/C-type ratios, positron/electron, antiproton, antihelium/antideuteron searches, deuteron/isotope separation). Covers AMS-specific selections, data quality, efficiency, acceptance, backgrounds, charge confusion, template fits, unfolding, systematics, likelihoods, time-dependent (solar-cycle, periodicity) analyses, structured analysis specifications and formal reviews, and source/"latest result" verification. Not for generic HEP, generic statistics, or other experiments unless the question is tied to AMS.
academic-diagrams
by hchouTW
"Use when generating, converting, reviewing, or simplifying academic diagrams and schematics for papers, theses, posters, and talks: research/analysis workflows, algorithm flowcharts, system and agent architectures, data-flow and sequence diagrams, model-architecture figures, causal/conceptual frameworks, experimental-setup schematics, and turning Mermaid/DOT/PlantUML into publication-quality SVG, PDF or TikZ. First-class HEP (analysis pipelines, Monte Carlo chains, decay trees, Feynman diagrams), statistics (plate notation, causal DAGs, Bayesian workflows) and CS architectures. Outputs editable sources plus caption and validation notes. Triggers on 'draw', 'diagram', 'flowchart', 'schematic', 'plate notation', 'causal DAG', 'Feynman diagram', 'architecture figure', 'Mermaid to SVG', 'turn this Methods section into a workflow', 'review this causal DAG', 'PlantUML to PDF', 'data flow of our training setup'. Not for data plots (histograms, spectra, ROC curves): see academic-papers and hep-analysis."
academic-papers
by hchouTW
"Use for any scientific-paper task, even a quick question - load it before answering from memory: summarizing or critiquing a paper or abstract, judging whether a claim or significance is justified, interpreting a figure, finding literature, lit reviews and related work, checking that a citation supports a claim or number, drafting or tightening a section (title, abstract, results, systematics), referee reports, responses and rebuttals, Comments/Replies and errata, formatting for a venue (REVTeX, JHEP, JCAP, EPJC, AASTeX, NeurIPS/ICML/ICLR/ACL, TMLR) and its length limits, BibTeX cleanup (INSPIRE-HEP/ADS/arXiv/DBLP), reproducibility, notation and consistency audits, turning research code into a methodology write-up, systematic-review screening, figure/table design, posters/talks, theses, proposals, data/code/ethics statements. Physics/HEP, astroparticle and statistics/ML conventions. Not for running fits or limits (hep-analysis), AMS-02 analyses (ams-analysis) or drawing diagrams (academic-diagrams)."
task-authoring
by hchouTW
"Use when asked to create or refine a development task, turn a short request into an implementation-ready Task Markdown document, or write a task/ticket/spec that another human or AI agent can implement without the original conversation. Triggers on 'create a task for', 'write a task/ticket for', 'draft a spec for X', 'turn this into a task', or a bare feature/bug/performance/research request that needs Background, Scope, Repository Context and Acceptance Criteria before anyone starts coding. Works across Claude Code, Codex and other repository-aware agents. Also covers authoring a canonical worked example for a skill's examples/ directory (eight archetypes, e.g. Contrast, Test-First, Postmortem), and designing or budgeting an LLM-call prompt (few-shot/chain-of-thought, structured-output contracts, token budgets, model tiers, agentic loops with ReAct, self-healing feedback and max-iteration guardrails). Not for implementing a change or scoping it live in the current conversation (use agile-development)."
deep-learning
by hchouTW
"Use for any PyTorch task, even a quick question or a one-line error - load it before answering from memory: writing, debugging, reviewing or explaining nn.Module models, training/eval loops, DataLoaders, losses, optimizers, mixed precision, gradient accumulation, checkpoint/resume, DDP/FSDP, determinism, inference, speed/memory; shape/device/dtype errors, NaN or exploding loss, slow dataloaders; Transformers, RNN/LSTM, VAE/GAN/diffusion, transfer learning, LoRA/quantization, autograd/hooks, export (ONNX/torch.compile). Also ML engineering decisions: architecture and scaling, parallelism strategy, compute/cost budgets, ablations and seed noise, dataset splits and leakage, eval design, serving capacity, drift, calibration (ECE), distribution shift, physics-informed/equivariant ML for HEP. PyTorch only, not TensorFlow/JAX/scikit-learn."
hep-analysis
by hchouTW
"Use when analyzing collider or astroparticle physics data/simulation, including quick formula lookups, how-to or conceptual questions - load it before answering from memory: ROOT C++/PyROOT/uproot/awkward/RDataFrame (crashes, CMake builds); event selections, cutflows, histograms, fits, efficiencies, yields, unfolding, systematics, ABCD/data-driven backgrounds, limits; RooFit/RooStats/pyhf/Combine (impacts, datacards). Jets/JES/JER, b-tagging, MET, pileup, tag-and-probe, BDT/NN classifiers. Detector subsystems and resolutions (tracker, ECAL/HCAL, TRD, TOF, RICH, dE/dx, muon), PID, reconstruction, Geant4, calibration. Astroparticle/cosmic-ray physics: spectrum and composition (knee/ankle/GZK, X_max), air showers, ground arrays, IACT, neutrino telescopes (effective area), space-based detection (geomagnetic cutoff, solar modulation), multi-messenger, Li & Ma/trials statistics, cosmic-ray flux from counts/exposure. Not for AMS-02-specific analyses (use ams-analysis) or unrelated 'root' (Linux/Android, certs)."
agile-development
by hchouTW
"Use when making any non-trivial software change - features, bug fixes, refactors, endpoints, UI work, migrations, dependency updates - turning a request into a small, verified, reviewable increment with acceptance criteria and a validation plan. Also covers reviewing a PR, incident response/postmortems, working safely in legacy code without tests, architectural decisions (boundaries, dependency direction, data ownership, ADRs), design docs, estimation and sprint planning (sizing stories, fitting a backlog to a sprint), and feature-flagged rollout planning. Triggers on 'implement', 'fix', 'add', 'refactor', 'migrate', 'update', or scoping/reviewing a change, without saying 'Agile'. Also covers implementation discipline (ask vs. assume on ambiguity, minimal changes, avoiding scope creep, verifiable done) - trigger for 'is this overcomplicated', 'am I over-engineering this'. Not for writing a standalone ticket/spec document for someone else to implement (use task-authoring)."