Use for Triton kernels — masking, tl.dot, autotuning, interpreter debugging, torch ops.
Use for MPI and OpenMP — decomposition, collectives, hybrid runs, binding, scaling.
Load before modeling tabular data — gradient boosting, categoricals, early stopping, calibration.
Use for unsafe Rust and C FFI — SAFETY comments, edition-2024 unsafe rules, raw pointers, Miri.
Use for editor internals or IDE features — ropes, undo, tree-sitter, LSP/DAP, terminals.
Use to set up, check or restore backups — restic, borg, database dumps, 3-2-1, restore drills.
Use to export trained models — torch.export, ONNX, Core ML, ExecuTorch, TensorRT, parity.
Use to debug a network problem layer by layer — routes, DNS, ports, TLS, MTU, packet captures.
Use before running ffmpeg or ffprobe — encodes, trims, concat, scaling, GIFs, loudnorm, subtitles.
Use for diagrams as text — Mermaid, Graphviz, D2, PlantUML, TikZ/tikz-cd; rendering, embedding.
Use for Android or cross-platform mobile apps — Kotlin, Gradle/AGP, Compose, tests, adb.
Use for geospatial data — CRS, GDAL, GeoPandas, rasterio, PostGIS, PMTiles, MapLibre maps.
Use for JVM code — Java 21+, Scala 3, Gradle/Maven/sbt, JUnit, JMH, GC tuning, GraalVM.
Use for LaTeX (papers, theses, beamer, arXiv) or Typst — engines, math, biblatex, errors.
Use for firmware and hardware-near work — MCUs, RTOS, drivers, flashing safety, FPGA, PCB.
Use for small molecules in RDKit — standardization, fingerprints, similarity, QSAR.
Use for uv projects and packaging — lockfiles, PEP 723 scripts, pyproject layout, wheels.
Use for multiplayer netcode — server authority, prediction, rollback, interpolation.
Use for dashboards and data apps — Streamlit, Dash, Panel or BI tools, layout, caching, freshness.
Use for Swift on Apple platforms — Swift 6 concurrency, SwiftPM, Swift Testing; iOS modules.
Use for Node packages and CLIs — ESM vs CJS, exports maps, version managers, pnpm/npm.
Use before driving Adobe, ONLYOFFICE or a native app by screen — routing, screen lock, shortcuts.
Use for RAG pipelines — chunking, embeddings, hybrid search, reranking, grounding, retrieval evals.
Use for a Linux ML/dev workstation (CachyOS, Ubuntu, RTX 50) — NVIDIA, CUDA, desktop, btrfs.
Use for Tailscale on servers — serve, Funnel, tags and grants, SSH, exit nodes, certs.
Load before creating or delivering vector files — SVG as code, svgo, tracing, plotters, cutters.
Use for memory growth, leaks, peak RSS — heaptrack, Instruments, memray, dhat, massif.
Use for C or C++ — UB traps, RAII and lifetimes, concurrency, sanitizers, clang-tidy, ABI.
Use for async Python — asyncio TaskGroup, cancellation, timeouts, anyio, free-threading.
Load before writing a system prompt, agent definition, CLAUDE.md, research prompt, schema or few-shots.
Use to find, download or publish Hugging Face models and datasets — licenses, gated repos, cache.
Use for web front ends — React 19, Next.js 16, Svelte 5, Vue, Astro, Tailwind v4, Web Vitals.
Use for custom GPU kernels — Metal/MLX vs CUDA vs Triton, roofline, correctness, profiling.
Use for scientific data I/O — HDF5, NetCDF, ADIOS2, Zarr, parallel writes, checkpoints.
Use when float accuracy matters — rounding, cancellation, compensated sums, fp16/bf16/fp8/fp4.
Use to lead or request an N-solver blind-vote run — code reducer, one reconcile round; leader procedure and classes.
Use for CUDA C++ kernels — sm_120 builds, warp reductions, Nsight, torch custom ops.
Use to integrate ODEs or SDEs — tolerances, stiffness, implicit and symplectic schemes, order.
Use for quantum chemistry — PySCF, Psi4, xtb, ORCA; methods, basis sets, convergence.
Use for computational biology or chemistry — biosecurity rules; genomics to quantum chemistry.
Use for sequencing data — QC, alignment, BAM/CRAM, variant calling, VCF, RNA-seq.
Use for SLURM jobs — sbatch scripts, arrays, GPUs, accounting, Spack; submits need consent.
Use when untrusted input reaches a sink — SQL/shell injection, paths, SSRF, XSS, CSRF, deserialization.
Load before quantum-circuit work — derivations, simulation, hardware runs, Qiskit, Cirq, stim, QEC.
Use to lock down a service, container, host, CI runner or local server — least privilege, sandboxing.
Use to make a book from a manuscript — print PDF, EPUB, DOCX, KDP/IngramSpark, cover, ISBN.
Use for latency and throughput under load — k6, vegeta, oha, open vs closed models, percentiles.
Load before training or evaluating robot policies — MuJoCo, Isaac Lab, RL, VLAs, sim-to-real.
Use to fuzz parsers and untrusted-input code — cargo-fuzz, atheris, AFL++, Go fuzzing, corpora.
Use for shaders — HLSL, GLSL, MSL, WGSL, Slang; cross-compiling, PBR, debugging.
Use to run, fine-tune or evaluate open image models — diffusers, mflux, ComfyUI, LoRA.
Use for Fortran — modern style, fpm, compilers, C/Python interop, legacy modernization.
Use for Julia — juliaup, Pkg, Test/Aqua/JET, type stability, benchmarks, GPU, SciML.
Use for observability — OpenTelemetry traces, metrics, logs, the Collector, SLO alerts.
Use to deploy or run Forgejo — compose stack, app.ini, Actions runner, dumps, upgrades.
Use before generating or editing images via image-studio — SVG and raster prompts, edits, QA, costs.
Load when a training run misbehaves — NaN loss, divergence, no learning, plateaus, overfitting.
Use for Houdini — VEX, HDAs, Pyro/FLIP/Vellum/RBD, caching, Solaris/Karma, hython, PDG.
Use for crashes, hangs or memory corruption in native code — sanitizers, lldb/gdb, rr.
Use for publication figures in matplotlib or seaborn — styles, fonts, sizes, facets, export.