agent-skills

Opinionated, scale-aware git workflow — Conventional Commits (English even under Chinese prompts), linear history via `pull --rebase` + fast-forward, when to commit to `main` vs open a PR, GitHub-Flow, worktrees for parallel agents, SemVer release tags, and pre-merge secret hygiene. Use when committing, branching, merging, opening a PR, tagging a release, setting up worktrees, or cleaning up stale branches — incl. "幫我 commit / 整理 git", "should I branch or PR", "git 工作流".

Set up and operate a research-experiment memory for ML/DL/Quant projects — LEDGER.md of numbered findings (what we currently believe, evidence links, overturn protocol), ROADMAP.md experiment queue triaged by expected payoff × cost × category × assumption dependencies, INBOX.md human scratchpad swept into the roadmap via targeted questions, an auto-rendered Mermaid map of how experiments/findings/queued work reference each other, and per-experiment REPORT.md with pre-registration, single-axis ablation contract, results tables and provenance blocks (git SHA, config hash, data window, seeds, optional MLflow refs). Use when the user wants to track which directions were explored vs not, record which succeeded or failed, avoid re-running dead-end experiments or wasting compute, make results discoverable and comparable, keep experiments reproducible, re-prioritize planned work when a conclusion changes, get a big-picture map of all experiments, jot a raw research idea somewhere, or asks to sweep the experiments inb

Commit SpecStory chat transcripts (`.specstory/history/*.md`), Claude Code plan files (`.claude/plans/*.md`, `plansDirectory`), and other coding-agent artifacts (`.cursor/plans/`, `.cursor/rules/`, `.opencode/plans/`, `.specify/`, `.codex/`) alongside the feature diff they produced — without leaking `.env` contents, API keys, or private-key PEM blocks into git history. Use when the user says "commit my chat", "save this specstory session", "stage the plan file", "scrub the transcript", "my .env leaked in chat", "bootstrap pre-commit for this project", or when you notice untracked `.specstory/history/*.md` or `.claude/plans/*.md` files while running `git status`. Also use after an accidental push of a secret to enforce rotate-first, rewrite-last remediation instead of reflexive `git push --force`.

Capture acceptance evidence — screenshots, terminal/screen recordings, HTTP logs — for a just-built feature into a gitignored .evidence/ bundle keyed to git branch/commit + coding-agent session, for async human validation (Cursor "Demos over diffs" style). Use when the user says "record a demo", "leave evidence", "capture a screenshot/video", "prove it works", wants async sign-off / 驗收, or asks to show a feature works without reading the diff or booting the app by hand. Playwright/asciinema/curl/ffmpeg.

Generic Slurm/HPC know-how for any cluster. Use when the user writes or debugs sbatch/srun batch scripts, requests resources (--mem/--cpus-per-task/--gres/--time/partitions), reads job state with squeue/sacct/sinfo/scontrol, cancels jobs, or reasons about resource isolation — cgroup CPU/RAM limits and GPU GRES (gpu vs shard vs mps vs MIG) and which of them actually fence VRAM so a misbehaving job fails alone.

Discover and drive a running Clash/mihomo proxy through its external-controller API — show the current node/mode/latency, switch proxy groups, set mode (rule/global/direct), toggle TUN, reload config, close connections, and toggle the OS system proxy. Use when the user asks to check or change their Clash/mihomo/Clash Verge (Rev)/ClashX proxy, switch nodes, test node latency, enable TUN, set up Service Mode or Mixin, toggle the system proxy on macOS/Linux, or fix an unreachable Clash API.

Bootstrap MkDocs Material docs sites with optional GitHub Pages deploy, uv-pinned tooling, llms.txt/copy-to-LLM support, page/nav helpers, and mkdocs-static-i18n languages such as zh-TW. Use when the user asks to set up docs, publish docs to GitHub Pages, create an MkDocs site, turn README or markdown notes into a site, add bilingual/multilingual docs, add zh-TW/Traditional Chinese, i18n, or translate docs. Consent-gated; records repo preferences and never auto-migrates existing docs.

FastAPI + AI engineering interview preparation — 100 self-contained Q&A across 10 topics (ASGI, Pydantic, REST, dependency injection, databases, security, testing, async, AI/ML/RAG/LLM serving, deployment) plus a quiz CLI. Use when preparing for a FastAPI/Python backend or AI-engineer interview, running or being given a mock interview, drilling a specific topic, or self-testing against senior-level answer rubrics.

Production patterns and gotchas for building FastAPI services, especially AI/ML/LLM serving. Use when you build, review, or debug a FastAPI app: choosing def vs async def, loading models in lifespan, preventing event-loop blocking, setting response_model to stop data leaks, enforcing object-level authz in queries, wrapping LLMs with streaming + Pydantic validation loops, or designing RAG/embedding endpoints.

Scaffold a production-grade FastAPI AI/ML service from a bundled, opinionated skeleton. Use when starting a new FastAPI inference/LLM/RAG backend, bootstrapping a FastAPI project with clean architecture, or setting up a model-serving API with the production basics pre-wired: router/service/repository layering, lifespan-loaded model + httpx client, /health + /ready probes, JWT auth, SQLModel + Alembic, an SSE LLM gateway, guardrails + a Pydantic validation loop, tests with dependency_overrides, Docker, and gunicorn.

Author a new agent skill or refactor an existing one to follow agentskills.io best practices — gotchas sections, output templates, validation loops, calibrated specificity (fragility-based), and agentic script design (--help, --dry-run, structured stdout, stderr diagnostics, PEP 723 inline deps, pinned uvx/npx versions). Use whenever the user wants to create a new skill from scratch, scaffold a SKILL.md, write a reference file, design a script meant to be invoked by an agent, lint a draft skill for quality, or convert an ad-hoc workflow into a reusable skill. For evaluating skill output quality with test cases, benchmarking, or optimizing the description trigger rate, defer to the `skill-creator` skill instead — this skill focuses on authoring, not evaluation.

Design and verify exercisable surfaces on apps, CLIs, services, and config. Use when authoring a new CLI/tool/library/service in Python (uv + tyro / click / argparse), Node (commander/yargs), or Bash to ensure it exposes `--help`, `--dry-run`, `--print-config`, isolated-state smoke entrypoints, and explicit exit-code contracts. Also use when editing app/tool config, CLI args/env parsing, dotfiles such as chezmoi, Ansible playbooks, CI/deploy manifests, or generated/rendered config — forces syntax/schema, then app-native loader/parser/debug/dry-run, then narrowest harmless runtime smoke. The invariant: a surface that cannot be exercised cheaply and harmlessly was not actually verified, regardless of how much linting passed.

Set up a structured project memory for any software project — TODO.md as priority/effort-tagged index of future work, backlog/ for resume-friendly research/design notes on P? items, and pitfalls/ as a symptom-grep-able knowledge base of past traps. Use when a user wants somewhere to record "maybe later" ideas, freeze troubleshooting state, capture trade-off analysis, or stop re-debugging the same problem.

Drive Nukesor/pueue (https://github.com/Nukesor/pueue) for queued, parallel, scheduled, and lightly-DAG'd shell jobs — wraps `pueue add --after`, `pueue status --json`, `pueue log --json`, group-level parallelism, and `pueued` daemon health. Use when the user wants to background long-running shell commands across reboots, queue dozens of jobs with capped parallelism, run a fan-out / fan-in pipeline of shell steps, says "pueue", "pueued", "pueue add", "pueue queue", "pueue group", "task queue for shell", "background this job", or asks how to schedule/parallelize CLI work without a real orchestrator (Airflow/Prefect/Dagster). Good fit for ML sweeps, long-running data pipelines, batched evaluations, scheduled `--delay` jobs, "wait for X then run Y" sequences.

Set up and operate MLflow for ML experiment tracking, model registry, and LLM tracing in any Python project. Covers three deployment modes — local file (`./mlruns`), local SQLite + `mlflow ui` (recommended for solo experiments), and a production Docker Compose stack with PostgreSQL + MinIO (vendored as a ready-to-copy `assets/docker-compose-stack/`). Use whenever the user wants to log params/metrics/artifacts, manage models with aliases (Champion/Challenger), enable framework autologging (sklearn/pytorch/lightning/xgboost/transformers/etc.), trace LLM calls (OpenAI/Anthropic/LangChain/LlamaIndex/DSPy), spin up a self-hosted tracking server, or asks about `mlflow.set_tracking_uri` / `mlflow.start_run` / `mlflow.<framework>.autolog` / `mlflow ui` / `MLFLOW_TRACKING_URI`. Always references the official docs at https://mlflow.org/docs/latest and the upstream repo https://github.com/mlflow/mlflow — MLflow ships fast, prefer fetching live docs over relying on memory.

Set up and operate a DVC (Data Version Control) workflow for ML projects — `dvc init`, `dvc.yaml` pipelines, `params.yaml`, `dvc exp run --queue` for parallel sweeps with metrics auto-bound to ephemeral git commits, and remote storage (S3/SSH/GDrive). Use whenever the user wants reproducible ML pipelines, data/model versioning that lives alongside git, parameter sweeps without standing up a tracking server, queued/parallel experiment execution, or asks about `dvc.yaml` / `dvc exp run` / `dvc queue` / `params.yaml` / `dvc add` / `dvc push` / `.dvc/cache`. Always references the official docs at https://dvc.org/doc and the upstream repo https://github.com/treeverse/dvc (Iterative was acquired by Treeverse in 2024 — `pip install dvc` resolves to this repo).

Prepare a marimo notebook to run as both an interactive UI and a scheduled batch script, using Tyro for CLI parsing (with dataclass or Pydantic params) and MLflow for experiment tracking. Use when the user wants the same marimo notebook to serve as both a UI for iteration and a CLI script for batch jobs / hyperparameter sweeps, and prefers Tyro over `mo.cli_args()` and MLflow over Weights and Biases. For the upstream W&B + `mo.cli_args()` variant, see marimo-team/skills/skills/marimo-batch.

assists in designing, evaluating and implementing factor-based investment strategies with Python