tao-research-skills

Use when turning recent agent-session evidence into durable skill updates. Triggers: "consolidate lessons", "update skills from sessions", "prevent repeated fixes", "recurring mistakes", "capture workflow learnings"

Use when running workloads on LUMI supercomputer, including GPU job submission, PyTorch with ROCm/AMD MI250X, container workflows, and LUMI-specific Slurm configuration. Triggers: "LUMI", "MI250X", "ROCm", "AMD GPU", "CSC", "LUMI-G"

Use when integrating W&B experiment tracking into ML training pipelines, including logging strategy, run configuration, and online/offline mode management. Triggers: "W&B", "wandb", "weights and biases", "experiment logging", "wandb.log", "wandb.init", "training dashboard"

"Use when extracting Overleaf login cookies from Chrome on macOS for the VSCode Overleaf Workshop extension. Decrypts Chrome's cookie store and outputs the session cookie in the exact format the extension expects. Triggers: "overleaf cookie", "overleaf login", "overleaf workshop cookie", "get overleaf cookie", "vscode overleaf login"

Use when preparing to launch any remote or multi-GPU ML experiment. Triggers include "preflight", "sanity check", "launch validation", "before training", "GPU check", "dataset check", "environment check"

Use when pushing changes in a repository that contains git submodules, especially when the submodule has its own commits that must not be left stranded. Triggers include "push", "git push", "submodule", "commit submodule", "workspace push"

Use when evaluating academic papers or surveying a research topic. Gathers venue, citations, GitHub stats, social buzz, reproducibility, and author signals to produce a scored markdown report. Triggers: "evaluate paper", "paper review", "research survey", "literature review", "is this paper good", "find papers on", "compare papers", "paper impact"

Use when analyzing the follow-up works of a specific paper — the papers that cite it. Resolves the seed paper via Semantic Scholar, paginates the full citing-paper list, triages by impact + recency, dispatches parallel subagents to tag each follow-up (extension / improvement / application / theoretical / criticism), and produces a clustered markdown report under docs/research/. Triggers: "followup analysis", "follow-up works", "papers that cite", "forward citations", "who cites this paper", "citation graph", "impact analysis", "downstream papers"

Use when promoting an idea from a fleeting thought into a tracked artifact with a lifecycle. Defines a per-idea on-disk directory (./idea_box/<slug>/) and a hard-gated state machine (explored → feasible/blocked → building → built/killed) that the existing reader/evaluator skills (idea-feasibility, ml-ablation-design, academic-deep-research, followup-analysis, arxiv-latex-reader, pdf-reader, github-reader, blog-reader) plug into. Public repo ships only the convention; idea data lives in the user's private idea-box repo. Triggers: "idea box", "idea-box", "new idea", "advance idea", "list ideas", "kill idea", "想法箱", "新想法", "推进想法", "列出想法"

Use when proposing new research ideas grounded in a seed paper or method — surfaces the gaps that the citation cone hasn't filled, the drawbacks the community already complains about in the official repo's GitHub issues, and the adjacent angles the literature suggests. Orchestrates followup-analysis (already-done exclusion), GitHub-issue mining (known drawbacks), and academic-deep-research (adjacent literature) into a ranked candidate-ideas report. Triggers: "idea explore", "idea-explore", "propose ideas", "explore ideas", "new directions", "research gaps", "what's missing", "build on this paper", "探索想法", "提出新想法", "研究空白"

Use when evaluating whether a research or product idea is actually feasible — buildable, evaluable, and de-risked by available checkpoints, code, datasets, and GPU budget. Normalizes the idea, gathers primary-source evidence (arXiv, GitHub, project pages, model hosts), scores it against four mandatory hard gates, and emits a verdict + falsifiable MVP. Triggers: "idea feasibility", "is this idea feasible", "feasibility check", "can we do this", "is this practical", "worth pursuing", "评估想法可行性", "可行性分析", "这个 idea 能做吗", "这个想法靠谱吗"

Use when designing ablation studies to compare model components, loss functions, or architectural choices. Covers synthetic data experiments, variant loops, production metrics, and W&B grouping. Triggers: "ablation", "ablation study", "variant comparison", "controlled experiment", "synthetic data experiment"

Use when setting up Claude Code on a new machine, configuring permissions, statusline, or plugins. Contains the standard settings.json and statusline script. Triggers: "Claude Code setup", "settings.json", "statusline", "permissions", "plugins", "enabledPlugins", "~/.claude"

Use when user says "login", "登录", "fix expired sessions", "refresh login", or needs to re-authenticate CDP browser sessions for any platform. Auto-detects expired platforms and guides interactive re-login via MCP browser tools. Triggers: "login", "登录", "CDP login", "refresh login", "expired session", "re-authenticate"

Use when setting up tmux on a new machine, looking up key bindings, debugging terminal colors, or restoring tmux dotfiles. Triggers: "tmux", "terminal multiplexer", "tmux.conf", "pane", "window split", "copy mode", "prefix key"

Use when setting up zsh and Oh My Zsh on a new machine, configuring shell plugins, PATH, or conda initialization. Triggers: "zsh", "zshrc", "Oh My Zsh", "shell config", "dotfiles", "conda init", "PATH"

Use when editing, modifying, or composing existing images with Google's Gemini image models ("nano-banana"). Covers single-image edits (add/remove/restyle), multi-image composition (outfit swap, subject-into-scene, style transfer), input encoding, preservation tricks, and iterative refinement. Triggers: "edit image", "gemini edit", "nano-banana edit", "image to image", "image composition", "outfit swap", "subject transfer", "style transfer", "inpaint", "outpaint"

Use when generating images from text prompts via Google's Gemini image models ("nano-banana") with the `google-genai` SDK or OpenRouter. Covers model choice, aspect ratio, prompt patterns, multi-candidate sampling, and retry logic. Triggers: "generate image", "gemini image", "nano-banana", "text to image", "gemini-3-pro-image", "gemini-2.5-flash-image", "imagen", "paper figure generation"

Use when running workloads on SURF Snellius supercomputer, including GPU job submission (NVIDIA A100/H100), conda/venv setup, Slurm configuration, and Snellius-specific infrastructure. Triggers: "Snellius", "SURF", "A100", "H100", "Dutch HPC"

Use when a repo should keep a single memory file in sync with generated AGENTS.md and a thin CLAUDE.md wrapper, or when memory drift needs to be checked or repaired. Triggers: "memory-sync", "sync memory", "AGENTS.md drift", "CLAUDE.md drift"

Use when installing tao-research-skills into a machine's global AI agent configuration. Symlinks every skill into Claude Code and Codex, injects a SessionStart git-pull hook for auto-updates, bootstraps memory, and appends workflow rules. Triggers: "install skills", "install to computer level", "setup skills globally", "meta-init", "cross-platform skills install", "bootstrap skills"

Use when a repo does not yet have canonical memory, AGENTS.md, or CLAUDE.md. Bootstraps the standard Codex + Claude Code project memory layout in one step. Triggers: "project-mem-init", "bootstrap memory", "initialize project memory"

Use when reading PDF papers via Claude Code CLI instead of Python converter tools. Delegates PDF-to-text conversion and summarization to Claude Code's native multimodal Read + WebFetch tools, handling multi-page papers via chunked reads and structured prompts. Triggers: "cc-pdf-reader", "claude code pdf", "read pdf with claude", "claude pdf reader", "cc pdf", "claude code paper"

Use when reading a YouTube video (especially an AI/ML interview, podcast, or technical talk) and producing a faithful, timestamped wiki entry the user can return to weeks later. Fetches the transcript via yt-dlp, sections by chapters or LLM-detected topics, summarizes per-section with parallel subagents, preserves verbatim quotes with speaker attribution, and runs a coverage test against the raw transcript. Triggers: "youtube wiki", "video wiki", "summarize this youtube", "watch this interview", "read this talk", "digest this video", "https://youtu.be/", "https://www.youtube.com/watch"

Use when capturing a single URL (X tweet, LinkedIn post, HN thread, short blog, news article, GitHub gist or issue) as a small persistent markdown entry the user can re-read later for motivation or grep across future sessions. Verbatim body + author + tags + a one-line "Why I saved this" hook. Writes to docs/wikis/YYYY-MM-DD-<slug>.md. WebFetch by default; CDP via Playwright MCP for login-walled hosts (x.com, twitter.com, linkedin.com). Redirects YouTube / arXiv / GitHub repo URLs to their specialized skills. Triggers: "insert wiki", "add to wiki", "insert this", "save this tweet", "capture this post", "add this to my wiki", "remember this link", "https://x.com/", "https://twitter.com/", "https://linkedin.com/posts/"

Use when reading a GitHub repository (especially a research code release with an accompanying paper) and producing a faithful digest that covers the implementation logic, the main insight, and the key reported results. Research-first with graceful fallback for non-paper repos. Handles arXiv link detection and delegates paper reading to arxiv-latex-reader / pdf-reader. Triggers: "read repo", "read this github", "analyze github", "digest repo", "github reader", "extract from github", "summarize this codebase", "read this code", "https://github.com/"

Use when building iOS apps with SwiftUI, UIKit bridging, SSH terminal integration, or voice input pipelines. Applies to Swift 5.9+/iOS 17+ projects using SPM, xcodegen, Citadel, SwiftTerm, or AVFoundation. Triggers: "SwiftUI", "UIViewRepresentable", "ObservableObject", "@MainActor", "xcodegen", "SwiftTerm", "Citadel", "SSH", "STT", "AVAudioEngine", "iOS app", "terminal emulator", "voice input"

Use when interacting with GitHub repos, PRs, issues, releases, or API data. Covers gh CLI usage patterns, authentication, and common queries. Triggers: "gh", "github", "pull request", "PR", "issue", "gh api", "gh pr", "gh issue", "github release"

Use when deploying Next.js apps to Vercel and costs are high, or when setting up a new Vercel project. Covers ISR-breaking patterns, function constraints, caching, Fluid Compute, build optimization. Triggers: "Vercel bill", "Vercel cost", "ISR broken", "dynamic rendering", "cache-control private", "x-vercel-cache MISS", "function invocations", "Fluid Compute", "GB-hours", "s-maxage", "stale-while-revalidate", "maxDuration", "build minutes"

Use when curating, uploading, or managing HuggingFace datasets for ML training, including offline caching, preflight verification, and data directory conventions. Triggers: "HuggingFace", "datasets", "push_to_hub", "load_dataset", "HF Hub", "dataset cache"

Use when structuring ML experiment configs with Hydra, adding new config groups, or debugging config resolution. Applies to any project using Hydra for hyperparameter management. Triggers: "Hydra", "config", "yaml config", "OmegaConf", "config groups", "defaults list", "config override"

Use when streaming large datasets from tar shards with WebDataset, replacing file-based DataLoaders, or precomputing encoder latents into shards. Triggers: "WebDataset", "tar shards", "wds", "streaming data", "ShardWriter", "latent shards"

Use when running ML training on HPC clusters with Slurm, including job submission, environment setup, monitoring, and failure triage. Applies to any GPU training workload on Slurm-managed clusters. Triggers: "sbatch", "srun", "Slurm", "SBATCH", "job submission", "HPC", "cluster", "walltime", "squeue"

Use when writing or improving CLAUDE.md, AGENTS.md, GEMINI.md, or any agent instruction file. Covers section structure, memory patterns, workflow rules, and anti-patterns. Triggers: "CLAUDE.md", "AGENTS.md", "agent instructions", "project memory", "MEMORY.md", "instruction file"

Use when designing ML training pipelines, data loaders, or inference systems. Enforces engineering discipline — no silent fallbacks, explicit errors on critical paths, config as single source of truth. Triggers: "silent failure", "fallback", "preflight", "assertion", "error handling", "fail fast", "config truth"

Use when reading large arxiv papers without context overflow. Progressive two-layer reading: index all sections (~2k tokens), then deep-read on demand. Never truncates. Triggers: "read paper", "paper sections", "section index", "progressive reading", "paper_content.json", "section summary"

Use when reading a long technical blog post (ML research, engineering deep-dives, Distill/Lil'Log-style posts) and producing a faithful, figure-aware summary. Handles context-over-limit via section-based chunking, captures important figures via multimodal Read, and runs a coverage test to catch missing information. Triggers: "read this blog", "summarize this post", "read blog", "digest this article", "long blog post", "read this article", "blog summary", "distill post", "summarize url", "chunk and summarize"

Use when reading PDF papers, reports, or long documents where text, figures, and tables must all be captured and chunk-summarized without truncation. Converts PDF to a markdown + paper_content.json workspace, extracts figures and tables as standalone files, then delegates to arxiv-latex-reader's progressive two-layer reading (section index + on-demand deep reads). Triggers: "read pdf", "pdf to markdown", "summarize pdf", "pdf paper", "extract figures from pdf", "extract tables from pdf", "marker pdf", "pymupdf", "docling", "paper digest", "pdf reader"

Use when evaluating generative models — choosing metrics (FID, IS, KID, sFID, FDD, FVD, PRDC, LPIPS, SSIM, AuthPct, Vendi), setting up online or offline evaluation, feature extractor selection, distributed computation, memory management during sampling. Triggers: "FID", "IS", "KID", "inception score", "frechet", "LPIPS", "SSIM", "evaluation metrics", "generative evaluation", "FVD"

Use when optimizing PyTorch training speed or memory on CUDA GPUs — global flags, torch.compile, fused optimizers, mixed precision, gradient checkpointing, kernel fusion, memory layout, or latent-space training. Applies to any PyTorch training workload. Triggers: "torch.compile", "TF32", "fused optimizer", "mixed precision", "bf16", "fp16", "gradient checkpointing", "Triton kernel", "CUDA flags", "GPU slow", "GPU memory"