Latest Agent Skills

Search SAM.gov for live notices matching a pursuit row (agency, incumbent, NAICS keywords) and save results to sam_scan.md. Use for admin monitor setup, presolicitation watch, or when user asks what is live on SAM for an expiring contract.

Run an extremely strict maintainability review for abstraction quality, giant files, and spaghetti-condition growth. Use for a thermo-nuclear code quality review, thermonuclear review, deep code quality audit, or especially harsh maintainability review.

Get a second opinion from Claude Code. Use from Codex when you want to cross-check your reasoning, validate a plan, or get an independent review from Claude.

Detect which HPC cluster you are on and load the correct job submission instructions. Use when the user asks to submit, run, or schedule a job, or when you need to know which cluster environment you are in.

Submit one or more wandb-logged training/finetuning runs to the HPC scheduler. `WANDB_PROJECT` is fixed per repo (snake_case basename); `WANDB_RUN_GROUP` is picked per invocation. The training script must take the experiment/group name as a config key (e.g. Hydra `meta.experiment_name=<group>`); the skill passes it on the command line. The working tree is committed first so each run pins to a real SHA. Delegates SLURM/PBS templating to `cluster-instructions`. Use when the user asks to submit, queue, launch, or kick off a wandb training/finetuning job.

Upload local image files to a Notion page using the Notion File Upload REST API. Use when the user wants to embed local images (PNG, JPG, etc.) into a Notion page.

Monitor SLURM/PBS jobs and their logs. Use when the user asks to monitor, check, or watch a submitted job. Automatically detects running/pending jobs, tails logs, and reports errors.

Address Codex review comments on an open PR, re-run thermo-nuclear-code-quality-review, sync with the base branch, and push the updates. Use after a Codex (or human) review has landed on a PR you opened — the skill resolves each comment in code, marks the GitHub review thread resolved, then rebases/merges any new base-branch commits before pushing.

Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".

Creates a GitHub pull request with an auto-generated summary, running mandatory pre-commit and code-quality review gates, plus optional deep-review subagents, and posts findings into the PR body.

Package an in-progress task on this cluster so a Claude agent on another cluster (MMM Young, CX3, HX1, Isambard) can pick it up, or resume a task handed off to this cluster. Use when the user says hand off / migrate / continue this on <cluster>, or asks you to pick up an existing handoff.

Fetches a BibTeX citation for a given DOI using doi.org content negotiation.

Tears down a merged PR's worktree and feature branch, verifying merge status on GitHub before deleting local and remote branches and switching back to the main worktree.

Delegates scientific literature searches to the Edison API, which autonomously searches hundreds of papers and returns a cited answer. Handles query submission, polling, and reporting results.

Delegates a task to OpenAI's Codex CLI for an independent review or second opinion, with guidance on prompt construction and handling output.

Benchmarks LLM bias by scoring responses to automatically generated open-ended questions across sensitive attributes, producing a composite fitness score from 0 to 5.

Benchmarks large audio-language models against adversarial audio attacks using the CAA dataset, computing WER, ROUGE-L, cosine similarity, and coherence scores to assess robustness in conversational settings.

Benchmarks language model agents on the C2C multi-agent negotiation task, reporting win rate across starting positions.

Evaluates speech language models on beyond-semantic speech attributes such as dialect comprehension, multi-turn context memory, emotion perception, age-aware response generation, and non-verbal cue handling, reporting accuracy and judge-based scores.

Benchmarks energy-function-based safe control algorithms on the BIS (Benchmark of Interactive Safety) dataset, scoring safety, efficiency, and hybrid performance in human-robot and robot co-working scenarios.

Evaluates social bias in question-answering models using the BBQ benchmark, measuring accuracy and a bias score across ambiguous and disambiguated contexts to reveal reliance on stereotypes.

Benchmarks zero-shot in-context learning on BIG-Bench Hard multiple-choice tasks, comparing self-generated demonstrations against direct prompting and chain-of-thought baselines, and reports accuracy.

Evaluates open-ended generation quality of multilingual LLMs across brainstorming, planning, and long-form tasks, using AYA and DOLLY datasets with qualitative fluency and quality scoring.

Benchmarks automatic speech translation and recognition on English-French and English-Romanian datasets, reporting BLEU and WER on tokenized outputs.

Benchmarks medical AI agents on synthetic EHR tasks, measuring success rates for data retrieval, temporal aggregation, and threshold-based conditional logic with exact-match scoring.

Benchmarks systems on the Abstraction and Reasoning Corpus (ARC) by requiring inference of abstract transformation rules from few input-output grid demonstrations and application to novel test cases, reporting the fraction of tasks solved.

Benchmarks automatic post-editing (APE) models on WMT'18 SMT, SubEdits, and MLQE-PE datasets, reporting BLEU, ChrF, and TER scores computed with SacreBLEU and TERCOM.

Computes the Anderson-Darling test statistic and p-value using scipy.stats.anderson for evaluating predictions against ground truth.

Benchmarks LLM agents fine-tuned with the Agent Data Protocol across software engineering, web browsing, OS/database tool use, and reasoning tasks, reporting unit test pass rates and task success rates.

Evaluates an AI judge system's pairwise ranking accuracy on generated commit messages against a heuristic ground truth from five automatic text metrics, using the MCMD dataset.

Benchmarks large language models on symbolic music understanding and instruction following using text-based ABC notation, covering syntax parsing, error detection, segment-level reasoning, and sequence-level musical analysis.

Configure repository-specific guidance for GitHub Copilot by creating and structuring .github/copilot-instructions.md files.

Query the U.S. Treasury Fiscal Data API for federal financial data including national debt, government spending, revenue, interest rates, exchange rates, and savings bonds. Access 54 datasets and 182 data tables with no API key required.

Evaluates speaker verification models by computing true positive rate at fixed false positive rate thresholds, probing embedding space separation of same-speaker versus different-speaker pairs.

Computes Theil's U (uncertainty coefficient) between predictions and ground truth using the torchmetrics implementation, handling categorical data and NaN strategies.

Benchmarks a text-to-text transformer across GLUE, SuperGLUE, CNN/Daily Mail, SQuAD, and WMT, reporting GLUE average, BLEU, ROUGE-2-F, and Exact Match scores.

Benchmarks inference latency and computational runtime of transformer models and MLX operations across Apple Silicon and NVIDIA GPU backends, with configurable input lengths and batch sizes.

Computes the R2Score metric using torchmetrics, handling single and multi-output predictions with options for adjusted and variance-weighted scores.

Evaluates how well internal model representations (hidden states) predict token-level information importance in summarization tasks, using NDCG@10 and Spearman's rank correlation.

Measures inference latency of binarized, 8-bit, and 32-bit convolutional layers on edge devices to evaluate the efficiency and speedup of the Larq Compute Engine framework compared to standard implementations.

Compute the Kruskal-Wallis H-test using scipy.stats.kruskal for independent samples, returning the H statistic and p-value.

Compute the F1Score metric using torchmetrics when predictions and ground-truth labels are available.

Evaluates the correlation between a zero-cost NAS metric (epsilon) and actual training accuracy across different neural architecture search spaces, testing the metric's ability to rank architectures without training. It probes whether output dispersion from constant weight initializations can serve as a reliable.

Evaluates the faithfulness of abstractive summaries by verifying if factual claims (masked as cloze questions) in the reference summary can be correctly answered using only the generated summary, compared against a gold-standard answer derived from the source context.

Benchmarks mobile GUI agents on multi-step tasks across 20 Android apps, measuring task completion and essential-state navigation with Task Success Rate and Essential State Achieved Rate.

Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard.

Reviews an academic paper using the NeurIPS review form with three reviewer personas, ensemble scoring, and reflection refinement. Extracts text from PDF, runs structured review, and outputs actionable feedback.

Provides reviewer-style audit and deep review for academic papers in LaTeX, Typst, and PDF formats, producing structured issue bundles, pass/fail gates, and revision roadmaps.

Evaluates text-to-image generation models by decomposing assessment into five specialized skills (object presence, count, spatial relations, scale, and text rendering) and open-ended prompts, producing interpretable binary scores with visual and textual explanations.

Evaluates spatial realism and temporal flow consistency of AI-generated videos using embedding spaces and Fourier transforms, producing bounded STREAM-S and STREAM-T scores.

Evaluates machine-generated log summaries without human-written references, using LLM judgment and dense embeddings to score relevance, informativeness, and coherence.

Computes the Recall metric using torchmetrics, including configuration for binary, multiclass, and multilabel tasks.

Computes the LogAUC metric using the torchmetrics implementation for binary, multiclass, or multilabel classification tasks.

Scores generated text for morphosyntactic well-formedness by measuring how closely it adheres to language-specific dependency rules extracted from treebanks.

Benchmarks long-context language models across 20 sub-tasks spanning 3k–200k tokens, covering retrieval, reasoning, summarization, and instruction understanding, with exact-match accuracy as the primary metric.

Computes the InfoLM metric from torchmetrics for evaluating text generation against ground truth, with configurable information measures and sentence-level scoring.

Evaluates the correlation between automatic text generation scores and human quality ratings, including robustness to domain and quality drift, using metrics like Kendall's Tau and Pearson correlation.

Applies a battle-tested bilingual web-design system with typography, responsive rules, and performance patterns for Yuval Avidani's projects.

Create publication-quality static, animated, and interactive plots with fine-grained control over every element, from basic charts to multi-panel figures, with export to PNG, PDF, and SVG.

Pretrains large language models at scale using PyTorch-native torchtitan with 4D parallelism, Float8, and distributed checkpointing.