skillsbench

Reference for COBOL COMP-3 (packed decimal / BCD) numeric storage -- layout on disk, declaring it in the FD, and pitfalls when carrying the decoded value through working storage and into later arithmetic. Useful when an input record holds a balance, rate, or carry-forward as packed decimal rather than a printable digit string.

Reference for the EBCDIC "overpunch" / zoned-decimal sign convention where the units position of a numeric field is replaced with a letter that encodes both a digit and a sign. Useful when reading mainframe-style fixed-length tapes whose amount fields appear as digits followed by a letter (e.g. "0000000000123D" or "0000000000045M").

Background on general-ledger batch posting codes (HD/DR/CR/RV plus signed "other" rows). Reference only -- follow the task instruction for the exact rules, which may differ per account type.

GnuCOBOL toolchain pointers for mainframe-style batch programs (LINE/RECORD SEQUENTIAL files, PIC clauses, cobc build flags).

Reference for paired-reversal handling on GL batch tapes: an RV row whose ref-trace points at an earlier row cancels BOTH legs (the RV and the row it references) — but ONLY when the two value-dates fall within the shop's reversal settlement window; an RV that references a too-old posting is NOT a cancellation and posts one-sided. Useful when the posting tape carries a reference-trace column alongside reversal codes and the running balance is off by twice the reversal magnitude, or by a whole leg on stale references.

Reference for multi-hop / triangulated currency conversion on batch rate tapes. When a rate row carries an explicit "via" currency, the row's rate is only one leg of the conversion, and the via currency may ITSELF be quoted through another via — so the effective rate is the product of every leg, resolved by walking the chain until a direct quote is reached. Useful when an FX feed lists some pairs directly to the reporting currency and others as triangulated (or doubly-triangulated) quotes through vehicle currencies.

Implement paper-defined attention variants by extracting mechanism invariants, mapping tensor shapes, preserving module interfaces, validating numerical behavior, and integrating the custom attention into an existing transformer stack.

Run Python code on cloud GPUs using Modal serverless platform. Use when you need A100/T4/A10G GPU access for training ML models. Covers Modal app setup, GPU selection, data downloading inside functions, and result handling.

PDF manipulation toolkit. Extract text/tables, create PDFs, merge/split, fill forms, for programmatic document processing and analysis.

Query PubChem via PUG-REST API/PubChemPy (110M+ compounds). Search by name/CID/SMILES, retrieve properties, similarity/substructure searches, bioactivity, for cheminformatics.

Cheminformatics toolkit for fine-grained molecular control. SMILES/SDF parsing, descriptors (MW, LogP, TPSA), fingerprints, substructure search, 2D/3D generation, similarity, reactions. For standard workflows with simpler interface, use datamol (wrapper around RDKit). Use rdkit for advanced control, custom sanitization, specialized algorithms.

The gh CLI is GitHub's official command line tool for interacting with GitHub repositories, issues, pull requests, and more. When needs to interact with GitHub repositories, issues, pull requests, and more, use this skill.

Choose placements that preserve useful residual capacity. Use for bin packing, GPU sharing, accelerator placement, and multi-resource scheduling where stranded capacity hurts future fit.

Validate and repair proposed resource allocations by replaying them against temporary capacity. Use when actions consume several resource dimensions such as CPU, memory, GPUs, or accelerators.

Design deterministic online scheduling policies from current observations. Use when assigning arriving work to limited resources without seeing future requests.

Implement mHC (Manifold-Constrained Hyper-Connections) for stabilizing deep network training. Use when implementing residual connection improvements with doubly stochastic matrices via Sinkhorn-Knopp algorithm. Based on DeepSeek's 2025 paper (arXiv:2512.24880).

Train GPT-2 scale models (~124M parameters) efficiently on a single GPU. Covers GPT-124M architecture, tokenized dataset loading (e.g., HuggingFace Hub shards), modern optimizers (Muon, AdamW), mixed precision training, and training loop implementation.

Methodology for clause-by-clause review of a contract against a structured deviation policy ("playbook"). Covers how to walk a playbook, locate the matching provision in the contract, apply rule types (max-value, must-be-present, must-be-absent, acceptable-set, must-have-feature), classify the result (ok / risk / reject), choose the prescribed action, and ground each finding in a verbatim excerpt. Use whenever reviewing any contract — NDA, MSA, vendor DD questionnaire, lease, DPA — against a structured rules-based playbook.

Reference for the standard clauses found in commercial non-disclosure agreements (mutual and one-way) — what each clause does, the surface forms it appears in, and how to recognise it in unfamiliar drafting. Use when reviewing, comparing, or extracting provisions from any confidentiality / NDA / mutual NDA / standstill-and-confidentiality agreement.

Read Microsoft Excel (.xlsx) files robustly with `openpyxl` (or `pandas`). Covers multi-sheet workbooks, header rows, empty cells, merged cells, comma-separated list cells, and converting a sheet to a list-of-dicts the rest of your code can consume. Use when a task input or reference document is an `.xlsx` file rather than JSON/CSV.

Analyze and count objects in videos using Google Gemini API (object counting, pedestrian detection, vehicle tracking, and surveillance video analysis).

Analyze videos with Google Gemini API (summaries, Q&A, transcription with timestamps + visual context, scene/timeline detection, video clipping, FPS control, multi-video comparison, and YouTube URL analysis).

Analyze images and multi-frame sequences using OpenAI GPT series

Extract frames from video files and save them as images using OpenCV

Create audiobooks from web content or text files. Handles content fetching, text processing, and TTS conversion with automatic fallback between ElevenLabs, OpenAI TTS, and gTTS.

ElevenLabs Text-to-Speech API for high-quality speech synthesis.

Google Text-to-Speech (gTTS) for converting text to audio. Use when creating audiobooks, podcasts, or speech synthesis from text. Handles long text by chunking at sentence boundaries and concatenating audio segments with pydub.

OpenAI Text-to-Speech API for high-quality speech synthesis. Use for generating natural-sounding audio from text with customizable voices and tones.

Parse scheduling constraints from a email text.

Manage Gmail - send, read, search emails, manage labels and drafts. Use when user wants to interact with their Gmail account for email operations.

Transcribe audio segments to text using Whisper models. Use larger models (small, base, medium, large-v3) for better accuracy, or faster-whisper for optimized performance. Always align transcription timestamps with diarization segments for accurate speaker-labeled subtitles.

Combine visual features (face detection, lip movement analysis) with audio features to improve speaker diarization accuracy in video files. Use OpenCV for face detection and lip movement tracking, then fuse visual cues with audio-based speaker embeddings. Essential when processing video files with multiple visible speakers or when audio-only diarization needs visual validation.

Choose and implement clustering algorithms for grouping speaker embeddings after VAD and embedding extraction. Compare Hierarchical clustering (auto-tunes speaker count), KMeans (fast, requires known count), and Agglomerative clustering (fixed clusters). Use Hierarchical clustering when speaker count is unknown, KMeans when count is known, and always normalize embeddings before clustering.

Detect speech segments in audio using VAD tools like Silero VAD, SpeechBrain VAD, or WebRTC VAD. Use when preprocessing audio for speaker diarization, filtering silence, or segmenting audio into speech chunks. Choose Silero VAD for short segments, SpeechBrain VAD for general purpose, or WebRTC VAD for lightweight applications.

Build unified multi-level category taxonomy from hierarchical product category paths from any e-commerce companies using embedding-based recursive clustering with intelligent category naming via weighted word frequency analysis.

Video editing with ffmpeg including cutting, trimming, concatenating segments, and re-encoding. Use when working with video files (.mp4, .mkv, .avi) for: removing segments, joining clips, extracting portions, or any video manipulation task.

Process filler word annotations to generate video edit lists. Use when working with timestamp annotations for removing speech disfluencies (um, uh, like, you know) from audio/video content.

Transcribe audio/video to text with word-level timestamps using OpenAI Whisper. Use when you need speech-to-text with accurate timing information for each word.

Transcribe video to timestamped text using Whisper tiny model (pre-installed).