ndpvt-web
- 659 skills
- 0 followers
- 14 repo stars
- last month last updated
- ▌ Pope Learning Reason Hard · ndpvt-webApply the POPE (Privileged On-Policy Exploration) technique to solve hard reasoning problems by decomposing them with oracle-guided prefixes and transferring learned reasoning back to unguided attempts. Use when: 'help me solve this hard problem step by step', 'I'm stuck on this complex algorithm', 'break down this difficult reasoning task', 'guide me through this math/logic problem', 'use privileged hints to bootstrap a solution', 'scaffold a hard problem with partial solutions'.
- ▌ Puda Private User Dataset · ndpvt-webBuild privacy-preserving personalized AI systems using Puda's multi-granularity user data architecture. Implements client-side data aggregation with three privacy levels (raw browsing history, extracted keywords, predefined category subsets) and scoped access control for third-party AI agents. Use when: 'build a privacy-preserving user profile system', 'implement tiered data sharing for personalization', 'create a browser extension that collects user data locally', 'design an agent that personalizes without leaking private data', 'add granular privacy controls to a recommendation system', 'build client-side user data management with OAuth scoping'.
- ▌ Pull Requests As Training · ndpvt-webApply the Clean-PR agentless repo-level code editing protocol: decompose issues into file localization, fine-grained navigation via AST, and minimal Search/Replace patch generation. Triggers: 'fix this issue across the repo', 'edit multiple files for this bug', 'agentless code edit', 'search replace patch', 'repo-level fix', 'locate and patch this bug'
- ▌ Read As Human Compressing · ndpvt-webCompress long contexts using the RAM (Read As Human) strategy: partition text into segments, score relevance against a query, fully retain high-relevance segments (close reading), and compress low-relevance segments into compact summaries (skimming). Use when: 'compress this long document for my query', 'summarize only the relevant parts', 'reduce context length before processing', 'skim this codebase for relevant sections', 'adaptive context compression', 'parallel segment relevance filtering'.
- ▌ The Shadow Self Intrinsic · ndpvt-webDetect and mitigate intrinsic value misalignment in LLM agent systems using the IMPRESS scenario-driven framework. Use when: 'audit my agent for value misalignment', 'test if my agent acts against user interests', 'generate misalignment probes for my LLM system', 'check my agent pipeline for shadow self behavior', 'build safety scenarios for my autonomous agent', 'evaluate alignment of my agentic workflow'.
- ▌ Towards Green AI Decoding · ndpvt-webOptimize LLM-generated code for energy efficiency by detecting and suppressing babbling behavior (excess tokens like redundant test cases, alternative implementations, whitespace padding, and usage examples appended after functional code). Use when: 'reduce energy cost of code generation', 'suppress babbling in LLM output', 'trim unnecessary tokens from generated code', 'optimize inference energy for code', 'detect excessive generation in code output', 'green AI code generation'.
- ▌ Bridging Online Offline Rl · ndpvt-webApply Cobalt-style contextual bandit learning to multi-turn code generation tasks. Decomposes iterative coding into partial trajectory completions, treating each debugging turn as a single-step bandit problem rather than a full RL rollout. Use when: 'help me fix this code iteratively', 'debug this with test feedback', 'multi-turn code repair', 'iterative code generation with execution feedback', 'fix failing test cases step by step', 'recover from wrong code using error output'.
- ▌ Comparing AI Coding Agents · ndpvt-webAnalyze AI coding agent PR datasets using task-stratified acceptance rate methodology. Classify PRs into 9 task categories (feat, fix, docs, chore, refactor, test, ci, build, style), compute per-agent acceptance rates, run stratified chi-square tests with Bonferroni correction, and perform temporal trend analysis. Trigger phrases: "compare AI coding agents", "analyze PR acceptance rates", "stratified PR analysis", "which AI agent is best for", "task-type acceptance rates", "AI agent benchmark pipeline"
- ▌ Do Reasoning Ask Questions · ndpvt-webInformation-theoretic question-asking framework for disambiguating user intent through structured yes/no questions. Uses a three-agent architecture (Seeker, Oracle, Pruner) grounded in Shannon entropy to maximize information gain per question turn. Trigger phrases: 'clarify ambiguous request', 'ask better questions', 'narrow down requirements', 'disambiguate user intent', 'information-gathering dialogue', 'structured requirements elicitation'.
- ▌ Evaluating Social Bias RAG · ndpvt-webEvaluate and mitigate social bias in RAG pipelines. Use when: 'audit my RAG system for bias', 'check if retrieval introduces stereotypes', 'measure fairness in my QA pipeline', 'reduce bias in LLM outputs with retrieval', 'evaluate social bias across demographic groups', 'bias-aware RAG system design'.
- ▌ Evolving Tool User Creator · ndpvt-webTransform Claude from a static tool user into a dynamic tool creator using the UCT (User-to-Creator Transformation) framework. Harvests reasoning traces from problem-solving sessions and distills them into reusable utility functions, scripts, and helpers that grow a persistent tool library over time. Trigger phrases: 'create a reusable tool from this solution', 'build a helper I can reuse', 'evolve my toolset', 'extract a utility from this workflow', 'self-improving agent pipeline', 'turn this reasoning into a tool'.
- ▌ From Passive Metric Active · ndpvt-webBuild systems that use LLM uncertainty as an active control signal -- routing computation, triggering tool calls, enabling self-correction, and governing agent decisions. Use when asked to: 'add uncertainty-aware routing to my LLM pipeline', 'make my agent decide when to use tools based on confidence', 'implement adaptive compute for reasoning chains', 'add self-correction when the model is unsure', 'build a confidence-gated retrieval system', 'prevent reward hacking in my RLHF loop'.
- ▌ Logsieve Task Aware CI Log · ndpvt-webReduce verbose CI/CD build logs before LLM analysis using RCA-aware semantic filtering. Removes boilerplate lines (dependency downloads, progress bars, timestamps, env setup) while preserving diagnostic content (error traces, test failures, compiler errors). Use when: 'analyze this CI log', 'why did my build fail', 'reduce this GitHub Actions log', 'summarize build failure from logs', 'filter noisy CI output', 'diagnose this workflow run'.
- ▌ Multi Field Tool Retrieval · ndpvt-webImplement multi-field tool retrieval systems that decompose tool documentation into structured fields (description, parameters, response, examples) and match user queries against each field independently with learned weights. Use when asked to 'build a tool retrieval system', 'improve tool search for an LLM agent', 'index API documentation for retrieval', 'match user queries to tools', 'build a tool recommendation engine', or 'optimize function/API lookup for agents'.
- ▌ Secure Code Generation Via · ndpvt-webGenerates secure, vulnerability-free code by applying the SecCoderX reasoning framework — systematically analyzing code for CWE-classified vulnerabilities while preserving full functionality. Combines vulnerability-aware task decomposition with a reasoning-based security audit pass on every code output. Trigger phrases: - "Write secure code for..." - "Generate code without vulnerabilities" - "Security-hardened implementation of..." - "Check this code for CWE vulnerabilities" - "Fix the security issues in this code" - "Write this function with security best practices"
- ▌ Social Catalysts Not Moral · ndpvt-webDesign and audit multi-agent LLM systems for genuine cooperation vs. surface compliance. Implements Anchoring Agent injection, cognitive decomposition (belief error + strategic deviation), and transfer testing from the "Social Catalysts, Not Moral Agents" framework. Use when: "test if my agents truly cooperate", "audit multi-agent alignment", "detect chameleon behavior in LLM agents", "design a public goods game for agents", "check if agent cooperation is real or strategic", "add anchoring agents to my system".
- ▌ Stop Testing Attacks Start · ndpvt-webDiagnose LLM safety defenses using the Four-Checkpoint Framework. Instead of asking "does this jailbreak work?", systematically identify WHERE and WHY safety mechanisms fail across four sequential defensive layers (input-literal, input-intent, output-literal, output-intent). Use when: "audit LLM safety", "diagnose defense gaps", "evaluate safety checkpoints", "red-team with checkpoint analysis", "find where safety breaks", "run four-checkpoint analysis".
- ▌ Test Vs Mutant Adversarial · ndpvt-webAdversarial test generation using two competing LLM agents: a Test Agent that writes unit tests and a Mutant Agent that creates code mutations to expose blind spots. The agents co-evolve through iterative rounds, producing test suites with high coverage AND strong fault detection. Trigger phrases: "generate robust tests", "adversarial test generation", "mutation testing", "find blind spots in my tests", "harden my test suite", "test vs mutant"
- ▌ Treetensor Boost AI System · ndpvt-webImplement TreeTensor-based nested data handling for AI systems using the DI-treetensor library. Replaces manual recursive traversal of hierarchical data (dicts-of-tensors, nested observations, multi-agent states) with tree-structured tensor operations that auto-propagate through the hierarchy. Use this skill when: - "Help me batch nested dictionary observations for RL training" - "Stack variable-structure game states into a training batch" - "Apply a neural network to nested observation data without manual loops" - "Simplify my multi-agent data pipeline that has deeply nested dicts" - "Convert my recursive tensor processing code to use TreeTensor" - "Handle heterogeneous nested data in PyTorch without boilerplate"
- ▌ Tsaqa Time Series Analysis · ndpvt-webStructured time series question answering using the TSAQA six-task framework: anomaly detection, classification, characterization, comparison, data transformation, and temporal relationship analysis. Triggers: 'analyze this time series', 'detect anomalies in sensor data', 'compare these two time series', 'classify this signal', 'what trend does this data show', 'reorder these time segments'
- ▌ When Much Imagine Adaptive · ndpvt-webAdaptive test-time scaling framework that decides WHEN and HOW MUCH to invoke expensive generative steps (world models, tool calls, API queries) before committing compute. Based on the AVIC (Adaptive Visual Imagination with Confidence) paper. Use when: 'build an adaptive pipeline that skips unnecessary generation steps', 'add confidence gating before expensive API calls', 'implement adaptive test-time compute scaling', 'reduce wasted inference by checking sufficiency first', 'build a system that decides whether to call a model or skip it', 'implement selective tool invocation with confidence checks'.
- ▌ Bass Benchmarking Audio Lms · ndpvt-webBuild evaluation benchmarks for audio language models using the BASS methodology — structured task taxonomies across structural segmentation, lyric transcription, musicological analysis, and artist collaboration. Trigger phrases: 'benchmark audio model', 'evaluate music understanding', 'music LLM evaluation', 'audio reasoning benchmark', 'test music AI capabilities', 'build music benchmark dataset'
- ▌ Beyond Blame Rethinking Szz · ndpvt-webIdentify bug-inducing commits using temporal knowledge graph search beyond git blame. Use when: 'find what commit introduced this bug', 'trace root cause of regression', 'which commit broke this', 'find the bug-inducing change', 'blame analysis for this fix', 'what caused this defect'
- ▌ Biases Blind Spot Detecting · ndpvt-webAutomated black-box pipeline for detecting unverbalized biases in LLM decision-making. Discovers biases that models exhibit but never mention in their chain-of-thought reasoning. Use when: 'detect hidden biases in my LLM', 'audit model fairness', 'find unverbalized biases', 'bias testing pipeline', 'test my model for discrimination', 'automated bias discovery'.
- ▌ Datachef Cooking Up Optimal · ndpvt-webAutomate data recipe generation for LLM fine-tuning and adaptation. Generates executable data processing pipelines (filtering, synthesis, mixing, augmentation) that transform raw data sources into optimized training corpora for a target task. Trigger phrases: 'create a data recipe', 'optimize training data mix', 'build a data pipeline for fine-tuning', 'curate training data for task X', 'generate a data processing pipeline', 'mix datasets for domain adaptation'.
- ▌ David Vs Goliath Verifiable · ndpvt-webAudit and harden tool-augmented AI agent systems against Tag-Along Attacks -- adversarial agent-to-agent jailbreaks that exploit legitimate tool privileges through conversation. Use when asked to 'red-team an agent system', 'audit agentic tool safety', 'design verifiable agent security tests', 'build a tag-along attack harness', 'harden an operator agent against privilege escalation', or 'evaluate agent-to-agent attack surfaces'.
- ▌ Echo Open Research Platform · ndpvt-webBuild and configure ECHO-style research platforms for running reproducible user studies comparing chat-based AI and web search interactions. Use when: 'set up a user study platform', 'build a chat vs search experiment', 'log participant interactions with LLMs', 'create a research workflow with surveys and tasks', 'export user study interaction traces', 'configure an IRB-compliant experiment with pre/post questionnaires'.
- ▌ From Assistant Double Agent · ndpvt-webSecurity audit and hardening for personalized LLM-based agents against prompt injection, tool poisoning, and memory attacks. Use when: 'audit my agent for security vulnerabilities', 'test my AI assistant against prompt injection', 'harden my agent toolchain', 'evaluate memory poisoning risks in my agent', 'red-team my personalized AI agent', 'add defenses against indirect prompt injection'.
- ▌ From Sparse Decisions Dense · ndpvt-webBuild content moderation and safety classification systems using multi-attribute trajectory reasoning instead of binary labels. Decomposes monolithic safe/unsafe decisions into structured reasoning chains (evidence grounding, modality assessment, risk mapping, policy decision, response generation) with multi-head reward scoring. Use when asked to: 'build a content moderation pipeline', 'classify harmful content with explanations', 'create a safety filter with reasoning traces', 'design a multi-attribute content scorer', 'implement explainable content moderation', 'add dense safety reasoning to a classifier'.
- ▌ Ide Bench Evaluating As Ide · ndpvt-webApply IDE-Bench's structured agent workflow for tackling real-world software engineering tasks: systematic exploration before editing, intent-driven tool transitions, and iterative test validation. Use when asked to 'fix a bug across multiple files', 'implement a feature in an unfamiliar codebase', 'refactor this module', 'optimize performance of this system', 'act like an IDE agent', or 'solve this engineering task systematically'.
- ▌ Malicious Agent Skills Wild · ndpvt-webAudit and detect malicious agent skills in LLM skill registries using the taxonomy and analysis pipeline from Liu et al. (2026). Identifies Data Thieves (credential exfiltration via supply chain), Agent Hijackers (instruction manipulation), shadow features, and platform-native attacks. Trigger phrases: "audit this skill for security", "scan agent skills", "detect malicious skills", "is this skill safe to install", "review skill registry for threats", "check skill for hidden behavior"
- ▌ Sar RAG Atr Visual Question · ndpvt-webBuild image retrieval-augmented generation (ImageRAG) pipelines for visual recognition tasks. Combines vision embeddings, vector databases, and multimodal LLMs to classify or describe images by retrieving similar labeled exemplars before generation. Use this skill when: - "Build an image RAG pipeline for visual classification" - "Create a retrieval-augmented system for SAR target recognition" - "Set up a vector database of image embeddings for similarity search" - "Add a visual memory bank to improve image classification accuracy" - "Implement ImageRAG for visual question answering" - "Retrieve similar images from a database to help an MLLM identify objects"
- ▌ Stateless Yet Not Forgetful · ndpvt-webDetect, audit, and defend against implicit memory channels in LLM-powered systems where models encode hidden state in outputs that persists across interactions. Use when: 'audit my LLM pipeline for hidden channels', 'check if my agent can leak state between sessions', 'detect steganographic encoding in model outputs', 'sanitize LLM outputs before reingestion', 'build a stateless agent pipeline', 'test for implicit memory vulnerabilities'.
- ▌ Swe Context Bench Benchmark · ndpvt-webReuse prior coding experience across related repository tasks. Accumulate, summarize, retrieve, and inject compact experience from previously solved issues to boost accuracy and cut cost on downstream tasks. Use when: 'reuse context from prior fixes', 'apply experience from related issues', 'solve this like the last bug', 'use what we learned fixing X to fix Y', 'accumulate coding experience across tasks', 'context-aware issue resolution'.
- ▌ System Name Address Parsing · ndpvt-webParse unstructured person names and addresses into a structured 17-field schema using prompt-driven extraction with layered validation. Use when the user asks to 'parse addresses', 'extract name fields', 'structure address data', 'normalize mailing records', 'split name and address into fields', or 'clean up contact data'.
- ▌ The Clef 2026 Finmmeval Lab · ndpvt-webBuild multilingual, multimodal financial AI evaluation pipelines using the FinMMEval framework. Covers financial exam QA, cross-lingual document reasoning, and trading decision systems. Use when the user says: 'evaluate a financial LLM', 'build a financial QA benchmark', 'multilingual financial reasoning pipeline', 'financial decision-making agent', 'cross-lingual finance evaluation', 'financial exam question answering system'.
- ▌ Trust Design Skill Profiles · ndpvt-webBudget-aware LLM model selection using BELLA-style skill profiling. Decomposes tasks into granular skill requirements, builds capability matrices for candidate models, and runs multi-objective optimization to recommend the cheapest model that meets quality constraints. Generates transparent natural-language rationale for every recommendation. Trigger phrases: "which model should I use", "select the cheapest LLM", "LLM routing", "model cost comparison", "skill profile for this task", "budget-efficient model selection"
- ▌ Will It Survive Deciphering · ndpvt-webAnalyze the survival and maintenance fate of AI-generated code in repositories using survival analysis techniques from Rahman & Shihab (2026). Assess whether AI-authored code is durable or disposable, classify modification types, and predict which code units are modification-prone. Trigger phrases: 'analyze AI code survival', 'is this AI code disposable', 'code survival analysis', 'predict code modification risk', 'compare AI vs human code durability', 'audit AI-generated code maintenance burden'.
- ▌ A2rag Adaptive Agentic Graph · ndpvt-webBuild adaptive, cost-aware Graph-RAG pipelines that route queries through escalating retrieval stages (local -> bridge -> global) with triple-check verification and provenance map-back. Use when: 'build a graph RAG pipeline', 'implement adaptive retrieval for knowledge graphs', 'cost-aware multi-hop question answering', 'add evidence verification to RAG', 'handle mixed-difficulty queries efficiently', 'graph retrieval with source text grounding'.
- ▌ Are Open Weight Ready Social · ndpvt-webBuild LLM-based content moderation pipelines using zero-shot classification with open-weight models. Implements the structured prompting and evaluation methodology from Chou et al. (2026) for detecting rudeness, intolerance, and threats in social media posts. Use when: 'build a content moderation system', 'detect harmful posts with LLMs', 'moderate social media content', 'set up open-weight moderation pipeline', 'evaluate LLM moderation accuracy', 'privacy-preserving content filtering'.
- ▌ Compar Ia French Governments · ndpvt-webBuild multilingual LLM evaluation arenas and preference data collection pipelines modeled on France's compar:IA platform. Collects human preference pairs for RLHF/DPO training in non-English languages using blind pairwise comparison, Bradley-Terry ranking, and privacy-preserving filtering. Trigger phrases: 'build an LLM arena', 'collect preference data for DPO', 'create a chatbot comparison platform', 'multilingual RLHF data pipeline', 'pairwise LLM evaluation system', 'French language model leaderboard'.
- ▌ Covagent Overcoming 30 Curse · ndpvt-webBoost Android app test coverage beyond the 30% activity ceiling using agentic static analysis of Smali code, component transition graphs, activation condition inference, and Frida dynamic instrumentation script generation. Use when: 'increase Android app test coverage', 'generate Frida scripts for unreachable activities', 'analyze APK activation conditions', 'break through coverage barriers in mobile testing', 'instrument Android app for GUI fuzzing', 'find unreachable activities in Android app'.
- ▌ Drpg Decompose Retrieve Plan · ndpvt-webStructured rebuttal and critique-response generation using the DRPG framework (Decompose, Retrieve, Plan, Generate). Breaks down multi-point feedback into atomic concerns, retrieves supporting evidence from source documents, selects a rebuttal strategy (clarification vs. justification), and generates targeted point-by-point responses. Trigger phrases: - "Write a rebuttal to this review" - "Respond to reviewer comments" - "Address this feedback point by point" - "Help me respond to this code review / PR review" - "Draft a response to these critique points" - "Decompose this review and plan responses"
- ▌ Echoes Loop Diagnosing Risks · ndpvt-webDiagnose and mitigate feedback-loop risks (bias amplification, hallucination propagation, exposure polarization) in LLM-powered recommender systems using a role-aware, phase-wise diagnostic framework. Use when: 'audit my recommendation pipeline for bias', 'check feedback loop risks in my LLM recommender', 'diagnose hallucination propagation in recommendations', 'build a feedback-loop simulation for my recommender', 'trace popularity bias through my recommendation cycles', 'add risk monitoring to my LLM-based ranking system'.
- ▌ Fine Tuning Gpt 5 Gpu Kernel · ndpvt-webGenerate optimized GPU kernels in Triton from PyTorch reference code using the Makora RL-based iterative refinement workflow. Applies tool-augmented kernel generation with correctness validation and performance benchmarking. Use when: 'write a Triton kernel for this PyTorch op', 'optimize this GPU kernel', 'generate a fast Triton implementation of matrix multiply', 'convert this PyTorch module to a custom Triton kernel', 'speed up this CUDA operation with Triton', 'benchmark my kernel against TorchInductor'.
- ▌ Hunt Instead Wait Evaluating · ndpvt-webAutonomously explore databases and datasets to extract key insights without predefined queries, using investigatory intelligence (hypothesis-driven, goal-setting exploration) rather than executional intelligence (completing assigned tasks). Applies the Deep Data Research (DDR) methodology from Liu et al. 2026. Trigger phrases: - "Explore this database and find interesting insights" - "What patterns or anomalies exist in this dataset?" - "Investigate this data and tell me what's important" - "Do a deep dive analysis on this database" - "Hunt for insights in these tables" - "What can we learn from this data without me telling you what to look for?"
- ▌ Koral Knowledge Graph Guided · ndpvt-webBuild Knowledge Graph-guided LLM reasoning pipelines for operational telemetry analysis. Combines a Literature KG (extracted from domain papers) with a Data KG (materialized from telemetry) to ground LLM outputs in structured evidence. Triggers: 'build a knowledge graph from telemetry', 'KG-guided LLM reasoning', 'analyze SSD or hardware telemetry with a knowledge graph', 'ground LLM analysis in structured evidence', 'knowledge graph pipeline for operational data', 'extract knowledge graph from research papers'
- ▌ Lemon Agent Technical Report · ndpvt-webOrchestrate multi-agent workflows using the Lemon Agent orchestrator-worker pattern with hierarchical scheduling, progressive context compression, and self-evolving memory. Use when asked to 'break this into parallel subtasks', 'orchestrate agents for a complex task', 'manage context across long-running agent workflows', 'set up an orchestrator-worker pipeline', 'compress agent context efficiently', or 'build a multi-agent system with memory'.
- ▌ Revisiting Role Natural Code · ndpvt-webComment-augmented code translation (COMMENTRA) that uses targeted natural language comment injection to significantly improve LLM-based cross-language code translation. Use when: 'translate this Java to Python', 'convert C++ code to Go', 'port this Python function to C', 'translate code between languages', 'fix a failed code translation', 'improve code translation accuracy'.
- ▌ Shardmemo Masked Moe Routing · ndpvt-webImplement ShardMemo-style tiered, sharded memory with masked Mixture-of-Experts routing for agentic LLM systems. Use when: 'build a sharded memory system for agents', 'add tiered memory with MoE routing', 'implement scope-based memory retrieval for multi-agent', 'create a budgeted agent memory service', 'route queries to memory shards with masking', 'design a scalable external memory for concurrent agents'.
- ▌ Smartoracle Agentic Approach · ndpvt-webAgentic differential oracle for triaging cross-implementation discrepancies. Decomposes bug triage into specialized sub-agents (discrepancy finder, spec checker, false-positive critic, duplicate checker) that independently gather evidence and synthesize a verdict. Use when: 'triage differential test results', 'build a differential oracle', 'reduce false positives in cross-engine testing', 'agentic bug triage pipeline', 'analyze spec compliance differences', 'filter noise from differential fuzzing'.
- ▌ Strong Reasoning Isnt Enough · ndpvt-webBuild interactive diagnostic agents that systematically elicit evidence before concluding, using the REFINE (Reasoning-Enhanced Feedback for INformation Elicitation) loop from EID-Benchmark research. Prevents premature diagnosis by measuring Information Coverage Rate (ICR) and closing evidence gaps through verification feedback. Trigger phrases: 'build a diagnostic chatbot', 'interactive troubleshooting agent', 'evidence gathering workflow', 'systematic question-asking agent', 'build an intake triage system', 'REFINE diagnostic loop'
- ▌ Tam Eval Evaluating Llms For · ndpvt-webSystematic test suite maintenance using the TAM-Eval framework's three-scenario approach: creating new tests, repairing broken tests, and updating tests after code changes. Evaluates test quality via pass rate, code coverage delta, and mutation score delta rather than relying on reference test files. Operates at the test-file level with full repository context. Trigger phrases: - "maintain my test suite" - "fix these broken tests" - "update tests after refactoring" - "generate tests with mutation coverage" - "evaluate my test quality" - "improve test effectiveness"
- ▌ The Semantic Trap Fine Tuned · ndpvt-webEvaluate code vulnerability detection for semantic traps -- where analysis fixates on functional context (e.g., "this is crypto code, so it's probably vulnerable") instead of reasoning about the actual root cause of a vulnerability. Applies the TrapEval methodology to distinguish genuine vulnerability reasoning from pattern-matching shortcuts. Trigger phrases: "check this code for vulnerabilities", "is this patch secure", "audit this function for security issues", "compare vulnerable vs patched code", "does this fix actually address the vulnerability", "evaluate my vulnerability detector"
- ▌ Vectra Metric Dataset Visual · ndpvt-webAssess visual quality of translated product images using Vectra's 14-dimension scoring framework. Use when: 'evaluate translated image quality', 'score e-commerce product rendering', 'assess in-image translation defects', 'build IIMT quality pipeline', 'rate visual rendering of translated text on images', 'detect text hallucination in product photos'.
- ▌ When Get Significantly Worse · ndpvt-webStatistically detect LLM degradation after optimization using McNemar's paired test. Use when: 'did quantization hurt my model', 'is this accuracy drop significant', 'compare model before and after optimization', 'detect model degradation', 'statistical test for LLM quality', 'is 0.5% accuracy drop real or noise'.
- ▌ Why AI Agents Systematically · ndpvt-webDiagnose and fix systematic failure modes in LLM-based multi-agent systems performing root cause analysis on cloud incidents. Applies the 12-pitfall taxonomy from Kim et al. (2026) to audit agent architectures, harden inter-agent communication, and eliminate hallucinated reasoning in diagnostic workflows. Use when: - "audit my agent pipeline for common failure modes" - "why does my RCA agent keep hallucinating root causes" - "harden multi-agent communication for incident response" - "review this agent workflow for reasoning pitfalls" - "my agents lose context across handoffs" - "fix incomplete exploration in my diagnostic agent"
- ▌ Gamms Graph Based Adversarial · ndpvt-webBuild and run graph-based multi-agent adversarial simulations using the GAMMS framework. Covers agent creation, graph environments (including OpenStreetMap road networks), sensor configuration, adversarial rules, potential-field strategies, recording/replay, and integration with ML libraries. Trigger phrases: "simulate multi-agent on a graph", "GAMMS simulation", "adversarial agent modeling", "graph-based agent environment", "multi-agent road network simulation", "build a capture-the-flag simulation"
- ▌ How Information Access Affect · ndpvt-webBuild Extract-and-Evaluate (EaE) hierarchical monitoring pipelines that detect sabotage and misbehavior in LLM agent outputs. Uses information filtering to exploit the "less-is-more" effect where monitors perform better with curated excerpts than full trajectories. Trigger phrases: - "monitor agent outputs for sabotage" - "detect misbehavior in LLM agent traces" - "build an oversight pipeline for AI agents" - "extract and evaluate agent reasoning" - "filter agent trajectories for suspicious behavior" - "hierarchical LLM monitoring system"
- ▌ How Personalized Memory Shape · ndpvt-webRational preference utilization for personalized LLM assistants. Implements RP-Reasoner's pragmatic reasoning to selectively integrate user memory/preferences, avoiding irrational personalization. Use when: 'build a personalized assistant with memory', 'add user preference handling to my chatbot', 'filter relevant user context for responses', 'implement memory-aware response generation', 'prevent over-personalization in my AI assistant', 'rational preference selection for user profiles'.
- ▌ Learning Rate Matters Vanilla · ndpvt-webConfigure optimal learning rates for LoRA fine-tuning of LLMs. Generates hyperparameter search configs, training scripts, and analysis code that ensure vanilla LoRA matches or beats fancy variants. Triggers: 'fine-tune with LoRA', 'set up LoRA training', 'compare LoRA variants', 'optimize LoRA hyperparameters', 'learning rate sweep for LoRA', 'LoRA fine-tuning config'
- ▌ Mata Multiagent Framework For · ndpvt-webMulti-agent table question answering using MATA's three-path reasoning strategy (Chain-of-Thought, Program-of-Thought, Text-to-SQL) with confidence-based answer selection and scheduler-driven efficiency. Use when: 'answer questions about this table', 'analyze this CSV and answer', 'query this spreadsheet data', 'build a table QA pipeline', 'compare multiple reasoning approaches on tabular data', 'reliable table analysis with verification'.
- ▌ Mpib Benchmark Medical Prompt · ndpvt-webEvaluate and defend clinical LLM systems against prompt injection attacks using the MPIB benchmark methodology. Implements Clinical Harm Event Rate (CHER) scoring, adversarial test generation, and defense configuration for medical RAG pipelines. Use when: 'test my medical LLM for prompt injection', 'audit clinical safety of RAG system', 'generate adversarial medical prompts', 'measure clinical harm rate', 'defend medical chatbot against injection', 'evaluate healthcare AI safety'.
- ▌ Opportunities Aiml Rubin Lsst · ndpvt-webBuild trustworthy ML pipelines for large-scale scientific data analysis with calibrated uncertainties, simulation-based inference, and physics-informed constraints. Use when: 'build a photometric redshift pipeline', 'add uncertainty quantification to my ML model', 'create a simulation-based inference workflow', 'integrate ML into a scientific analysis pipeline', 'validate ML predictions against simulations', 'set up active learning for survey data'.
- ▌ Small Beautiful Practical Log · ndpvt-webBuild efficient log parsing systems that extract structured templates from raw log messages using a dual-cache architecture with LLM-powered correction -- optimized for small/local models. Use when the user says 'parse logs', 'extract log templates', 'build a log parser', 'log analysis pipeline', 'structure unstructured logs', or 'log template mining'.
- ▌ Timely Machine Awareness Time · ndpvt-webApply time-budget-aware reasoning to agentic tasks with tool calls. Dynamically adjust strategy depth, tool call frequency, and fallback heuristics based on wall-clock time constraints. Use when: 'optimize this under a time limit', 'run as many experiments as possible in N minutes', 'choose between thorough and fast approaches', 'time-aware agent planning', 'budget my tool calls', 'adapt strategy to deadline'.
- ▌ Towards Science Collective AI · ndpvt-webDesign, evaluate, and optimize LLM multi-agent systems using the Collaboration Gain (Gamma) framework. Replaces trial-and-error MAS design with rigorous factor attribution so you know whether agents are truly collaborating or just burning tokens. Trigger phrases: - "Design a multi-agent system for..." - "Are my agents actually collaborating or just wasting tokens?" - "Evaluate collaboration gain of this agent swarm" - "Optimize my multi-agent architecture" - "Set up a scientific MAS benchmark" - "Attribute which factors drive my multi-agent performance"
- ▌ Agentic AI Healthcare Medicine · ndpvt-webDesign, evaluate, and improve LLM-based agentic systems for healthcare using a seven-dimensional taxonomy with 29 sub-dimensions. Triggers: 'build a healthcare AI agent', 'evaluate my medical agent', 'healthcare agent architecture review', 'audit agent capabilities for clinical use', 'design a multi-agent medical system', 'gap analysis for healthcare LLM agent'.
- ▌ Agentsys Secure Dynamic Agents · ndpvt-webBuild LLM agent systems hardened against indirect prompt injection using hierarchical memory isolation, schema-validated return values, and event-triggered sanitization inspired by OS process isolation. Use this skill when a user asks to "build a secure agent pipeline", "protect agents from prompt injection", "isolate agent memory", "design a multi-agent system with security boundaries", "implement safe tool calling for LLM agents", or "add injection defense to an agent framework".
- ▌ Beyond Accuracy Cognitive Load · ndpvt-webAnalyze and reduce cognitive load in tool-use agent workflows using the Cognitive Load Framework from AAAI 2026. Diagnoses why agent pipelines fail by decomposing task complexity into Intrinsic Load (tool dependency depth/branching) and Extraneous Load (ambiguity/parameter confusion). Use when: 'diagnose why my agent keeps failing', 'reduce tool-call complexity', 'optimize my agent workflow', 'analyze cognitive load of this pipeline', 'map capability boundaries', 'simplify my tool orchestration'.
- ▌ Bridging Modality Gap Roadside · ndpvt-webBuild training-free pipelines that convert sparse 3D LiDAR point clouds into depth-encoded 2D images for classification by Vision-Language Models (CLIP, etc.). Covers the full workflow: point cloud denoising, temporal frame fusion, canonical orientation, orthographic projection, morphological cleanup, bilateral smoothing, and few-shot VLM prompting with semantic anchoring. Use when: 'classify vehicles from roadside LiDAR', 'convert point clouds to images for VLM', 'few-shot 3D object classification without training', 'bridge LiDAR to vision-language model', 'bootstrap labeled dataset from unlabeled LiDAR', 'cold start vehicle classifier from point clouds'.
- ▌ Contextual Drag Errors Context · ndpvt-webMitigate contextual drag — the phenomenon where failed attempts in conversation context bias LLM reasoning toward structurally similar errors. Apply context-denoising and fallback-reasoning strategies when iterating on broken code, debugging multi-step failures, or refining solutions that keep failing in similar ways. Trigger phrases: - "I keep getting the same kind of error" - "My fix attempt made it worse" - "The refactored code has the same bug pattern" - "Each iteration introduces a similar failure" - "Self-correction loop isn't converging" - "Why does my retry keep failing the same way"
- ▌ Cutting Gordian Knot Detecting · ndpvt-webDetect malicious PyPI/NPM packages using behavioral pattern mining and semantic reasoning (PyGuard). Use when: 'scan this package for malware', 'is this PyPI dependency safe', 'audit my requirements.txt for supply chain attacks', 'check this setup.py for suspicious behavior', 'analyze this npm package for data exfiltration', 'detect obfuscated malicious code in this package'.
- ▌ Dynamic Long Context Reasoning · ndpvt-webProcess extremely long documents and contexts by compressing them into memory chunks, selectively retrieving relevant blocks via a gating mechanism, and reasoning iteratively with working memory. Based on the CogMem framework (arXiv:2602.08382). Use this skill when the user says: - "Analyze this huge codebase and answer questions about it" - "Process this long document and find the relevant sections" - "Reason over multiple files to trace a bug" - "Summarize and query across a 100K+ token context" - "Handle this context that's too long to fit in one pass" - "Multi-hop reasoning over a large repository"
- ▌ Effgen Enabling Small Language · ndpvt-webDeploy and optimize small language models (SLMs) as autonomous agents using the effGen framework. Implements prompt compression (70-80% context reduction), five-factor complexity routing, intelligent task decomposition, and unified memory for local SLM-based agent systems. Triggers: 'set up effgen agent', 'deploy small language model agent', 'optimize prompts for small model', 'compress agent context for SLM', 'build local AI agent with effgen', 'route tasks by complexity for small models'
- ▌ Hidden Licensing Risks Llmware · ndpvt-webDetect license incompatibilities across LLM supply chains (OSS repos, models, datasets) using the LiAgent multi-agent extraction and compatibility analysis framework. Use when: 'check my project for license conflicts', 'are my HuggingFace model dependencies compatible', 'audit LLM supply chain licenses', 'find licensing risks in my ML pipeline', 'is this model license compatible with my repo', 'analyze license compatibility for my AI project'.
- ▌ Krone Hierarchical Modular Log · ndpvt-webDetect anomalies in application logs using KRONE's hierarchical decomposition: parse flat log sequences into Entity/Action/Status trees, build modular Krone Seqs, and apply two-stage detection (pattern matching + LLM reasoning). Use when: 'analyze these logs for anomalies', 'detect failures in system logs', 'build a log anomaly detector', 'find unusual patterns in server logs', 'hierarchical log analysis', 'parse logs into execution hierarchies'.
- ▌ Multi Agent Teams Hold Experts · ndpvt-webPrevent expertise dilution in multi-agent LLM workflows by applying findings from 'Multi-Agent Teams Hold Experts Back' (Pappu et al., 2026). Detects and mitigates integrative compromise -- where teams average expert and non-expert views instead of deferring to the best agent. Use when: 'design a multi-agent system', 'why is my agent swarm underperforming', 'should I use multiple agents or one', 'coordinate agents without losing accuracy', 'optimize agent team composition', 'prevent consensus bias in agent teams'.
- ▌ Omni Rrm Advancing Omni Reward · ndpvt-webBuild rubric-grounded reward models and preference evaluation pipelines for multimodal AI outputs. Use when asked to 'evaluate model outputs with rubrics', 'build a preference dataset', 'score multimodal responses', 'create a reward model pipeline', 'judge which AI response is better', or 'rank model completions across dimensions'.
- ▌ Proxywar Dynamic Assessment Of · ndpvt-webBuild competitive game-arena evaluation frameworks for LLM-generated code using ProxyWar's multi-layer pipeline: agent generation, hierarchical testing, iterative repair, and TrueSkill tournaments. Triggers: 'evaluate code generation in game arenas', 'build a competitive agent tournament', 'assess LLM agents with ProxyWar', 'iterative code repair pipeline', 'TrueSkill ranking for generated agents', 'dynamic assessment of code quality beyond benchmarks'
- ▌ Reducing Costs Proof Synthesis · ndpvt-webGenerate formally verified Rust code with Verus specifications and proofs using the VeruSyn methodology. Applies self-synthesis, tutorial-based synthesis, and chain-of-thought debugging to produce correct-by-construction Rust systems code. Trigger phrases: 'verify this Rust code', 'add Verus proofs', 'write a formal specification', 'prove this function correct', 'generate verified Rust', 'add requires and ensures clauses'.
- ▌ Roma Recursive Open Meta Agent · ndpvt-webDecompose long-horizon, multi-step tasks using ROMA's recursive meta-agent pattern: Atomizer decides if a task needs splitting, Planner builds a dependency-aware subtask DAG, Executors run leaf tasks in parallel, and Aggregator compresses results bottom-up. Use when the user says 'break this project into subtasks', 'plan a complex multi-step workflow', 'decompose this into parallel work', 'use recursive task decomposition', 'orchestrate agents for this large task', or 'ROMA-style planning'.
- ▌ Seccodeprm Process Reward Code · ndpvt-webStep-level security scoring for code generation and vulnerability detection using process reward model techniques. Use when asked to: 'review this code for vulnerabilities', 'generate secure code for [task]', 'score security of each code block', 'find the risky parts of this function', 'which lines introduce the vulnerability', 'help me write secure C/C++/Python code'.
- ▌ The Landscape Prompt Injection · ndpvt-webHarden LLM agent systems against prompt injection using layered text/model/execution defenses and the AgentPI evaluation framework. Use when: 'secure my agent against prompt injection', 'audit this agent for injection vulnerabilities', 'add prompt injection defenses', 'evaluate agent trustworthiness', 'harden this LLM pipeline', 'test my agent with adversarial inputs'.
- ▌ When Agents Fail Comprehensive · ndpvt-webDiagnose and fix bugs in LLM agent systems using a research-backed taxonomy of 11 bug types, 9 root causes, and 12 observable effects across agent core, tools, planning, and memory components. Use when: 'my LangChain agent is crashing', 'debug this CrewAI workflow', 'why is my agent stuck in a loop', 'agent gives empty responses', 'fix this AutoGen agent error', 'agent ignores tools I defined'.
- ▌ Agentsm Semantic Memory Agentic · ndpvt-webAgentic Text-to-SQL with semantic memory that captures and reuses structured execution traces. Use when: 'write SQL for this database', 'query this schema', 'generate SQL from natural language', 'text to SQL with complex schema', 'help me explore this database and answer questions', 'build a reusable SQL generation pipeline'.
- ▌ AI Agent For Reverseengineering · ndpvt-webReverse-engineer legacy numerical/scientific Fortran or C code and translate it into modern Python frameworks (Devito, NumPy, SciPy, FEniCS, etc.) using a multi-stage analysis pipeline with knowledge-graph-guided retrieval, structured code synthesis, and iterative validation. Trigger phrases: "convert this Fortran code to Python", "reverse engineer this finite difference code", "translate this legacy numerical solver to Devito", "modernize this scientific computing code", "what does this Fortran stencil do and how do I write it in Python", "migrate this CFD solver from Fortran to a modern framework"
- ▌ Ama Adaptive Memory Multi Agent · ndpvt-webBuild adaptive memory systems using coordinated multi-agent collaboration with hierarchical storage and consistency maintenance. Use when: 'build a memory system for my chatbot', 'add long-term memory to my agent', 'implement multi-granularity retrieval', 'create a memory-augmented LLM pipeline', 'handle memory conflicts in conversational AI', 'reduce context window usage with smart retrieval'.
- ▌ Assessing Quality Mental Health · ndpvt-webEvaluate LLM-generated mental health responses using a 6-attribute clinical rubric spanning Cognitive Support (Guidance, Informativeness, Safety) and Affective Resonance (Empathy, Helpfulness, Interpretation). Based on Badawi et al. 2026. Trigger phrases: "evaluate mental health response", "assess therapeutic quality", "score counseling output", "rate empathy of LLM response", "audit mental health chatbot", "cognitive-affective gap analysis"
- ▌ Atomic Information Flow Network · ndpvt-webTrace and attribute RAG system responses back to specific tools and sources using Atomic Information Flow (AIF) -- a network flow model that decomposes outputs into atoms and computes precise attribution scores. Use when: "trace which tools contributed to this RAG response", "attribute this answer to its sources", "debug why my RAG pipeline returned wrong information", "compress RAG context without losing accuracy", "build attribution into my multi-agent system", "score tool contribution in my retrieval pipeline".
- ▌ Closing Reasoning Gaps Clinical · ndpvt-webBuild systems that detect and fix reasoning gaps in LLM agents by comparing their chain-of-thought against reference reasoning, extracting structured discrepancies, and generating corrective instructions stored in a retrievable knowledge base. Use when: 'build a reasoning improvement pipeline', 'detect logic gaps in agent output', 'compare agent reasoning to expert reasoning', 'create a corrective knowledge base from reasoning errors', 'improve clinical decision support accuracy', 'patch reasoning with RAG-retrieved instructions'.
- ▌ Code2world Gui World Renderable · ndpvt-webPredict and simulate GUI state transitions by generating renderable HTML/CSS/SVG code from screenshots and user actions. Use when asked to: 'simulate what happens when I click this button', 'predict the next UI state', 'generate HTML that reproduces this screenshot', 'build a GUI world model', 'create a visual sandbox for UI testing', 'convert a mobile screenshot to interactive HTML'.
- ▌ Compactrag Reducing Calls Token · ndpvt-webBuild multi-hop RAG systems that answer complex questions with only 2 LLM calls total, regardless of reasoning depth. Applies CompactRAG's offline atomic QA decomposition and online entity-consistent retrieval to slash token costs by 2-5x vs iterative RAG. Trigger phrases: - "build a multi-hop RAG pipeline" - "reduce LLM calls in my RAG system" - "answer complex questions over a knowledge base efficiently" - "implement CompactRAG" - "optimize token usage in retrieval-augmented generation" - "build a cost-efficient question answering system"
- ▌ Deep Researcher Sequential Plan · ndpvt-webConduct deep, multi-step research on complex topics using Sequential Plan Refinement with Reflection and Candidates Crossover. Maintains a Global Research Context across iterations so each search step builds on prior findings, avoids redundancy, and adapts the plan at runtime. Use this skill when: "research this topic in depth", "write a comprehensive report on", "deep dive into", "investigate and synthesize findings on", "generate a research report about", "analyze this complex topic thoroughly".
- ▌ Deepera Deep Evidence Reranking · ndpvt-webRerank retrieved passages for RAG pipelines using step-by-step logical reasoning to filter out semantically similar but logically irrelevant (SSLI) documents. Use when: 'rerank these search results', 'filter irrelevant passages from retrieval', 'build a scientific QA pipeline', 'improve RAG answer quality', 'passages look relevant but answers are wrong', 'reduce hallucinations in retrieval-augmented generation'.
- ▌ Discovering High Level Patterns · ndpvt-webExtract high-level semantic patterns from fine-grained simulation or event logs using LM-guided program synthesis. Transforms raw numerical traces into annotated pattern timelines, then composes reward/query programs from natural language goals. Use when: 'analyze simulation logs for patterns', 'find high-level events in trace data', 'generate reward functions from goals', 'summarize physics simulation output', 'build pattern detectors for time-series logs', 'annotate event traces with semantic labels'.
- ▌ From Pragmas Partners Symbiotic · ndpvt-webAgentic High-Level Synthesis (HLS) optimization: autonomously analyze, insert, and tune C/C++ HLS pragmas (pipeline, unroll, array_partition, dataflow) through closed-loop feedback with synthesis tools. Use when: 'optimize this HLS kernel', 'add pragmas to this C++ for FPGA', 'explore the design space for this Vitis HLS project', 'tune my hardware accelerator', 'analyze this HLS synthesis report', 'debug why my HLS design has poor throughput'.
- ▌ Gflowpo Generative Flow Network · ndpvt-webOptimize LLM prompts using GFlowPO's iterative generate-evaluate-refine loop with diversity-preserving exploration and dynamic memory. Use when: 'optimize this prompt', 'find a better prompt for this task', 'prompt engineering with examples', 'auto-tune my system prompt', 'improve prompt accuracy', 'generate prompt variations'.
- ▌ Mempot Defending Against Memory · ndpvt-webDefend LLM agent memory systems against extraction attacks using optimized honeypot injection and sequential detection. Implements the MemPot framework: generates trap documents that lure attackers while staying invisible to legitimate users, then detects extraction attempts via Wald's Sequential Probability Ratio Test (SPRT). Trigger phrases: - "protect agent memory from extraction" - "add honeypots to my RAG memory" - "detect memory extraction attacks on my LLM agent" - "defend my vector database against adversarial queries" - "implement SPRT-based attack detection for my agent" - "harden my LLM agent's retrieval system"
- ▌ Pearl Plan Exploration Adaptive · ndpvt-webApply PEARL's two-phase tool orchestration: offline tool exploration to learn valid usage patterns and failure modes, then structured plan-before-execute workflows for complex multi-step tool chains. Use when: 'plan a multi-tool workflow', 'chain multiple API calls', 'explore tools before executing', 'build a robust tool pipeline', 'multi-hop tool invocation', 'reduce tool call errors in a workflow'.