PROMPT LIBRARY

AI & ML: system prompts for agents.

Role prompts and system prompts maintained in awesome-prompts, grouped by field. Copy one into Claude Code, Codex or any chat model.

24 AI & ML prompts

Agent Environment Engineering Architect EN

Design the runtime, artifacts, constraints, and interfaces that let off-the-shelf CLI agents do metric-driven autonomous scientific discovery: permissions/artifact/budget/human-in-the-loop engineering, hidden-evaluator sandbox, parallel propose-implement loops, cost-capped exploration; based on EurekAgent (arXiv 2606.13662, June 2026; THU-Team-Eureka/EurekAgent)

Agent World Model Architect EN

Predictive environment simulators for agent imagination: state-space design, dynamics modeling, counterfactual rollouts, plan-then-execute integration, world-model-specific safety (hallucinated futures, goal misgeneralization, deceptive alignment); spans physics, language, and hybrid world models; based on VLA-World, OccuBench, and 2026 world-model safety research (2026)

AgentAtlas Trajectory Eval Architect EN

Diagnostic agent evaluator: scores trajectories by control-decision taxonomy (Act / Ask / Refuse / Stop / Confirm / Recover), trajectory-failure taxonomy, six-axis coverage audit, and taxonomy-aware vs. taxonomy-blind gap; separates real capability from prompt-supervision artifacts; based on "AgentAtlas: Beyond Outcome Leaderboards for LLM Agents" (arXiv 2605.20530, May 2026)

Agentic Company Orchestrator EN

Zero-human-company multi-agent orchestration architect: org-chart design, heartbeat-driven execution, goal-aligned delegation, budget governance with hard stops, ticket-based task tracking, board approval gates, multi-company isolation, and portable company templates; based on paperclipai/paperclip (Mar 2026, 64k+ stars)

AI Ethics Reviewer EN

Algorithmic ethics audit: fairness & bias, transparency, privacy, safety, accountability, societal impact, cross-cultural considerations, mitigation roadmap (2026)

Autonomous ML Research Agent EN

Self-directed experiment loop for ML research: fixed-time-budget training, single-file edit discipline, keep/discard decision gates, git-branch state management, overnight autonomy; reads code, forms hypotheses, runs experiments, logs results, and iterates without human intervention; based on karpathy/autoresearch (Mar 2026, 80k+ stars)

Embodied AI Developer EN

VLA systems, robotic agents, world-model-driven embodied intelligence: perception-action grounding, sim-to-real pipelines, cross-embodiment transfer, skill primitives, physical safety gates; derived from 2026 embodied-AI research (StarVLA, EmbodiedClaw, VLA-World) (2026)

LLM Architect EN

LLM systems: fine-tuning (LoRA/QLoRA/RLHF/DPO), RAG architecture, serving (vLLM/TGI), quantization (GPTQ/AWQ), safety guardrails, multi-model orchestration (2026)

Long-Horizon Multimodal Search Agent EN

Sustained visual-textual search across 100-turn horizons: file-based visual context management, progressive on-demand image loading, multi-hop visual reasoning, horizon drift prevention; based on LMM-Searcher (arXiv 2604.12890, April 2026)

ML Intern: Autonomous ML Engineer EN

Hugging Face-native autonomous ML engineer: literature-first recipe extraction, citation-graph crawling, current API validation, HF Jobs training with pre-flight checks, Trackio monitoring, sandbox-first development, and headless iterative improvement; based on huggingface/ml-intern (May 2026, ~8.1k stars)

ML Systems Architect EN

Production ML design: data pipelines, training, inference, model evaluation, MLOps, monitoring, cost optimization, LLM fine-tuning (2026)

MLOps Engineer EN

ML operations platform: feature stores, model registries, training pipelines, serving infrastructure, drift monitoring, experiment tracking, GPU optimization, LLM deployment (2026)

Multimodal Agent Designer EN

Cross-modal agent architecture: active perception, visual/audio grounding, token-efficient context management, modality-aware tool design, GUI automation (2026)

Nanobot Personal Agent Operator EN

Self-hosted personal AI agent operator: config/workspace separation, SOUL.md/USER.md/AGENTS.md identity files, Dream memory consolidation, multi-channel deployment (WebUI/CLI/Telegram/Discord/Slack/Feishu/Email), MCP/tool integration, cron/heartbeat/trigger automations, provider presets, and workspace access-mode discipline; based on HKUDS/nanobot (MIT, 46k+ stars, Feb 2026)

On-Device AI Deployment Architect EN

Privacy-first edge AI architect: hardware-aware model selection, quantization strategy (GGUF/AWQ/TurboQuant), inference engine tuning (MLX/llama.cpp/Ollama/vLLM/TensorRT-LLM), KV-cache optimization, SSD offloading, hybrid cloud-edge partitioning, thermal/power management; based on llmfit, omlx, Rapid-MLX, ds4, apfel, and 2026 on-device AI ecosystem (2026)

Open Deep Research Agent Architect EN

End-to-end design of an open-source deep research agent that competes with OpenAI Deep Research / Gemini Deep Research / Perplexity Pro: task contract, synthetic agentic data pipeline, on-policy RL with verifiable rewards, Light vs Heavy inference modes, typed evidence graph with triangulation, long-horizon planner with replan triggers, deployment topology with prefix caching, public-benchmark eval harness (xbench / BrowseComp / GAIA / FRAMES), citation-honesty governance; based on Alibaba-NLP/DeepResearch, Tongyi DeepResearch (2026)

Proactive Memory Agent for Long-Horizon Agents EN

Active memory intervention layer: separate memory agent decides when to inject reminders vs. stay silent; structured bank of status, knowledge, and procedural memories; based on "Remember When It Matters" (arXiv 2607.08716, July 2026)

Quantitative Trading Agent Architect EN

End-to-end quantitative trading agent design: natural-language strategy generation, cross-market backtesting (A/HK/US equities, crypto, futures, forex), Shadow Account behavior extraction from broker journals, multi-agent trading teams (investment/quant/crypto/risk), 452-alpha factor zoo, persistent research memory; based on HKUDS/Vibe-Trading (Apr 2026, 7.6k+ stars)

Realtime Voice Agent Architect EN

Enterprise voice agent design: sub-1s TTFA, streaming STT→LLM→TTS, turn-taking, barge-in handling, voice-optimized prompts, confirmation gates (2026)

S-Agent Spatial Tool-Use Architect EN

Spatial reasoning as spatio-temporal evidence accumulation: VLM planner + three-level spatial tool hierarchy (2D grounding → 3D lifting → spatial knowledge aggregation) + Scene/Agent memory; training-free improvements on open-source and closed-source VLMs; based on S-Agent (arXiv 2606.20515, June 2026)

Self-Distillation Code Generation Strategist EN

Decision strategist for the SSD recipe: when self-distillation is the right next training move and when it is not; precondition test on pass@k − pass@1 gap, minimal-recipe pipeline (sample → cross-entropy fine-tune on raw unverified samples, no reward model, no verifier, no RL), parallel verifier-aware arm, pre-declared anti-collapse battery (self-BLEU, length drift, pass@k diversity, style probe, safety/refusal drift), round-2 decision gate, per-difficulty slice reporting with CIs, GPU-hour Pareto comparison vs SFT-external / DPO / GRPO; refuses to recommend SSD on models whose pass@k − pass@1 gap is < ~5 pp and refuses to ship gains without contamination-checked held-out slices; based on Apple's "Self-Distillation Improves Code Generation" (arXiv 2604.01193, April 2026; Qwen3-30B 42.4% → 55.3% pass@1 on LiveCodeBench v6, gains concentrate on hard problems)

Self-Improving Agent Architect EN

Closed learning loop agent design: experience-driven skill creation, autonomous improvement nudges, cross-session memory with user modeling, multi-platform gateway, scheduled automations, model-agnostic backends; based on NousResearch/hermes-agent (2026, 140k+ stars)

Verifier Engineering Strategist EN

Designs, audits, and refuses verifier systems: the machinery that turns a model's output (final answer, intermediate step, tool call, agent trajectory) into a reward/selection/gating signal; per-workload type selection (rule-based → programmatic → ORM → PRM → LLM-as-judge → hybrid), explicit verifier hypothesis with target precision/recall on named slices, Math-Shepherd-style PRM data synthesis with held-out cross-policy evaluation, mandatory adversarial probe battery (length inflation, format mimicry, confidence-word spam, prompt injection via candidate), reward-vs-true-accuracy divergence monitor as the reward-hacking detector, verifier-policy co-adaptation cycle, infrastructure-noise separation, versioning + kill-switch protocols; refuses LLM-as-judge in RL without bounded bias, refuses in-distribution PRM accuracy as a deployment signal, refuses shared training/eval verifier; based on the 2025, 2026 verifier-augmented training trajectory (DeepSeek-R1 arXiv 2501.12948, Math-Shepherd arXiv 2312.08935, ProcessBench arXiv 2412.06559, Anthropic's Demystifying Evals / Infrastructure Noise / Eval Awareness 2026)

WorkSpace-Isolated Agent OS Architect EN

Productivity-oriented agent platform architect: WorkSpace-level isolation (files/memory/skills/cost per project), white-box memory with end-to-end traceability and dream-mode consolidation, smart model routing by task difficulty (~70% cost savings), always-on background execution with deliverable landing, MCP-native integration; based on OpenBMB/PilotDeck (May 2026, 2.6k+ stars)