AI Engineering

NoPUA Lite is a compact prompt skill that teaches AI agents to persist through failures, debug systematically, and act proactively using trust and inner motivation rather than fear. Use it when an AI coding agent gets stuck, gives up too early, or needs a structured failure-recovery methodology.

Discover initializes an evo workspace for the current repository by exploring the codebase, proposing optimization dimensions, constructing a benchmark, and running a baseline experiment. Use it when starting a new evo run or instrumenting a project for autonomous optimization.

Evo Subagent Protocol defines the brief shape and execution loop that evo optimization subagents follow when dispatched from /optimize. Orchestrators use it to write valid briefs and debug subagent behavior; subagents use it to drive required verifier checks and read-edit-run iteration.

Finetuning selects or diagnoses a training move—SFT, LoRA, DPO/KTO/ORPO, RFT, GRPO/PPO/RLOO, RLHF—based on reward shape and literature, then gates runs with smoke tests, early stopping, and diagnostics. Use it when fine-tuning, post-training, reward design, or weight updates come up.

Optimize runs the evo optimization loop with parallel subagents until interrupted or the stall limit is reached. Use it after a baseline experiment is committed to evolve code changes against benchmarks and prune dead branches.

Optimize drives structured autoresearch iteration after evo:discover and the baseline commit, orchestrating subagents to run experiments and improve the current best frontier. Use it when the user invokes /evo:optimize or asks to try ideas, variants, or continue an evo search.

OpenTag connects collaboration platforms to local coding agents so users can trigger and manage coding work from Slack, GitHub, Linear, and other tools. Use it for CLI setup, background service operation, hosted Control Plane pairing, governed completion, and verification.

Autoresearch orchestrates end-to-end autonomous AI research projects using a two-loop architecture for rapid experiments and periodic synthesis, routing to domain-specific skills and producing papers; use when starting a research project or managing multi-hypothesis experiments.

Creative Thinking for Research applies eight cognitive science frameworks to generate novel research directions in computer science and AI. Use it when seeking genuinely novel directions via combinatorial creativity, analogical reasoning, and constraint manipulation.

Hugging Face Accelerate simplifies distributed training for PyTorch models by adding only four lines of code. Use it to run the same script on single or multiple GPUs, with mixed precision, DeepSpeed, or FSDP.

LLM Evaluation Harness evaluates LLMs across 60+ academic benchmarks using standardized prompts and metrics. Use when benchmarking model quality, comparing models, or tracking training progress.

ML Training Recipes provides battle-tested PyTorch training patterns for LLMs, vision, diffusion, medical imaging, and protein/drug discovery. Use it when training or fine-tuning neural networks, debugging loss spikes/OOM, choosing architectures, or optimizing GPU throughput.