@r1z4x/tezgah 1.0.0

package/skills/ai-research/SKILL.md

@@ -0,0 +1,101 @@
+---
+name: ai-research
+description: >
+  The vendored AI-research-SKILLs library - 98 upstream skills in 23 categories,
+  MIT - that now ships with tezgah. Use when a research line needs AI or ML
+  machinery: training or fine-tuning a model, serving or quantizing one,
+  evaluating a benchmark, reading model internals, building a retrieval or agent
+  pipeline, tracking runs, or writing the result up. Read the one entry the work
+  needs, never the tree. Do NOT use for code discovery ("where is X", "who calls
+  Y") - that is the code graph.
+---
+
+# ai-research
+
+Ninety-eight skills from Orchestra Research's `AI-research-SKILLs`, vendored into
+this repository at revision `773a529` (2026-06-15, MIT) and kept in the upstream
+layout, so a path here is the path there. Re-vendor, or verify the tree, with
+`bin/tezgah-import-ai-research`; what was left out and why is in `SOURCE`.
+
+Tezgah owns the research loop - the two-loop rhythm, the workspace under
+`.tezgah/research/<slug>/`, the protocol committed before the run, the six
+dimension review (`skills/research`). This library owns the machinery the
+experiment runs on. Reach for it after the question and the metric are locked, not
+before.
+
+## Read path - three steps, and only the last one costs anything
+
+1. **`index/<stage>.md`** - six files, one per stage of a research line. Read the
+   one that matches the work in hand; it lists every entry with its purpose.
+2. **the entry's `SKILL.md`** - the file the index line names, e.g.
+   `11-evaluation/lm-evaluation-harness/SKILL.md`. That is the upstream body.
+3. **its `references/`** - the detail the body points at, when the body is not
+   enough.
+
+| index file | stage |
+|---|---|
+| `index/1-frame.md` | framing the question, ideation, the research artifact |
+| `index/2-data.md` | data, retrieval, tokenization |
+| `index/3-train.md` | model, training, post-training, compression |
+| `index/4-measure.md` | evaluation, safety, interpretability, observability |
+| `index/5-run.md` | compute, serving, run tracking |
+| `index/6-write.md` | agents, structured output, media, paper writing |
+
+## Read the flags before you trust a body
+
+Each index line can carry a flag, and they are not decoration:
+
+- **generated** - upstream built this body from scraped documentation. It is thin,
+  sometimes mangled, and one of them still carries the placeholder *"Quick
+  reference patterns will be added as you use the skill."* Read the entry for its
+  coverage and its reference files, not for a workflow.
+- **stale-api** - the body names an API its library has since moved
+  (`langchain` chains, pre-2.5 `dspy`, `outlines.models.*`, AutoAWQ). The
+  concepts survive; the call signatures may not. Check the current docs before
+  running a snippet.
+- **dead-links: N** - the body links N reference files upstream does not ship. The
+  live links in the same entry still work.
+
+An entry with no body vendored would be a line without a file - there is none:
+every one of the 98 ships its `SKILL.md`.
+
+## Where tezgah's own rules still win
+
+The library knows ML. It does not know this contract. A recipe inside it never
+overrides: the locked metric and baseline in `state.json`; `protocol.md` committed
+before the run; a claim's falsification criterion; the review a claim passes
+before it is reported. Vendored advice that starts with "just run ..." on a
+cluster is advisory, not a plan, and anything it recommends that touches a live
+account, a paid service or the user's data still needs the user's yes first.
+
+## Entries worth knowing about before you need them
+
+- `10-optimization/ml-training-recipes/references/experiment-loop.md` - a
+  fixed-time-budget keep/discard loop, which is the closest external analogue of
+  this contract's protocol-before-run rule.
+- `20-ml-paper-writing/ml-paper-writing/references/citation-workflow.md` - how to
+  verify a reference instead of writing one from memory.
+- `20-ml-paper-writing/academic-plotting/SKILL.md` - chart-type rules, a
+  colourblind-safe palette and venue column widths; every line ends in a figure.
+- `22-agent-native-research-artifact/rigor-reviewer/SKILL.md` - the 1-5 scoring
+  anchors and severity inventory behind this contract's six-dimension review.
+- `22-agent-native-research-artifact/compiler/SKILL.md` - evidence fidelity: exact
+  numbers, a derived subset never named `Table N`, every evidence file carrying
+  its source.
+- `11-evaluation/lm-evaluation-harness/SKILL.md` - the runnable benchmark recipe a
+  research line usually quotes.
+
+## Provenance
+
+Vendored, not adapted: the bodies are byte-for-byte upstream, frontmatter
+included - the author line is upstream's, and two entries carry a contributing
+author's name rather than Orchestra Research's
+(`10-optimization/ml-training-recipes`, `14-agents/a-evolve`); a markdown file
+with no frontmatter carries one attribution comment.
+`SOURCE` records the revision and the exact drop list, `library.json` is the
+manifest (`bin/tezgah-import-ai-research --check` verifies every digest), and
+`NOTICE` records the licence. Neither the library nor any entry in it is a
+tezgah file: file an upstream bug upstream.
+
+Kill switch: `research-off` disables the routing that reaches this library, the
+same switch that turns off the research rule; the files stay on disk.

package/skills/ai-research/SOURCE

@@ -0,0 +1,38 @@
+# Source
+
+Vendored from https://github.com/orchestra-research/AI-research-SKILLs at `773a52944ba4747a18bd4ae9ade53fff041adcbc` (2026-06-15), MIT licence. Every body keeps its upstream
+frontmatter, `author` line included: 96 name Orchestra Research, and two
+(`10-optimization/ml-training-recipes`, `14-agents/a-evolve`) name the
+contributing author of that entry. Checkout for a re-vendor:
+`git clone https://github.com/orchestra-research/AI-research-SKILLs && git -C <clone> checkout 773a52944ba4747a18bd4ae9ade53fff041adcbc`.
+
+## What is vendored
+
+Every skill's `SKILL.md` and every markdown file beside it, in the upstream
+layout, because a vendored path is its provenance. 98 skills across 23
+categories, 365 files, 4202519 bytes.
+
+## What is not, and why
+
+Files: `03-fine-tuning/unsloth/references/llms-full.md`, `03-fine-tuning/unsloth/references/llms-txt.md`, `08-distributed-training/deepspeed/references/tutorials.md` - machine-scraped reference dumps, which the entry's own
+`SKILL.md` already distils.
+
+Directories: `20-ml-paper-writing/ml-paper-writing/templates`, `20-ml-paper-writing/systems-paper-writing/templates` - LaTeX class/bst/bib trees and example PDFs,
+re-downloadable from each venue.
+
+Everything else that is not markdown (PDFs, images, `.gitkeep`, scripts)
+is dropped by the importer's `*.md` filter, and `demos/`, `docs/`,
+`packages/`, `video-promo/` are upstream's repo furniture, not content.
+
+## The only modification
+
+A markdown file that carries no YAML frontmatter gets one attribution
+comment on its first line. `SKILL.md` files are byte-for-byte upstream,
+frontmatter included. `skills/ai-research/SKILL.md` is tezgah's own and is
+not part of the manifest.
+
+## Re-vendoring
+
+`bin/tezgah-import-ai-research --source <clone> --rev <sha>`, then commit
+the diff. `--check` verifies the committed tree against `library.json`
+without the clone, and is what CI runs.

package/skills/ai-research/index/1-frame.md

@@ -0,0 +1,18 @@
+<!-- generated by bin/tezgah-import-ai-research; do not edit -->
+
+# 1-frame. Frame the question and run the loop
+
+Vendored from [AI-research-SKILLs](https://github.com/orchestra-research/AI-research-SKILLs) at `773a52944ba4747a18bd4ae9ade53fff041adcbc` (MIT). Read the row you need, then the
+file it names; the body sits beside it. Flags: **generated** - upstream
+built the body from scraped docs, so it is thin; **stale-api** - the body
+names an API the library has since moved; **dead-links: N** - the body
+links N reference files upstream does not ship.
+
+| entry | file | what it covers | flags |
+|---|---|---|---|
+| `0-autoresearch-skill` | `0-autoresearch-skill/SKILL.md` | Orchestrates end-to-end autonomous AI research projects using a two-loop architecture. |  |
+| `brainstorming-research-ideas` | `21-research-ideation/brainstorming-research-ideas/SKILL.md` | Guides researchers through structured ideation frameworks to discover high-impact research directions. |  |
+| `creative-thinking-for-research` | `21-research-ideation/creative-thinking-for-research/SKILL.md` | Applies cognitive science frameworks for creative thinking to CS and AI research ideation. |  |
+| `compiler` | `22-agent-native-research-artifact/compiler/SKILL.md` | Compiles any research input — PDF papers, GitHub repositories, experiment logs, code directories, or raw notes — into a complete Agent-Native Research Artifact (ARA) with ... |  |
+| `research-manager` | `22-agent-native-research-artifact/research-manager/SKILL.md` | Records research provenance as a post-task epilogue, scanning conversation history at the end of a coding or research session to extract decisions, experiments, dead ends, claims ... |  |
+| `rigor-reviewer` | `22-agent-native-research-artifact/rigor-reviewer/SKILL.md` | Performs ARA Seal Level 2 semantic epistemic review on Agent-Native Research Artifacts, scoring six dimensions (evidence relevance, falsifiability, scope calibration, argument ... |  |

package/skills/ai-research/index/2-data.md

@@ -0,0 +1,21 @@
+<!-- generated by bin/tezgah-import-ai-research; do not edit -->
+
+# 2-data. Data, retrieval and tokenization
+
+Vendored from [AI-research-SKILLs](https://github.com/orchestra-research/AI-research-SKILLs) at `773a52944ba4747a18bd4ae9ade53fff041adcbc` (MIT). Read the row you need, then the
+file it names; the body sits beside it. Flags: **generated** - upstream
+built the body from scraped docs, so it is thin; **stale-api** - the body
+names an API the library has since moved; **dead-links: N** - the body
+links N reference files upstream does not ship.
+
+| entry | file | what it covers | flags |
+|---|---|---|---|
+| `huggingface-tokenizers` | `02-tokenization/huggingface-tokenizers/SKILL.md` | Fast tokenizers optimized for research and production. |  |
+| `sentencepiece` | `02-tokenization/sentencepiece/SKILL.md` | Language-independent tokenizer treating text as raw Unicode. |  |
+| `nemo-curator` | `05-data-processing/nemo-curator/SKILL.md` | GPU-accelerated data curation for LLM training. |  |
+| `ray-data` | `05-data-processing/ray-data/SKILL.md` | Scalable data processing for ML workloads. |  |
+| `chroma` | `15-rag/chroma/SKILL.md` | Open-source embedding database for AI applications. |  |
+| `faiss` | `15-rag/faiss/SKILL.md` | Facebook's library for efficient similarity search and clustering of dense vectors. |  |
+| `pinecone` | `15-rag/pinecone/SKILL.md` | Managed vector database for production AI applications. |  |
+| `qdrant` | `15-rag/qdrant/SKILL.md` | High-performance vector similarity search engine for RAG and semantic search. |  |
+| `sentence-transformers` | `15-rag/sentence-transformers/SKILL.md` | Framework for state-of-the-art sentence, text, and image embeddings. |  |

package/skills/ai-research/index/3-train.md

@@ -0,0 +1,48 @@
+<!-- generated by bin/tezgah-import-ai-research; do not edit -->
+
+# 3-train. Model, training, post-training and optimization
+
+Vendored from [AI-research-SKILLs](https://github.com/orchestra-research/AI-research-SKILLs) at `773a52944ba4747a18bd4ae9ade53fff041adcbc` (MIT). Read the row you need, then the
+file it names; the body sits beside it. Flags: **generated** - upstream
+built the body from scraped docs, so it is thin; **stale-api** - the body
+names an API the library has since moved; **dead-links: N** - the body
+links N reference files upstream does not ship.
+
+| entry | file | what it covers | flags |
+|---|---|---|---|
+| `litgpt` | `01-model-architecture/litgpt/SKILL.md` | Implements and trains LLMs using Lightning AI's LitGPT with 20+ pretrained architectures (Llama, Gemma, Phi, Qwen, Mistral). |  |
+| `mamba` | `01-model-architecture/mamba/SKILL.md` | State-space model with O(n) complexity vs Transformers' O(n²). | dead-links: 2 |
+| `nanogpt` | `01-model-architecture/nanogpt/SKILL.md` | Educational GPT implementation in ~300 lines. |  |
+| `rwkv` | `01-model-architecture/rwkv/SKILL.md` | RNN+Transformer hybrid with O(n) inference. | stale-api |
+| `torchtitan` | `01-model-architecture/torchtitan/SKILL.md` | Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). |  |
+| `axolotl` | `03-fine-tuning/axolotl/SKILL.md` | Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support | generated |
+| `llama-factory` | `03-fine-tuning/llama-factory/SKILL.md` | Expert guidance for fine-tuning LLMs with LLaMA-Factory - WebUI no-code, 100+ models, 2/3/4/5/6/8-bit QLoRA, multimodal support | generated |
+| `peft` | `03-fine-tuning/peft/SKILL.md` | Parameter-efficient fine-tuning for LLMs using LoRA, QLoRA, and 25+ methods. |  |
+| `unsloth` | `03-fine-tuning/unsloth/SKILL.md` | Expert guidance for fast fine-tuning with Unsloth - 2-5x faster training, 50-80% less memory, LoRA/QLoRA optimization | generated |
+| `grpo-rl-training` | `06-post-training/grpo-rl-training/SKILL.md` | Expert guidance for GRPO/RL fine-tuning with TRL for reasoning and task-specific model training |  |
+| `miles` | `06-post-training/miles/SKILL.md` | Provides guidance for enterprise-grade RL training using miles, a production-ready fork of slime. |  |
+| `openrlhf` | `06-post-training/openrlhf/SKILL.md` | High-performance RLHF framework with Ray+vLLM acceleration. |  |
+| `simpo` | `06-post-training/simpo/SKILL.md` | Simple Preference Optimization for LLM alignment. |  |
+| `slime` | `06-post-training/slime/SKILL.md` | Provides guidance for LLM post-training with RL using slime, a Megatron+SGLang framework. |  |
+| `torchforge` | `06-post-training/torchforge/SKILL.md` | Provides guidance for PyTorch-native agentic RL using torchforge, Meta's library separating infra from algorithms. |  |
+| `trl-fine-tuning` | `06-post-training/trl-fine-tuning/SKILL.md` | Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training. |  |
+| `verl` | `06-post-training/verl/SKILL.md` | Provides guidance for training LLMs with reinforcement learning using verl (Volcano Engine RL). | dead-links: 1 |
+| `accelerate` | `08-distributed-training/accelerate/SKILL.md` | Simplest distributed training API. |  |
+| `deepspeed` | `08-distributed-training/deepspeed/SKILL.md` | Expert guidance for distributed training with DeepSpeed - ZeRO optimization stages, pipeline parallelism, FP16/BF16/FP8, 1-bit Adam, sparse attention | generated |
+| `megatron-core` | `08-distributed-training/megatron-core/SKILL.md` | Trains large language models (2B-462B parameters) using NVIDIA Megatron-Core with advanced parallelism strategies. |  |
+| `pytorch-fsdp2` | `08-distributed-training/pytorch-fsdp2/SKILL.md` | Adds PyTorch FSDP2 (fully_shard) to training scripts with correct init, sharding, mixed precision/offload config, and distributed checkpointing. |  |
+| `pytorch-lightning` | `08-distributed-training/pytorch-lightning/SKILL.md` | High-level PyTorch framework with Trainer class, automatic distributed training (DDP/FSDP/DeepSpeed), callbacks system, and minimal boilerplate. |  |
+| `ray-train` | `08-distributed-training/ray-train/SKILL.md` | Distributed training orchestration across clusters. | dead-links: 1 |
+| `awq` | `10-optimization/awq/SKILL.md` | Activation-aware weight quantization for 4-bit LLM compression with 3x speedup and minimal accuracy loss. | stale-api |
+| `bitsandbytes` | `10-optimization/bitsandbytes/SKILL.md` | Quantizes LLMs to 8-bit or 4-bit for 50-75% memory reduction with minimal accuracy loss. |  |
+| `flash-attention` | `10-optimization/flash-attention/SKILL.md` | Optimizes transformer attention with Flash Attention for 2-4x speedup and 10-20x memory reduction. | dead-links: 2 |
+| `gguf` | `10-optimization/gguf/SKILL.md` | GGUF format and llama.cpp quantization for efficient CPU/GPU inference. |  |
+| `gptq` | `10-optimization/gptq/SKILL.md` | Post-training 4-bit quantization for LLMs with minimal accuracy loss. |  |
+| `hqq` | `10-optimization/hqq/SKILL.md` | Half-Quadratic Quantization for LLMs without calibration data. |  |
+| `ml-training-recipes` | `10-optimization/ml-training-recipes/SKILL.md` | Battle-tested PyTorch training recipes for all domains — LLMs, vision, diffusion, medical imaging, protein/drug discovery, spatial omics, genomics. |  |
+| `knowledge-distillation` | `19-emerging-techniques/knowledge-distillation/SKILL.md` | Compress large language models using knowledge distillation from teacher to student models. |  |
+| `long-context` | `19-emerging-techniques/long-context/SKILL.md` | Extend context windows of transformer models using RoPE, YaRN, ALiBi, and position interpolation techniques. |  |
+| `model-merging` | `19-emerging-techniques/model-merging/SKILL.md` | Merge multiple fine-tuned models using mergekit to combine capabilities without retraining. |  |
+| `model-pruning` | `19-emerging-techniques/model-pruning/SKILL.md` | Reduce LLM size and accelerate inference using pruning techniques like Wanda and SparseGPT. |  |
+| `moe-training` | `19-emerging-techniques/moe-training/SKILL.md` | Train Mixture of Experts (MoE) models using DeepSpeed or HuggingFace. |  |
+| `speculative-decoding` | `19-emerging-techniques/speculative-decoding/SKILL.md` | Accelerate LLM inference using speculative decoding, Medusa multiple heads, and lookahead decoding techniques. | dead-links: 1 |

package/skills/ai-research/index/4-measure.md

@@ -0,0 +1,25 @@
+<!-- generated by bin/tezgah-import-ai-research; do not edit -->
+
+# 4-measure. Measurement, safety and interpretability
+
+Vendored from [AI-research-SKILLs](https://github.com/orchestra-research/AI-research-SKILLs) at `773a52944ba4747a18bd4ae9ade53fff041adcbc` (MIT). Read the row you need, then the
+file it names; the body sits beside it. Flags: **generated** - upstream
+built the body from scraped docs, so it is thin; **stale-api** - the body
+names an API the library has since moved; **dead-links: N** - the body
+links N reference files upstream does not ship.
+
+| entry | file | what it covers | flags |
+|---|---|---|---|
+| `nnsight` | `04-mechanistic-interpretability/nnsight/SKILL.md` | Provides guidance for interpreting and manipulating neural network internals using nnsight with optional NDIF remote execution. |  |
+| `pyvene` | `04-mechanistic-interpretability/pyvene/SKILL.md` | Provides guidance for performing causal interventions on PyTorch models using pyvene's declarative intervention framework. |  |
+| `saelens` | `04-mechanistic-interpretability/saelens/SKILL.md` | Provides guidance for training and analyzing Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features. |  |
+| `transformer-lens` | `04-mechanistic-interpretability/transformer-lens/SKILL.md` | Provides guidance for mechanistic interpretability research using TransformerLens to inspect and manipulate transformer internals via HookPoints and activation caching. |  |
+| `constitutional-ai` | `07-safety-alignment/constitutional-ai/SKILL.md` | Anthropic's method for training harmless AI through self-improvement. | dead-links: 3 |
+| `llamaguard` | `07-safety-alignment/llamaguard/SKILL.md` | Meta's 7-8B specialized moderation model for LLM input/output filtering. | dead-links: 1 |
+| `nemo-guardrails` | `07-safety-alignment/nemo-guardrails/SKILL.md` | NVIDIA's runtime safety framework for LLM applications. | dead-links: 1 |
+| `prompt-guard` | `07-safety-alignment/prompt-guard/SKILL.md` | Meta's 86M prompt injection and jailbreak detector. |  |
+| `bigcode-evaluation-harness` | `11-evaluation/bigcode-evaluation-harness/SKILL.md` | Evaluates code generation models across HumanEval, MBPP, MultiPL-E, and 15+ benchmarks with pass@k metrics. |  |
+| `lm-evaluation-harness` | `11-evaluation/lm-evaluation-harness/SKILL.md` | Evaluates LLMs across 60+ academic benchmarks (MMLU, HumanEval, GSM8K, TruthfulQA, HellaSwag). |  |
+| `nemo-evaluator` | `11-evaluation/nemo-evaluator/SKILL.md` | Evaluates LLMs across 100+ benchmarks from 18+ harnesses (MMLU, HumanEval, GSM8K, safety, VLM) with multi-backend execution. |  |
+| `langsmith` | `17-observability/langsmith/SKILL.md` | LLM observability platform for tracing, evaluation, and monitoring. |  |
+| `phoenix` | `17-observability/phoenix/SKILL.md` | Open-source AI observability platform for LLM tracing, evaluation, and monitoring. |  |

package/skills/ai-research/index/5-run.md

@@ -0,0 +1,23 @@
+<!-- generated by bin/tezgah-import-ai-research; do not edit -->
+
+# 5-run. Compute, serving and tracking
+
+Vendored from [AI-research-SKILLs](https://github.com/orchestra-research/AI-research-SKILLs) at `773a52944ba4747a18bd4ae9ade53fff041adcbc` (MIT). Read the row you need, then the
+file it names; the body sits beside it. Flags: **generated** - upstream
+built the body from scraped docs, so it is thin; **stale-api** - the body
+names an API the library has since moved; **dead-links: N** - the body
+links N reference files upstream does not ship.
+
+| entry | file | what it covers | flags |
+|---|---|---|---|
+| `lambda-labs` | `09-infrastructure/lambda-labs/SKILL.md` | Reserved and on-demand GPU cloud instances for ML training and inference. |  |
+| `modal` | `09-infrastructure/modal/SKILL.md` | Serverless GPU cloud platform for running ML workloads. |  |
+| `skypilot` | `09-infrastructure/skypilot/SKILL.md` | Multi-cloud orchestration for ML workloads with automatic cost optimization. |  |
+| `llama-cpp` | `12-inference-serving/llama-cpp/SKILL.md` | Runs LLM inference on CPU, Apple Silicon, and consumer GPUs without NVIDIA hardware. |  |
+| `sglang` | `12-inference-serving/sglang/SKILL.md` | Fast structured generation and serving for LLMs with RadixAttention prefix caching. |  |
+| `tensorrt-llm` | `12-inference-serving/tensorrt-llm/SKILL.md` | Optimizes LLM inference with NVIDIA TensorRT for maximum throughput and lowest latency. |  |
+| `vllm` | `12-inference-serving/vllm/SKILL.md` | Serves LLMs with high throughput using vLLM's PagedAttention and continuous batching. |  |
+| `mlflow` | `13-mlops/mlflow/SKILL.md` | Track ML experiments, manage model registry with versioning, deploy models to production, and reproduce experiments with MLflow - framework-agnostic ML lifecycle platform |  |
+| `swanlab` | `13-mlops/swanlab/SKILL.md` | Provides guidance for experiment tracking with SwanLab. |  |
+| `tensorboard` | `13-mlops/tensorboard/SKILL.md` | Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard - Google's ML visualization toolkit |  |
+| `weights-and-biases` | `13-mlops/weights-and-biases/SKILL.md` | Track ML experiments with automatic logging, visualize training in real-time, optimize hyperparameters with sweeps, and manage model registry with W&B - collaborative MLOps ... |  |

package/skills/ai-research/index/6-write.md

@@ -0,0 +1,35 @@
+<!-- generated by bin/tezgah-import-ai-research; do not edit -->
+
+# 6-write. Agents, structured output, media and writing
+
+Vendored from [AI-research-SKILLs](https://github.com/orchestra-research/AI-research-SKILLs) at `773a52944ba4747a18bd4ae9ade53fff041adcbc` (MIT). Read the row you need, then the
+file it names; the body sits beside it. Flags: **generated** - upstream
+built the body from scraped docs, so it is thin; **stale-api** - the body
+names an API the library has since moved; **dead-links: N** - the body
+links N reference files upstream does not ship.
+
+| entry | file | what it covers | flags |
+|---|---|---|---|
+| `a-evolve` | `14-agents/a-evolve/SKILL.md` | Provides guidance for automatically evolving and optimizing AI agents across any domain using LLM-driven evolution algorithms. |  |
+| `autogpt` | `14-agents/autogpt/SKILL.md` | Autonomous AI agent platform for building and deploying continuous agents. |  |
+| `crewai` | `14-agents/crewai/SKILL.md` | Multi-agent orchestration framework for autonomous AI collaboration. |  |
+| `langchain` | `14-agents/langchain/SKILL.md` | Framework for building LLM-powered applications with agents, chains, and RAG. | stale-api |
+| `llamaindex` | `14-agents/llamaindex/SKILL.md` | Data framework for building LLM applications with RAG. |  |
+| `dspy` | `16-prompt-engineering/dspy/SKILL.md` | Build complex AI systems with declarative programming, optimize prompts automatically, create modular RAG systems and agents with DSPy - Stanford NLP's framework for systematic LM ... | stale-api |
+| `guidance` | `16-prompt-engineering/guidance/SKILL.md` | Control LLM output with regex and grammars, guarantee valid JSON/XML/code generation, enforce structured formats, and build multi-step workflows with Guidance - Microsoft ... | stale-api |
+| `instructor` | `16-prompt-engineering/instructor/SKILL.md` | Extract structured data from LLM responses with Pydantic validation, retry failed extractions automatically, parse complex JSON with type safety, and stream partial results with ... |  |
+| `outlines` | `16-prompt-engineering/outlines/SKILL.md` | Guarantee valid JSON/XML/code structure during generation, use Pydantic models for type-safe outputs, support local models (Transformers, vLLM), and maximize inference speed with ... | stale-api |
+| `audiocraft` | `18-multimodal/audiocraft/SKILL.md` | PyTorch library for audio generation including text-to-music (MusicGen) and text-to-sound (AudioGen). |  |
+| `blip-2` | `18-multimodal/blip-2/SKILL.md` | Vision-language pre-training framework bridging frozen image encoders and LLMs. |  |
+| `clip` | `18-multimodal/clip/SKILL.md` | OpenAI's model connecting vision and language. |  |
+| `cosmos-policy` | `18-multimodal/cosmos-policy/SKILL.md` | Evaluates NVIDIA Cosmos Policy on LIBERO and RoboCasa simulation environments. |  |
+| `llava` | `18-multimodal/llava/SKILL.md` | Large Language and Vision Assistant. |  |
+| `openpi` | `18-multimodal/openpi/SKILL.md` | Fine-tune and serve Physical Intelligence OpenPI models (pi0, pi0-fast, pi0.5) using JAX or PyTorch backends for robot policy inference across ALOHA, DROID, and LIBERO ... |  |
+| `openvla-oft` | `18-multimodal/openvla-oft/SKILL.md` | Fine-tunes and evaluates OpenVLA-OFT and OpenVLA-OFT+ policies for robot action generation with continuous action heads, LoRA adaptation, and FiLM conditioning on LIBERO ... |  |
+| `segment-anything` | `18-multimodal/segment-anything/SKILL.md` | Foundation model for image segmentation with zero-shot transfer. |  |
+| `stable-diffusion` | `18-multimodal/stable-diffusion/SKILL.md` | State-of-the-art text-to-image generation with Stable Diffusion models via HuggingFace Diffusers. |  |
+| `whisper` | `18-multimodal/whisper/SKILL.md` | OpenAI's general-purpose speech recognition model. |  |
+| `academic-plotting` | `20-ml-paper-writing/academic-plotting/SKILL.md` | Generates publication-quality figures for ML papers from research context. |  |
+| `ml-paper-writing` | `20-ml-paper-writing/ml-paper-writing/SKILL.md` | Write publication-ready ML/AI papers for NeurIPS, ICML, ICLR, ACL, AAAI, COLM. |  |
+| `presenting-conference-talks` | `20-ml-paper-writing/presenting-conference-talks/SKILL.md` | Generates conference presentation slides (Beamer LaTeX PDF and editable PPTX) from a compiled paper with speaker notes and talk script. |  |
+| `systems-paper-writing` | `20-ml-paper-writing/systems-paper-writing/SKILL.md` | Comprehensive guide for writing systems papers targeting OSDI, SOSP, ASPLOS, NSDI, and EuroSys. |  |