ml-dev-principles

ml-dev-principles is a skill for Claude Code from YoungjaeDev/my-claude-plugins. It costs 159 tokens per session (1,856 once invoked), scanned A, original, MIT.

A set of working rules for machine-learning, multimodal, and computer-vision development. It covers choosing models and datasets, checking data, training, evaluation, error analysis, and GPU processing.

In plain words
What is it for?
Use it when selecting models or datasets, preparing experiments, training or fine-tuning models, separating training and test data, analysing false positives and false negatives, and improving multi-GPU pipelines.
Why use it?
It helps prevent problems such as unsuitable licenses, test-data leakage, weak data checks, and misleading evaluation results.

Skill for Claude Code

Written for Claude Code: shipped in a Claude Code plugin. Also seen: mentions Codex.

Part of the ml plugin — 4 skills shipped together

Good fit Use it when selecting models or datasets, preparing experiments, training or fine-tuning models, separating training and test data, analysing false positives and false negatives, and improving multi-GPU pipelines.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/youngjaedev/my-claude-plugins/ml-dev-principles
Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent
npx skills add YoungjaeDev/my-claude-plugins --skill ml-dev-principles
Clone the repo
git clone --depth 1 https://github.com/YoungjaeDev/my-claude-plugins

Made for: Claude Code.

Or install ml, the plugin that ships this one along with the rest of its 4 skills.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for ml-dev-principles

README.md
[![agentmods](https://agentmods.dev/badge/skills/youngjaedev/my-claude-plugins/ml-dev-principles/github.svg)](https://agentmods.dev/skills/youngjaedev/my-claude-plugins/ml-dev-principles)
Your own site
<a href="https://agentmods.dev/skills/youngjaedev/my-claude-plugins/ml-dev-principles"><img src="https://agentmods.dev/badge/skills/youngjaedev/my-claude-plugins/ml-dev-principles/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for ml-dev-principles

Your own site · 80×15
<a href="https://agentmods.dev/skills/youngjaedev/my-claude-plugins/ml-dev-principles"><img src="https://agentmods.dev/badge/skills/youngjaedev/my-claude-plugins/ml-dev-principles.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 159 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,856 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00159 $0.01856
Opus 5.5 $0.00064 $0.00742
Sonnet 5.5 $0.00032 $0.00371
Haiku 4.5 $0.00016 $0.00186

Measured 1mo ago against content hash 2d9671cec9a5, method: parsed. Prices are Anthropic first-party input rates as of 2026-10-07, from the pricing page.

Security

Grade A, and why

ml-dev-principles scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 1mo ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

plugins/ml/skills/ml-dev-principles/SKILL.md · 56 lines

How it starts

The opening of the file, as written. The whole thing — 56 lines — stays where its author put it; the contents beside it link to each section on GitHub.

ML / 멀티모달 개발 원칙

ML·멀티모달·CV 작업을 어떻게 진행할지에 대한 범용 규율(어떤 라이브러리를 쓰느냐가 아니라). 작업 시작 시 한 번 훑고 적용한다. 한 문장 요약: 집계/규칙 텍스트만 보지 말고 실제(이미지·원문·util)를 직접 봐라.

메타: 규칙에는 타입이 있다

hard-block(컴플라이언스·안전·산출물 누설)과 process-caution(보수성·스타일·관행)을 구분한다. 모든 규칙을 반사적으로 최강 해석으로 적용하면 정당한 빠른 경로까지 막힌다. 분류부터 하고 hard-block만 게이트로 삼는다. 나머지는 trade-off로 판단한다.

1. 후보 선정: 라이선스/자격은 규칙 원문부터 읽고 거른다

  • 모델·데이터를 "비상업/연구 라이선스라 제외" 식으로 선제 과필터링 금지. 과제가 연구/대회이고 규칙 원문이 실제로 금지하지 않으면 후보에 남긴다. permissive(Apache/MIT)는 동급일 때 tiebreaker지 gate가 아니다.
  • 왜: 선정 단계의 잘못된 배제는 파이프라인 전체로 전파되고 방향 자체를 틀어버린다. 되돌리는 비용이 가장 크다.
  • 적용: 자격 제약을 문자 그대로 나열 → 각각 hard-block / preference로 분류 → hard-block만 배제.

2. 평가 격리: 하드룰과 프로세스 주의를 분리한다

  • 하드룰(절대 금지): 제출·배포되는 산출 모델을 held-out/test 라벨로 학습·early-stop·모델선택·튜닝하지 않는다. 누설이다.
  • 허용·권장: test-GT(또는 임의 held-out proxy)를 flag-gated dev 진단으로 써서 파이프라인이 end-to-end로 도는지·숫자가 말이 되는지 sanity-check 하는 건 정당하다. 순수성을 이유로 무해한 파이프라인 점검을 거부하지 마라. 진단 경로는 flag 뒤로 격리해 산출물에 절대 흘러들지 않게 한다.
  • 왜: "산출물에 누설 금지"와 "배관 점검용으로도 라벨을 보지 마라"를 혼동하면, 시간 촉박할 때 빠른 피드백을 스스로 차단한다.
  • 적용: 제출 모델의 학습/선택 신호는 train-disjoint independent val에서만. test-GT는 산출물로 흐르는 어떤 것도 건드리지 않는 격리 경로에만.

3. EDA는 필수: 멀티모달이면 사람이 raw/overlay 이미지를 직접 본다

  • 모델링 전 EDA는 비협상. 라벨 있는 멀티모달/CV는 분포 통계로 끝내지 말고 사람(또는 Claude vision)이 raw 이미지(detection/seg면 overlay)를 실제로 본다.
  • 왜: 집계 통계는 "다 봤다"는 거짓 인상을 준다. 멀티모달은 이미지가 판별 근거를 들고 있어 라벨-이미지 정합은 통계로 알 수 없다.
  • 적용: interactive viewer(예: ml:cv-notebook의 exploration mode)로 strata별 샘플을 눈으로, 숫자 신뢰 전에.

4. 오류 분석: 지표가 아니라 실제 FP/FN 이미지를 본다

  • 틀린 케이스(FP/FN)는 실제 이미지를 열어 (질문·선택지·정답)과 대조해 왜 틀렸는지 추론한다(overlay 포함). 그 인사이트를 다음 데이터/모델 변경의 입력으로 삼는다.
  • 왜: 지표는 "얼마나 틀렸나"만 주고 "왜"는 못 준다. 왜(라벨 오류 / 이미지 모호 / 모델이 이미지 무시하고 text-prior 의존)는 픽셀에만 보이고, 그게 다음 iteration을 끈다.
  • 적용: eval 후 FP/FN을 이미지와 함께 덤프하는 코드를 실행 → 샘플 육안 리뷰 → 구체적 데이터/aug/prompt 변경으로 환원.

5. GPU 처리량: 모델 크기에 맞는 병렬 전략 + eff-batch 불변식

  • 병렬을 모델 크기에 맞춘다: 소형(약 3B 이하)은 순수 DDP, ZeRO/deepspeed는 끈다(파티셔닝 오버헤드 > 이득). ZeRO 카고컬트 금지.
  • eff-batch 불변식(per_device x accum x n_gpu = 상수)을 유지한 채 per-device batch만 올려 수렴 동역학은 안 바꾸고 속도만 얻는다.
  • LoRA rank·flash-attn·grad-checkpointing·bf16이 표준 레버. rank는 메모리 천장에 맞춰 튜닝.
  • 모델 여러 개를 병렬 추론할 땐 손수 짠 멀티GPU 런처보다 CUDA_VISIBLE_DEVICES 격리 + per-process가 단순·충분, 배칭은 엔진(예: vLLM)에 맡긴다.
  • 왜: "GPU 기계장치 더 동원" != 빠름. 핵심은 모델에 맞는 전략 + eff-batch 고정으로 정확도 회귀 없는 공짜 속도.
  • 적용: util 프로파일 → 90% 미만이면 exotic 병렬 꺼내기 전에 deepspeed/ZeRO 오버헤드나 과소 batch부터 의심.
  • 데이터셋 전체를 여러 GPU로 나눠 처리해야 하는 경우(위 "여러 모델 병렬 추론"과는 다른 케이스, 단일 모델로 대량 이미지를 처리하는 배치 작업)의 ProcessPool/CUDA_VISIBLE_DEVICES 격리, CUDA Streams, I/O+compute 파이프라인 구현 패턴은 references/gpu-parallel.md 참고.
  • 배치 추론 안티패턴(파일 단위 루프), BGR/RGB 순서, Ultralytics YOLO 의 내부 변환 예외, WandB 토글 등 라이브러리 레벨 함정은 references/ml-pitfalls.md 참고.

Read the full file on GitHub · 56 lines

Files

What ships with it

2 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 1mo ago Changed · +1 lines 2d9671cec9a5
  2. 1mo ago First seen · 55 lines · 159 tokens per session scan A 87698ed01f0c

Subscribe to this mod's changes

ml-dev-principles is a skill published in the GitHub repository YoungjaeDev/my-claude-plugins (2 stars, last pushed yesterday), licensed MIT. It adds 159 tokens to every session and 1,856 once invoked, about $0.0006 per session on Opus 5.5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-05.

Related

Other skills, from other repositories

arboreto

Infers candidate gene regulatory networks from bulk or single-cell expression data using AertsLab Arboreto GRNBoost2 and GENIE3. Use for transcription factor-target association ranking, compatible Dask execution, sparse expression inputs, and network stability checks.

K-Dense-AI/scientific-agent-skills · 55 tokens

torchdrug

Builds and troubleshoots TorchDrug 0.2.1 workflows for molecular graphs, property prediction, self-supervised pretraining, molecule generation, retrosynthesis, protein representation learning, and knowledge graph reasoning. Use when code imports torchdrug or needs its datasets, models, tasks, or Engine.

K-Dense-AI/scientific-agent-skills · 64 tokens

deepspot-m

Generates transcriptome-wide virtual spatial transcriptomics from H&E histology with DeepSpot-M. Used for predicted log1p-CPM expression from 224x224 tiles at about 20x, querying the released protein-coding gene panel by symbol, and whole-slide prediction after resolution-aware tiling with histolab.

K-Dense-AI/scientific-agent-skills · 70 tokens

pyhealth

Builds and validates PyHealth clinical machine-learning pipelines for EHR, signals, imaging, and medical codes. Use for PyHealth dataset loading, MIMIC-III/IV, eICU or OMOP prediction tasks, patient-level evaluation, mortality/readmission/length-of-stay modeling, medication recommendation, sleep staging, Trainer…

K-Dense-AI/scientific-agent-skills · 84 tokens

ml-training-recipes

Battle-tested PyTorch training recipes for all domains — LLMs, vision, diffusion, medical imaging, protein/drug discovery, spatial omics, genomics. Covers training loops, optimizer selection (AdamW, Muon), LR scheduling, mixed precision, debugging, and systematic experimentation. Use when training or fine-tuning…

Orchestra-Research/AI-Research-SKILLs · 88 tokens

nemo-mbridge-perf-expert-parallel-overlap

Validate and use MoE expert-parallel communication overlap in Megatron-Bridge, including overlapmoeexpertparallelcomm, delaywgradcompute, and flex dispatcher backends such as DeepEP and HybridEP.

NVIDIA/skills · 56 tokens