Guide · Agent skill
How to use the Image Classification Best Practices agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
The cv-classification skill is a reference sheet that an AI agent reads when it works on image classification tasks, such as CIFAR-10, CIFAR-100 or ImageNet benchmarks. It gives the agent typical model choices for a given dataset size, common optimizer and learning-rate settings, standard augmentation methods, and approximate accuracy numbers to compare against. This guide shows you how to install it, trigger it with a normal request, and know what it can and cannot do.
Image Classification Best Practicesaiming-lab/AutoResearchClaw · MIT Download the skill Quick start
- 1Download the ZIP. Download the cv-classification ZIP file from this page.
- 2Unzip into the skills folder. Unzip it into ~/.claude/skills/ for personal use, or into .claude/skills/ inside a project for project-only use.
- 3Or upload in the Claude apps. In the Claude apps, go to Settings → Capabilities → Skills and upload the ZIP.
- 4Describe your classification task. Ask in plain language and name the dataset, such as CIFAR-10 or ImageNet.
- 5Review the suggestions. Check the architecture, optimizer and augmentation advice against your constraints.
- 6Verify the numbers. Compare your accuracy to the approximate benchmarks and confirm details against the cited papers.
What does the skill make the agent do?
Agent skills are folders of instructions that an AI agent can load when a task matches them. This skill's main file, SKILL.md, contains guidance only. When the agent is working on a vision classification task, it can draw on that guidance to suggest a starting point.
The skill covers three areas. For architecture selection, it groups models by scale: ResNet-18/34, WideResNet and Simple ViT for small scale (CIFAR-10/100); ResNet-50, EfficientNet-B0/B1 and DeiT-Small for medium scale; and ViT-B/16, ConvNeXt and Swin Transformer for large scale. For the training recipe, it lists optimizer options (AdamW with lr 1e-3 to 3e-4, or SGD with lr 0.1 and cosine decay), weight decay ranges, augmentations such as RandomCrop, RandomHorizontalFlip and Cutout/CutMix, warmup for transformers, and batch size ranges.
It also lists approximate benchmark accuracies, such as about 96% for ResNet-18 on CIFAR-10 and about 76% for ResNet-50 on ImageNet, so you can check whether a result looks reasonable.
How do you install the skill?
You have two options. Choose the one that fits how you work with your agent.
For a personal setup in Claude Code, download the ZIP file for this skill from this page and unzip it into ~/.claude/skills/. To limit it to one project, unzip it into the .claude/skills/ folder inside that project instead.
In the Claude apps, download the ZIP, then go to Settings → Capabilities → Skills and upload it.
- Personal (all projects): ~/.claude/skills/
- Single project: .claude/skills/ inside the project folder
- Claude apps: Settings → Capabilities → Skills, then upload the ZIP
What does the skill need before you start?
This skill has no listed requirements. It contains one file, SKILL.md, and it does not run scripts or use the network. You do not need to install any packages to use the guidance itself.
Running an actual training experiment is a separate job, and the skill does not provide it. You will need your own environment, datasets and code for that.
- Requirements listed: none
- Runs scripts: no
- Uses network: no
- Files included: SKILL.md
How do you trigger the skill with a natural request?
You do not need special syntax. Describe the classification task in plain language and mention the dataset or benchmark. The skill's description points to CIFAR, ImageNet and other classification benchmarks, and its trigger keywords include classification, image, cifar, imagenet, resnet, vision, cnn and vit.
Naming the dataset or model family in your request helps the agent match the skill to your task.
- "I'm setting up a CIFAR-10 baseline. Which architecture and optimizer should I start with?"
- "What learning rate and weight decay are typical for training ResNet-50 on ImageNet?"
- "My ViT gets 70% on CIFAR-100. Is that in the range I should expect?"
- "Which data augmentations are common for a small image classification dataset?"
What can you expect from the output?
Expect the agent to answer with suggestions drawn from the skill's reference lists. For example, for a small CIFAR dataset it will point toward ResNet-18/34 or a Simple ViT, and for a large ImageNet-scale task it will point toward ViT-B/16, ConvNeXt or Swin Transformer.
The numbers are approximate. Treat them as a sanity check, not a target you must hit, and confirm them against the original papers or your own runs before you report anything.
- Architecture suggestions matched to dataset scale
- Optimizer, learning-rate and weight decay starting points
- Augmentation and warmup suggestions
- Approximate accuracy figures for comparison
What are the safety notes?
The skill is text guidance. It runs no scripts and makes no network calls, according to the listing. That means installing it does not execute code on your machine.
The skill's own metadata cites two papers as references: He et al., Deep Residual Learning, CVPR 2016, and Dosovitskiy et al., An Image is Worth 16x16 Words, ICLR 2021. Check those sources if you need precise details.
- No scripts are run by the skill
- No network access is used by the skill
- Accuracy figures are approximate and should be verified
What tips help you get the most from it?
Use the skill as a starting point for a baseline, then tune from there. The guidance lists ranges, not single tested configurations, so pick a value within the range and adjust based on your validation results.
Match the model to your data size. The skill separates small, medium and large scale, so state your dataset and resolution clearly. Remember that batch size guidance applies only if your memory allows it.
- State the dataset name and scale in your request
- Start within the listed ranges and tune on validation data
- Compare results to the approximate benchmark figures, not exact targets
What people use it for
- Setting up a CIFAR-10 baseline with a ResNet-18 style model and a standard augmentation set
- Choosing between a CNN and a transformer for a medium-sized dataset
- Sanity-checking whether a ResNet-50 ImageNet result is in the expected range
- Learning which optimizer and learning-rate ranges are typical for vision training
- Picking augmentations such as RandomCrop, RandomHorizontalFlip or CutMix for a new dataset
Frequently asked questions
How do I install the cv-classification skill?
Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use, or into .claude/skills/ inside a project. In the Claude apps, upload the ZIP under Settings → Capabilities → Skills.
Does the cv-classification skill require anything else to install?
The listing shows no requirements. The skill contains one file, SKILL.md, runs no scripts and uses no network.
Is the cv-classification skill free?
The source material for this listing is licensed under MIT. The skill itself does not list a price, so check the repository for licence terms.
Can the skill train a model for me?
No. It is guidance text only and does not run anything itself. You still need your own code, data and compute to train a model.
Is it accurate enough for production decisions?
No. The file gives general starting points, not tested configurations for specific datasets, and its numbers are approximate. It is not meant for production deployment decisions.
Does it work for text or tabular classification?
No. The content covers only vision models and benchmarks, so it is not suited to text or tabular classification tasks.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
This skill is a useful reference for researchers setting up a CIFAR or ImageNet baseline, students learning typical architectures and training settings, and engineers who want rough accuracy figures to sanity-check results. It is not a source of tested configurations, so do not use it to make production deployment decisions or for non-image tasks.
Download the skill