nanogpt

Educational GPT implementation in ~300 lines. Reproduces GPT-2 (124M) on OpenWebText. Clean, hackable code for learning transformers. By Andrej Karpathy. Perfect for understanding GPT architecture from scratch. Train on Shakespeare (CPU) or OpenWebText (multi-GPU).

Quality

73%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Low

Low-risk findings worth noting

Fix and improve this skill with Tessl

tessl review fix ./skills/model-architecture/nanogpt/SKILL.md

Quality

Content

80%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

The body is highly actionable and well-structured with appropriate progressive disclosure to real reference files, but its multi-step training workflows omit explicit validation/verification checkpoints expected for batch operations.

Suggestions

Add validation checkpoints to the training workflows (e.g., after data prep, verify train.bin/val.bin exist and have expected size; after training, confirm sample.py produces coherent output before declaring success).

Include a brief feedback loop for the most common failure (e.g., CUDA OOM → reduce batch_size/block_size → retry) as an explicit workflow checkpoint rather than only in the Common issues section.

Trim the opening line 'nanoGPT is a simplified GPT implementation designed for learning and experimentation' since it restates the description and adds no new guidance.

Dimension	Reasoning	Score
Conciseness	The body is mostly efficient — commands, configs, and code dominate with little concept explanation Claude doesn't already know — but contains minor padding such as 'nanoGPT is a simplified GPT implementation designed for learning and experimentation', fitting score 4 rather than the fully lean 5.	4 / 5
Actionability	Provides fully executable, copy-paste-ready commands and configs (e.g., 'torchrun --standalone --nproc_per_node=8 train.py config/train_gpt2.py' and complete config blocks) covering the common Shakespeare, GPT-2, fine-tune, and custom-dataset cases, matching the score-5 anchor.	5 / 5
Workflow Clarity	Workflows have a clear Step 1/2/3 sequence but lack explicit validation checkpoints for these batch/long-running training operations, and per the rubric cap a batch skill without validation cannot score above 3.	3 / 5
Progressive Disclosure	The body is well-organized into Quick start, Common workflows, Common issues, and Advanced topics, with the latter linking one level deep to three real, clearly-signaled reference files (architecture.md, training.md, data.md), matching the score-5 anchor.	5 / 5
	Total	17 / 20 Passed

Description

66%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

The description is specific and action-oriented with strong natural trigger terms and a clear niche, but it lacks an explicit 'Use when...' clause that would tell Claude precisely when to invoke the skill.

Suggestions

Add an explicit 'Use when...' clause (e.g., 'Use when learning how GPT works, reproducing GPT-2, or training transformers from scratch') to lift completeness above 3.

Include common synonyms or file/dataset extensions users might mention (e.g., '.bin' tokenized data, 'char-level') to push trigger-term coverage toward 5.

Tighten the 'Perfect for understanding GPT architecture from scratch' phrasing into concrete trigger guidance rather than a marketing-style claim.

Dimension	Reasoning	Score
Specificity	Lists several concrete actions — 'Educational GPT implementation in ~300 lines', 'Reproduces GPT-2 (124M) on OpenWebText', 'Train on Shakespeare (CPU) or OpenWebText (multi-GPU)' — with only minor coverage gaps, fitting the score-4 anchor rather than the fully comprehensive 5.	4 / 5
Completeness	The 'what' is clear but there is no explicit 'Use when...' trigger clause; 'Perfect for understanding GPT architecture from scratch' only weakly implies the when, so per the rubric cap a missing explicit trigger guidance holds completeness at 3.	3 / 5
Trigger Term Quality	Includes natural terms users would say ('GPT', 'GPT-2', 'transformers', 'Shakespeare', 'OpenWebText', 'from scratch') with good coverage, though a few synonyms/extensions are missing, so it sits at 4 rather than 5.	4 / 5
Distinctiveness Conflict Risk	The nanoGPT/GPT-2-reproduction/educational niche is distinct with clear triggers, but there is minor overlap risk with general transformer-training skills, placing it at 4 rather than the minimal-conflict 5.	4 / 5
	Total	15 / 20 Passed

Validation

93%

Warnings & errors only

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 15 / 16 Passed

Validation for skill structure

Criteria	Description	Result
frontmatter_unknown_keys	Unknown frontmatter key(s) found; consider removing or moving to metadata	Warning

	Total	15 / 16 Passed

Repository: OpenLAIR/dr-claw
Path: skills/model-architecture/nanogpt/SKILL.md
Commit: 4b35756

Reviewed: 1 day ago

Table of Contents

Discovery Implementation Validation

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.