sparse-autoencoder-training

Provides guidance for training and analyzing Sparse Autoencoders (SAEs) using SAELens to decompose neural network activations into interpretable features. Use when discovering interpretable features, analyzing superposition, or studying monosemantic representations in language models.

Quality

81%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

Securityby

Passed

No findings from the security scan

Quality

Content

75%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 skill body is highly actionable with complete executable workflows and well-organized progressive disclosure to real reference files. It loses points mainly on background prose that over-explains known concepts and on missing explicit validation feedback loops in the training workflow.

Suggestions

Trim the 'Polysemanticity & Superposition', 'Key Validation (Anthropic Research)', and 'Based on Anthropic's groundbreaking research' narrative prose that explains concepts Claude already knows.

Add explicit validation checkpoints and a fix->retry feedback loop to the training workflow (e.g., 'after N tokens, check L0/CE/dead-feature metrics; if dead features >5%, adjust L1 warm-up or enable ghost grads and resume').

Fix the broken reference: references/README.md links to a non-existent papers.md — either create it or remove the link.

Dimension	Reasoning	Score
Conciseness	The body is mostly efficient with dense code and tables, but includes padded background prose explaining concepts Claude already knows (e.g., the 'Polysemanticity & Superposition' section, 'Based on Anthropic's groundbreaking research', and 'Key Validation' narrative) that could be trimmed.	3 / 5
Actionability	Provides multiple complete, copy-paste-ready workflows (loading, training, feature analysis, steering, attribution) with specific config values, hyperparameter tables, and metrics targets covering the common cases.	5 / 5
Workflow Clarity	Three workflows have numbered step sequences and per-workflow checklists, but the training workflow is a long batch operation lacking explicit validation checkpoints and fix->retry feedback loops beyond a checklist mention.	4 / 5
Progressive Disclosure	Good overview-to-reference structure with a clearly signaled one-level-deep reference table linking to real files (README.md, api.md, tutorials.md); minor gaps include a broken reference (README links to a missing papers.md) and some API-style tables inlined in SKILL.md.	4 / 5
	Total	16 / 20 Passed

Description

87%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 clear, specific, and provides explicit 'what' and 'when' guidance with concrete trigger phrases well-matched to its specialized domain. Minor room for tighter, more concrete action enumeration and additional synonyms.

Dimension	Reasoning	Score
Specificity	Names the domain (SAE training and analysis via SAELens) and several concrete actions ('decompose neural network activations into interpretable features', 'training and analyzing'), though the actions remain somewhat abstract rather than enumerating specific operations.	4 / 5
Completeness	Clearly states what the skill does ('training and analyzing Sparse Autoencoders ... to decompose neural network activations into interpretable features') and provides an explicit 'Use when...' clause with concrete triggers.	5 / 5
Trigger Term Quality	Includes natural phrases a mechanistic-interpretability user would say ('discovering interpretable features', 'analyzing superposition', 'monosemantic representations') with good coverage, but a few common synonyms or file/extension cues are absent.	4 / 5
Distinctiveness Conflict Risk	Targets a narrow, well-defined niche (SAEs / SAELens for mechanistic interpretability) with distinct trigger language, giving minimal conflict risk with other skills.	5 / 5
	Total	18 / 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: foryourhealth111-pixel/Vibe-Skills
Path: bundled/skills/sparse-autoencoder-training/SKILL.md
Commit: d5ae560

Reviewed: 2 days 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.