CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

nnsight-remote-interpretability

Provides guidance for interpreting and manipulating neural network internals using nnsight with optional NDIF remote execution. Use when needing to run interpretability experiments on massive models (70B+) without local GPU resources, or when working with any PyTorch architecture.

Quality

78%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

High

Do not use without reviewing

Fix and improve this skill with Tessl

tessl review fix ./skills/mechanistic-interpretability/nnsight/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

67%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.

A well-structured, highly actionable skill: five concrete step-by-step workflows, executable code, honest limitation notes ('Gradient access not supported for vLLM or remote execution'), and a clearly signaled one-level-deep references bundle. Main weaknesses are repetition of model-loading boilerplate, a non-executable sweep snippet, and content in the body that duplicates the references bundle.

Suggestions

Factor the repeated `LanguageModel("gpt2", device_map="auto")` loading boilerplate out of every workflow code block (state it once) and drop the 'Key Value Proposition' code sample that duplicates Workflow 3, cutting significant tokens.

Make the Systematic Patching Sweep executable by defining `seq_len`, `corrupted`, and `compute_metric` or explicitly labeling it as a template, and add an explicit validation step (e.g., assert saved-tensor shapes after each trace).

Move the Key API Reference table and the Workflow 1 tutorial-style walkthrough (which duplicates Tutorial 1 in references/tutorials.md) into references/, keeping the body as a lean overview of the five workflows.

DimensionReasoningScore

Conciseness

The body is mostly efficient (no basic-concept explanations, tool-specific guidance throughout), but it repeats boilerplate — `LanguageModel("gpt2", device_map="auto")` loading appears in nearly every code block, and the 'Key Value Proposition' remote example duplicates Workflow 3 — plus time-sensitive trivia ('730+ stars') and two comparison tables that could be trimmed or moved.

3 / 5

Actionability

Workflows 1, 2, 4, and 5 give concrete, executable code with numbered steps, and Common Issues provides copy-paste fixes ('model.transformer.h[8].output[0][:] = 0', 'with model.trace("prompt", remote=True, timeout=300)'). Not a 5 because the Systematic Patching Sweep uses undefined variables (`seq_len`, `compute_metric`, `corrupted`) and reads as pseudocode.

4 / 5

Workflow Clarity

Each workflow is a clearly numbered step-by-step sequence ('# 1. Get clean activations', '# 2. Patch clean into corrupted run') with a checklist for Workflow 1 and a troubleshooting section for error recovery. Not a 5 because explicit validation checkpoints (e.g., verify saved tensors/shapes before analysis) are only implicit outside the single checklist.

4 / 5

Progressive Disclosure

The Reference Documentation table clearly signals all three real bundle files (references/README.md, api.md, tutorials.md) with one-line content descriptions, and they are one level deep. Not a 5 because the 430-line body inlines substantial material (full API table, a Workflow 1 that substantially duplicates Tutorial 1 in references/tutorials.md) that belongs in the bundle.

4 / 5

Total

15

/

20

Passed

Description

80%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.

A strong description that clearly and explicitly states both what the skill does and when to use it, with concrete trigger phrases and natural terminology. Its main weaknesses are the over-broad 'any PyTorch architecture' trigger (conflict risk with generic PyTorch work) and missing mention of distinctive capabilities like activation patching that would sharpen both specificity and trigger coverage.

DimensionReasoningScore

Specificity

Names several concrete capabilities — 'interpreting and manipulating neural network internals', 'NDIF remote execution', 'run interpretability experiments on massive models (70B+)' — in third person. Not a 5 because coverage has gaps: no mention of signature capabilities like activation patching, gradient access, or cross-prompt sharing that the body demonstrates.

4 / 5

Completeness

Explicitly answers both parts: what ('Provides guidance for interpreting and manipulating neural network internals using nnsight with optional NDIF remote execution') and when ('Use when needing to run interpretability experiments on massive models (70B+) without local GPU resources, or when working with any PyTorch architecture') with concrete trigger phrases.

5 / 5

Trigger Term Quality

Good natural-term coverage: 'interpretability experiments', 'massive models (70B+)', 'without local GPU resources', 'PyTorch architecture' — phrases a user with this need would plausibly say. Not a 5 because common variations like 'mechanistic interpretability', 'activation patching', 'access activations', or model names (Llama, GPT) are missing.

4 / 5

Distinctiveness Conflict Risk

The core niche (nnsight/NDIF interpretability) is clear, but the trigger 'or when working with any PyTorch architecture' is broad enough to fire on generic PyTorch development tasks, creating real overlap risk with general PyTorch/ML skills — more than the 'minor overlap' of the 4 anchor.

3 / 5

Total

16

/

20

Passed

Validation

93%

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

CriteriaDescriptionResult

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/mechanistic-interpretability/nnsight/SKILL.md

Reviewed

Table of Contents

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.