mixed-precision

Use FP16/BF16 mixed precision to accelerate training and reduce memory. Use when optimizing GPU performance.

Quality

67%

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

Fix and improve this skill with Tessl

tessl review fix ./researchclaw/skills/builtin/tooling/mixed-precision/SKILL.md

Quality

Content

72%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 exceptionally concise and well-structured for a simple single-purpose skill. It would benefit from moving the executable code template into the body and making the NaN-gradient feedback loop explicit.

Suggestions

Move the executable GradScaler/autocast training loop from the frontmatter code-template into the body so Claude has copy-paste-ready code.

Make the NaN-gradient handling an explicit feedback loop: 'If NaN gradients appear, reduce learning rate and retry' rather than a single warning bullet.

Number the core steps (wrap forward in autocast -> scale loss -> backward -> step -> update scaler) to make the sequence unambiguous.

Dimension	Reasoning	Score
Conciseness	The body is a lean bullet list with a single header and no padding or explanation of concepts Claude already knows; every token earns its place, matching the lean-and-efficient anchor.	5 / 5
Actionability	The bullets name concrete APIs ('torch.cuda.amp.autocast()', 'GradScaler') but the body lacks a copy-paste executable code block (the full snippet lives in frontmatter), so guidance is concrete yet incomplete.	3 / 5
Workflow Clarity	A rough sequence is implied (autocast the forward pass, scale loss, step optimizer) and a NaN-gradient warning hints at a feedback loop, but steps are unnumbered and the validate/retry checkpoint is only implicit.	3 / 5
Progressive Disclosure	The skill is under 50 lines, requires no external bundle files (references/scripts/assets are absent), and is organized under a single clear section header, satisfying the simple-skill exception for well-organized sections.	5 / 5
	Total	16 / 20 Passed

Description

62%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 concise and answers both what and when, with a clear technical niche. Its main weakness is a generic trigger clause and missing natural synonyms that would improve discoverability and reduce overlap risk.

Suggestions

Expand the 'Use when' clause to list concrete triggers, e.g. 'Use when the user mentions FP16, BF16, mixed precision, half precision, or wants to speed up training / reduce GPU memory'.

Add natural synonyms such as 'half precision' and 'AMP/autocast' so the description matches phrases users actually say.

Specify a concrete action verb beyond outcomes, e.g. 'Apply torch.cuda.amp autocast and GradScaler to accelerate training and reduce memory.'

Dimension	Reasoning	Score
Specificity	Names the domain ('FP16/BF16 mixed precision') and a couple of concrete outcomes ('accelerate training and reduce memory'), but these are benefits rather than a comprehensive list of operations, matching the 1-2 concrete actions anchor.	3 / 5
Completeness	It states both what ('accelerate training and reduce memory' via mixed precision) and when ('Use when optimizing GPU performance'), but the 'when' is a single generic trigger that could be more explicit with concrete phrases, matching the both-present-but-could-be-more-specific anchor.	4 / 5
Trigger Term Quality	Several relevant domain keywords appear ('FP16/BF16', 'mixed precision', 'GPU', 'memory', 'training'), but common natural variations such as 'half precision', 'speed up training', and 'amp' are missing, fitting the some-relevant-keywords-but-missing-synonyms anchor.	3 / 5
Distinctiveness Conflict Risk	Mixed precision training is a clear niche with distinct technical triggers, but the generic 'optimizing GPU performance' trigger creates minor overlap risk with other GPU-optimization skills, fitting the mostly-distinct-minor-overlap anchor.	4 / 5
	Total	14 / 20 Passed

Validation

100%

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 — 16 / 16 Passed

Validation for skill structure

No warnings or errors.

Repository: aiming-lab/AutoResearchClaw
Path: researchclaw/skills/builtin/tooling/mixed-precision/SKILL.md
Commit: be4ba47

Reviewed: 5 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.