veomni-profile

Use this skill for performance profiling and optimization. Two modes: (1) Analyze existing profile files (Chrome traces, memory snapshots) — write scripts to parse and summarize metrics per user requirements. (2) Generate profiles during development — configure ProfileConfig, run training, collect traces, analyze bottlenecks, and suggest optimizations. Trigger: 'profile', 'performance', 'slow', 'MFU', 'throughput', 'bottleneck', 'memory usage', 'trace', 'optimize training speed'.

Quality

84%

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

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 content is a well-structured, highly actionable guide: concrete paths, config examples, commands, and a diagnostic table let Claude execute both analysis and profiling workflows immediately. All four dimensions land at 4 rather than 5 due to minor repetition, a few high-level suggestions, a missing error-recovery loop after validation, and content that could be offloaded to reference files.

Suggestions

Deduplicate the analysis guidance: fold the Mode 2 'Quick analysis shortcuts' into Mode 1's script-writing step (or reference it) so kernel-breakdown and merge guidance appear once.

Add a brief error-recovery branch to Mode 2 Step 5 (e.g. what to check and revert if the loss curve diverges from baseline after an optimization).

Move the profile-config field details and the NPU differences into a reference file (e.g. references/npu.md) to keep SKILL.md a leaner overview, since no bundle files currently exist.

Dimension	Reasoning	Score
Conciseness	The body is dense and infrastructure-specific (file paths, config fields, output naming) with no padding explaining concepts Claude already knows. It is not 5 because there is noticeable repetition — kernel-time breakdown appears in Mode 1 step 3 and again in Mode 2's 'Quick analysis shortcuts', and merge guidance is stated twice.	4 / 5
Actionability	Provides a component table with exact file paths, a complete YAML profile config, runnable bash and Python snippets, and a bottleneck-to-solutions mapping table. It is not 5 because some guidance remains high-level ('Increase compute/comm overlap', 'check for memory leaks') and commands use placeholders like '<model>.yaml'.	4 / 5
Workflow Clarity	Both modes are clearly sequenced with numbered steps, and Mode 2 includes an explicit validation step ('Re-profile with the same config to compare before/after', 'Verify training correctness is preserved (loss matches baseline)'). It is not 5 because there is no error-recovery feedback loop for when validation fails (e.g. what to do if loss diverges after an optimization).	4 / 5
Progressive Disclosure	A single-file skill with no bundle files present; the body is well-organized into clear sections (components, Mode 1, Mode 2, NPU) with no nested references. It is not 5 because at ~150 lines some content — the full profile-config field reference and the NPU differences — would be more appropriately split into reference files to keep SKILL.md a leaner overview.	4 / 5
	Total	16 / 20 Passed

Description

88%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: it names the domain, enumerates concrete capabilities for two distinct modes, and provides an explicit trigger list with natural user phrasing. The only weakness is trigger coverage that misses a few synonyms and file-extension terms, and some generic triggers that create minor overlap risk.

Dimension	Reasoning	Score
Specificity	The description lists multiple concrete actions across both modes — 'Analyze existing profile files (Chrome traces, memory snapshots)', 'write scripts to parse and summarize metrics', 'configure ProfileConfig, run training, collect traces, analyze bottlenecks, and suggest optimizations' — matching the comprehensive-coverage anchor. It is not 4 because there are no meaningful gaps in the action inventory for the two modes.	5 / 5
Completeness	It explicitly answers 'what' with two concretely described modes and 'when' with a dedicated 'Trigger:' clause listing natural phrases, matching the top anchor exactly. Not 4: the 'when' guidance is explicit and specific rather than merely present.	5 / 5
Trigger Term Quality	Trigger list ('profile', 'performance', 'slow', 'MFU', 'throughput', 'bottleneck', 'memory usage', 'trace', 'optimize training speed') has good natural-term coverage including colloquial phrases like 'slow'. It falls short of 5 because common synonyms and file extensions users might mention (e.g. 'latency', 'speed up', 'chrome trace', '.json.gz', '.pkl') are absent.	4 / 5
Distinctiveness Conflict Risk	Niche terms like 'MFU', 'throughput', and Chrome-trace references make the skill mostly distinct with minor overlap risk against general performance-optimization skills. It is not 5 because broad triggers like 'performance', 'slow', and 'optimize' could plausibly fire for non-profiling optimization requests.	4 / 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
referenced_paths_exist	Referenced path issues: 3 missing, 3 deeper-than-1-level	Warning

	Total	15 / 16 Passed

Repository: ByteDance-Seed/VeOmni
Path: .agents/skills/veomni-profile/SKILL.md
Commit: 86c5ec1

Reviewed: 8 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.