veomni-patchgen-model

Author or refresh a VeOmni model's patchgen-generated modeling under generated/ — GPU and/or NPU config, dense or MoE, text / VLM / Omni. Covers the patchgen decorators, sharing patches across sibling models via name_map, MoE fused-expert weight loading, Ulysses SP in multimodal forwards, __init__.py registration, running codegen, and the test cases. This is the modeling step of adding a new model, not only of refreshing an existing one. Trigger: 'add patchgen for a model', 'write a patch_gen_config', 'regenerate the generated modeling', 'add NPU patchgen', 'port a model to patchgen', 'transformers v5 migration'. Never hand-edit anything under generated/.

Quality

92%

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

92%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.

An executable, well-sequenced protocol with strong validation feedback loops and clean progressive disclosure into real reference files. Its only weak dimension is conciseness, dragged slightly by overall length and a few repeated emphases.

Suggestions

Consolidate the 'never edit generated/ by hand' rule to a single authoritative statement; the rule currently recurs in Phase 5, Common Pitfalls, and the description, and repeating it inflates length without adding clarity.

Tighten the empty-class-body (`: ...` vs `pass`) pitfall: lead with the actionable takeaway (verify the generated file imports cleanly when overriding a synthetic class) and move the 9-file enumeration and codegen.py internals into a reference file.

Dedupe the e2e `-k` keyword guidance, which appears in both Phase 6/Phase 7 tables and the Common Pitfalls list; a single canonical keyword-rules table would cut repetition.

Dimension	Reasoning	Score
Conciseness	The body is dense with genuinely non-obvious operational knowledge (patchgen decorators, OpSlots, NPU KERNEL_REGISTRY gaps, empty-class-body syntax quirks) and assumes Claude's competence rather than restating known concepts, but its ~750-line length and some repetition (the 'never edit generated/' rule stated 3+ times, the verbose empty-class-body pitfall) leave minor trimming room.	4 / 5
Actionability	Provides copy-paste-ready bash commands, executable PatchConfig/register-pattern code skeletons, concrete file paths, and specific test tables to append to, covering the common cases for GPU/NPU and dense/MoE/VLM/Omni variants.	5 / 5
Workflow Clarity	A clear Phase 0–8 sequence carries explicit Validation checkpoints (end of Phase 1, Phase 2, Phase 5), Phase 7 acceptance criteria, and Phase 8 review-gate feedback loops (safe/needs-attention/risky plus the patchgen --check --fix and pytest --collect-only error-recovery paths).	5 / 5
Progressive Disclosure	SKILL.md is a well-signaled spine with a capability-routed navigation table pointing one level deep to references/model-examples.md, moe.md, and multimodal.md (all present with the referenced sections), keeping category-specific bulk out of the spine and saving ~380 lines for simple text-only dense models.	5 / 5
	Total	19 / 20 Passed

Description

92%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 highly specific, well-scoped description with explicit trigger guidance and clear conflict boundaries. The only minor gap is keyword synonym coverage, which keeps trigger term quality at 4 rather than 5.

Dimension	Reasoning	Score
Specificity	Lists many concrete actions comprehensively across dense/MoE and text/VLM/Omni plus GPU/NPU — 'Author or refresh', 'patchgen decorators', 'sharing patches across sibling models via name_map', 'MoE fused-expert weight loading', 'Ulysses SP in multimodal forwards', '__init__.py registration', 'running codegen', 'test cases'.	5 / 5
Completeness	Explicitly answers both what (author/refresh modeling plus the enumerated coverage areas) and when (a dedicated Trigger clause with concrete phrases), matching the anchor that requires both.	5 / 5
Trigger Term Quality	An explicit Trigger clause gives six natural phrasings ('add patchgen for a model', 'write a patch_gen_config', 'regenerate the generated modeling', 'add NPU patchgen', 'port a model to patchgen', 'transformers v5 migration'), but these are power-user/domain phrases and a few common synonyms are absent.	4 / 5
Distinctiveness Conflict Risk	Occupies a clear niche (VeOmni patchgen-generated transformers modeling) with distinct triggers and explicit scope boundaries (non-transformers flux/movqgan/wan out of scope), minimizing conflict with sibling skills.	5 / 5
	Total	19 / 20 Passed

Validation

87%

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 — 14 / 16 Passed

Validation for skill structure

Criteria	Description	Result
skill_md_line_count	SKILL.md is long (759 lines); consider splitting into references/ and linking	Warning
relative_links	Relative link issues: 1 suspicious	Warning

	Total	14 / 16 Passed

Repository: ByteDance-Seed/VeOmni
Path: .agents/skills/veomni-patchgen-model/SKILL.md
Commit: b8a3edc

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.