verify-conversion

E2E gate — verifies that applied patches produce a working, numerically sane end-to-end inference through the OpenVINO plugin. Handles HuggingFace/optimum-intel, native OV conversion (ovc/convert_model), and ONNX. Used by orchestrators as the mandatory gate before any PR is published.

Quality

83%

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.

A tight, executable gate skill: concrete commands and code for every conversion path, an explicit numerical sanity checkpoint, and a structured failure-reporting contract. Its weaknesses are modest — duplicated snippets, placeholder-laden PyTorch code, a step-numbering gap, and inlined detail that could live in a reference file.

Suggestions

Fix the step numbering (Step 1 is followed by "Step 3"; add an explicit Step 2 for the conversion/inference phase) and consider a small retry/feedback loop for transient export failures (timeout/OOM already has an int4 fallback — make that pattern explicit).

Deduplicate the zeros-fill inference snippet shared by the ONNX and ovc sections (extract once and reference it) and trim the opening blockquote that restates the description, to improve conciseness.

Move the PIPELINE_TAG_MAP and per-path recipes into a references/ file (e.g. conversion-paths.md) linked one level deep, keeping SKILL.md as a lean overview.

Dimension	Reasoning	Score
Conciseness	The body is lean and code-first with almost no explanation of concepts Claude already knows — a decision table, short snippets, and a structured result format. Not a 5 because the zeros-fill infer snippet is duplicated verbatim in the ONNX and ovc sections, and the opening blockquote restates the frontmatter description.	4 / 5
Actionability	Mostly executable: concrete optimum-cli and ovc commands, runnable Python per path, ready-to-use sanity-check functions, and an explicit JSON result schema. Not a 5 because the PyTorch path contains placeholders ("# Load your torch model", "torch_model = ...", "adjust shape") and "path/to/model.onnx" requires substitution, so it is not fully copy-paste ready across all paths.	4 / 5
Workflow Clarity	Clear sequence (determine path → convert → infer → sanity-check → report) with an explicit validation checkpoint (the numerical sanity check) and explicit failure handling ("do not silently swallow the failure", structured verify_result.json). Not a 5 because step numbering is incoherent (Step 1 is followed by "Step 3" with no labeled Step 2) and there is no feedback loop for error recovery — failures are reported and delegated to the orchestrator.	4 / 5
Progressive Disclosure	No bundle files exist, and the body is well-sectioned with self-contained per-path instructions that are easy to navigate. Not a 5 because the ~12-entry PIPELINE_TAG_MAP and per-path detail are inlined in a ~200-line body where a one-level-deep reference file would keep SKILL.md closer to an overview.	4 / 5
	Total	16 / 20 Passed

Description

83%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, concrete description that states what the skill does, which conversion stacks it handles, and exactly when it must be invoked. The only room for improvement is broader natural synonym coverage (e.g. OpenVINO IR/model.xml) and slightly sharper separation from generic conversion skills.

Suggestions

Add common synonyms like "OpenVINO IR", "model.xml", or "convert a model to OpenVINO" to broaden natural trigger coverage for the trigger_term_quality dimension.

Sharpen the distinction from a general conversion skill (e.g. mention it only verifies and gates rather than performing exploratory conversion matrix runs) to reduce overlap risk.

Dimension	Reasoning	Score
Specificity	"verifies that applied patches produce a working, numerically sane end-to-end inference through the OpenVINO plugin" plus "Handles HuggingFace/optimum-intel, native OV conversion (ovc/convert_model), and ONNX" names several concrete, specific capabilities. Not a 5 because the action list is effectively one verification action with a path enumeration, rather than multiple comprehensive distinct actions.	4 / 5
Completeness	Clearly answers "what" ("verifies... a working, numerically sane end-to-end inference through the OpenVINO plugin") and "when" explicitly ("Used by orchestrators as the mandatory gate before any PR is published") with a concrete trigger condition. The explicit 'when' clause is equivalent to a 'Use when...' clause, so the completeness cap of 3 does not apply.	5 / 5
Trigger Term Quality	Good natural term coverage: "E2E gate", "OpenVINO plugin", "HuggingFace", "optimum-intel", "ovc", "convert_model", "ONNX", "PR" — terms an orchestrator or user working on OV conversion would actually say. A few domain synonyms are missing (e.g. "OpenVINO IR", "model.xml/.bin", "convert"), keeping it below the comprehensive 5 anchor.	4 / 5
Distinctiveness Conflict Risk	A clear niche (OpenVINO E2E verification gate) with concrete distinguishing tool names; the "E2E gate" framing separates it from plain conversion skills. Minor overlap risk with a general model-conversion or 'try-conversion' skill since it covers the same conversion paths, which keeps it just below the 5 anchor's minimal-conflict bar.	4 / 5
	Total	17 / 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: openvinotoolkit/openvino
Path: .github/agents-prototype/skills/verify-conversion/SKILL.md
Commit: f766feb

Reviewed: 3 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.