quantized-export

Export a promoted fine-tuned model in the right deployment format — merged safetensors, LoRA-only, GGUF with imatrix, or FP8. Use after a checkpoint passes promotion, when choosing a quantization format for a target device, or when an exported model fails its smoke test.

Quality

82%

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

77%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 delivers excellent workflow clarity with a mandatory, precisely-gated smoke test and good progressive disclosure through a single well-signaled reference file. Its weakness is token efficiency: key warnings (GB10/NVFP4, INT4 on long-context workloads) are repeated multiple times and some explanations cover ground Claude already knows.

Suggestions

State the GB10/NVFP4 exception once in the Format Map and reference it from the Worked Picks table, YAML snippet, and Spark-users note ('skip NVFP4 on GB10 — see Format Map') instead of restating the ~32% slowdown and its cause in multiple places.

Condense the merged-vs-LoRA-only bullet to the decision rule (portability vs. footprint / multi-adapter serving, plus the wrong-revision-base hazard) without explaining that merging folds the adapter into base weights.

Merge the Worked Picks table's 'long-context/code/math — never INT4' row into the Workload Overrides section it duplicates, keeping a single canonical statement of the INT4 failure mode.

Dimension	Reasoning	Score
Conciseness	The GB10/NVFP4 warning is repeated four times (Format Map bullet, Worked Picks row, YAML comment, Spark-users paragraph) and the INT4-long-context point twice, while the merged-vs-LoRA bullet restates mechanics Claude already knows. The genuinely non-obvious content (imatrix builds, the SM121 cvt.e2m1x2 detail, failure signatures) earns its tokens, but the repetition goes beyond 'minor instances that could be trimmed'.	3 / 5
Actionability	The body provides a copy-paste smoke-test command with exit-code semantics, a quick-decision YAML snippet, and a worked-picks lookup table, with per-format export commands correctly deferred to the real references/export-commands.md (verified to contain complete runnable sequences). The gap keeping it from 5 is that the body itself contains only one executable command and the format-choice guidance is decision-level rather than runnable.	4 / 5
Workflow Clarity	The sequence is explicit with strong validation: format selection with workload overrides, then a mandatory numbered smoke test with exact gates (byte-match for lossless exports, grader-verdict agreement for lossy), non-zero-exit gating, failure signatures mapping symptoms to causes, and re-run triggers on quant-method or runtime version bumps.	5 / 5
Progressive Disclosure	The bundle is one clearly-signaled, one-level-deep file (references/export-commands.md, verified present) referenced twice with an accurate description of its contents; decision rationale stays inline while runnable commands are externalized, and sections are well-organized and easy to navigate.	5 / 5
	Total	17 / 20 Passed

Description

87%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 states what the skill does and gives three explicit, natural trigger conditions in third-person voice. The only weakness is that the smoke-test half of the skill appears solely as a trigger rather than a stated capability, leaving slightly incomplete capability coverage.

Dimension	Reasoning	Score
Specificity	The description enumerates four concrete deployment formats ('merged safetensors, LoRA-only, GGUF with imatrix, or FP8') and names the core action, but the capability set is a single verb — the smoke-test workflow that constitutes half the skill surfaces only as a trigger phrase, a minor coverage gap.	4 / 5
Completeness	It explicitly answers both questions: the 'what' is 'Export a promoted fine-tuned model in the right deployment format' with formats enumerated, and the 'when' is an explicit 'Use after... when... or when...' clause with three concrete trigger phrases.	5 / 5
Trigger Term Quality	Natural domain phrases like 'choosing a quantization format for a target device', 'an exported model fails its smoke test', and 'checkpoint passes promotion' are present alongside format names (GGUF, FP8, LoRA), but common user variations such as 'convert a model', 'compress a model', or 'run locally' are missing.	4 / 5
Distinctiveness Conflict Risk	It occupies a clear niche — post-promotion export of fine-tuned models — with triggers tied to the checkpoint-promotion pipeline, making confusion with adjacent fine-tuning or serving skills unlikely.	5 / 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: 1 missing	Warning

	Total	15 / 16 Passed

Repository: wshobson/agents
Path: plugins/llm-finetuning/skills/quantized-export/SKILL.md
Commit: 4236bb9

Reviewed: 6 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.