remote-compute-modal

Run GPU jobs on the user's own Modal account via host.compute.create('byoc:modal', ...). Covers the create→submit→wait_for_notification flow, the compute_provider kernel for env setup, image/volume resolution, and the two approval cards. Load once you've decided to dispatch to Modal.

Quality

81%

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

77%

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

Highly actionable with executable examples, clear sequencing, and strong validation/recovery loops for a complex GPU-dispatch workflow. The main drawbacks are verbosity with some repetition, and a monolithic structure whose offloaded references cannot be confirmed against a present bundle.

Suggestions

Trim architectural justification and consolidate the timeout/harvest/idle-watchdog explanations so each fact appears once; the rationale prose ("That's the reason there are two approval cards…") can be shortened without losing operational value.

Move the long egress-fence, error-kind reference, and GPU-tier tables into separate reference files (e.g. egress.md, errors.md) so SKILL.md stays a navigable overview and the one-level references are verifiable.

Dimension	Reasoning	Score
Conciseness	Mostly efficient and largely specialized operational detail Claude would not already know, but the ~790-line body includes architectural rationale ("That's the reason there are two approval cards…") and repeats the timeout/harvest logic across several sections, so it could be tightened.	2 / 3
Actionability	Multiple complete, executable Python blocks carry real parameter values (e.g. `'image': r['image']`, `timeout_seconds=300`, `volumes={'/cache': 'af2-params'}`) with inline explanations, and the error-kind→remedy mapping gives copy-paste-ready guidance.	3 / 3
Workflow Clarity	A clear sequence is stated (read ledger → build env → submit → wait → harvest → close), backed by an explicit pre-check (`compute_provider_config()`), error-recovery feedback loops per `kind`, and a teardown checklist ("Before you finish the task: every handle closed…").	3 / 3
Progressive Disclosure	The body is well-sectioned and signals one-level-deep references to `env-setup.md` and `envs/*.py`, but it is largely a monolithic wall of text with egress/error/tier detail inlined that could be split out, and no `references/`/`scripts/`/`assets/` bundle exists to verify the referenced files are real.	2 / 3
	Total	10 / 12 Passed

Description

85%

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 specific, well-scoped description that answers both what and when with an explicit load trigger and a distinct Modal/GPU niche. Its one weakness is trigger-term breadth: it relies on internal API jargon more than on the range of natural phrases a user might say.

Suggestions

Broaden the natural trigger terms beyond "GPU jobs" and "Modal" — add user-facing phrasing like "run a job on my Modal account", "use Modal for a GPU run", or "dispatch to Modal" so the skill matches more variations of how a user asks.

Consider trimming the internal jargon ("byoc:modal", "compute_provider kernel") from the description or moving it after the trigger clause, so the first thing a reader sees is the natural task, not the API surface.

Dimension	Reasoning	Score
Specificity	Names multiple concrete actions — "Run GPU jobs", "create→submit→wait_for_notification flow", "compute_provider kernel for env setup", "image/volume resolution", and "the two approval cards" — matching the anchor for listing several specific concrete actions rather than a vague domain label.	3 / 3
Completeness	It clearly answers what the skill does and carries an explicit trigger clause ("Load once you've decided to dispatch to Modal") that is the equivalent of a "Use when…" trigger, so it is not capped at 2 for missing trigger guidance.	3 / 3
Trigger Term Quality	Some natural terms a user would say are present ("GPU jobs", "Modal", "dispatch to Modal"), but coverage is thin on common variations and the description leans heavily on internal jargon like "byoc:modal" and "compute_provider kernel" rather than broader natural trigger language.	2 / 3
Distinctiveness Conflict Risk	The niche is sharply bounded — "the user's own Modal account", the byoc:modal provider, and the two approval cards — making it clearly distinguishable and unlikely to trigger for a different skill.	3 / 3
	Total	11 / 12 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
skill_md_line_count	SKILL.md is long (790 lines); consider splitting into references/ and linking	Warning

	Total	15 / 16 Passed

Repository: UnicomAI/wanwu
Path: configs/microservice/bff-service/configs/agent-skills/claude-science/remote-compute-modal/SKILL.md
Commit: f618458

Reviewed: about 14 hours 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.