CtrlK
BlogDocsLog inGet started
Tessl Logo

veomni-debug

Use this skill for ANY bug, error, crash, wrong output, loss divergence, gradient explosion, test failure, CUDA error, distributed training hang, checkpoint load failure, or unexpected behavior. Covers both quick fixes (clear root cause) and complex debugging (unclear cause). Trigger: 'fix bug', 'fix error', 'broken', 'crash', 'doesn't work', 'fails with', 'loss NaN', 'training hangs', 'FSDP error', 'OOM'.

71

Quality

87%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide

SecuritybySnyk

Passed

No findings from the security scan

SKILL.md
Quality
Evals
Security

Quality

Content

88%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 highly actionable, well-sequenced debugging protocol with executable commands, explicit verification gates, feedback loops, and checklists. The only soft spots are modest verbosity in the bisect section and a monolithic single-file structure with no progressive disclosure into bundle reference files.

Suggestions

Consider moving the detailed transformers/package bisect procedure into a references/ file and summarizing the decision rule inline, to tighten conciseness and improve progressive disclosure.

Trim or fold the prose justifications around the bisect commands into brief inline comments so the load-bearing flags (--active, two-worktree rationale) stay but the surrounding narration shrinks.

DimensionReasoningScore

Conciseness

The body is high-signal protocol with no general-concept padding (no 'what a gradient is' filler), and the bisect explanations are repo-specific load-bearing gotchas ("--active is load-bearing", "two worktrees and two venvs") rather than trivia; it is efficient with only minor trimming opportunities in the long transformers-bisect section, so it sits above the midpoint but short of the every-token-earns-its-place anchor.

4 / 5

Actionability

It provides copy-paste-ready, executable commands (uv venv/sync with VIRTUAL_ENV, git worktree add, make patchgen, make quality, pytest tests/<module>/) and concrete file paths (veomni/distributed/parallel_plan.py, .agents/knowledge/constraints.md), with justified placeholders (<reproducer>, <package>==<other-version>) covering the common bisect cases.

5 / 5

Workflow Clarity

It sequences a clear Quick Path and a Phase 1-5 Full Protocol with explicit validation checkpoints (Phase 3 "Verification gate", Phase 4 verify/collateral checks), feedback loops (Quick Path validate-retry, Phase 3 hypothesis loop), stop conditions, and checklists (Phase 5, Domain-Specific Checklists), matching the anchor for explicit validation steps and error-recovery loops.

5 / 5

Progressive Disclosure

Content is well-organized under clear headers (Quick Path vs Full Protocol, Phases, Stop Conditions, Pitfalls, Checklists, Appendix) and signals one-level-deep references to repo knowledge files ("Read .agents/knowledge/constraints.md"), but no bundle files exist and the long bisect procedure is inlined rather than split into a reference, leaving minor organization gaps short of the well-split anchor.

4 / 5

Total

18

/

20

Passed

Description

86%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, trigger-rich description that clearly states both scope and when to invoke it, with excellent natural trigger phrases. Its main weakness is that it catalogs bug types rather than the concrete debugging actions Claude will perform, and it mixes very generic triggers with the domain-specific ones.

Suggestions

Add 1-2 concrete actions Claude performs (e.g., 'reproduce, bisect dependencies, isolate root cause, fix, and capture knowledge') to lift specificity beyond domain-naming.

Qualify the generic triggers with the veomni/distributed-training scope (e.g., "'fix bug' in veomni training runs") to reduce overlap with general debugging skills.

DimensionReasoningScore

Specificity

It enumerates concrete bug types ("loss divergence, gradient explosion, ... checkpoint load failure") and names two work categories ("quick fixes (clear root cause)" / "complex debugging (unclear cause)"), but does not list the concrete actions Claude performs (reproduce, bisect, root-cause, fix, capture), so it sits at the domain-named/1-2-actions anchor rather than the multi-action anchor above.

3 / 5

Completeness

It explicitly answers both what ("Covers both quick fixes (clear root cause) and complex debugging (unclear cause)" plus enumerated bug types) and when ("Use this skill for ANY bug, error, crash..." with a concrete Trigger: clause), matching the anchor for clear what-and-when with concrete triggers.

5 / 5

Trigger Term Quality

An explicit Trigger: clause supplies natural user phrases with synonyms and domain terms ("'fix bug', 'fix error', 'broken', 'crash', 'doesn't work', 'fails with', 'loss NaN', 'training hangs', 'FSDP error', 'OOM'"), giving comprehensive coverage of terms a user would actually say.

5 / 5

Distinctiveness Conflict Risk

Domain-specific triggers ("FSDP error", "loss NaN", "training hangs", "gradient explosion") carve a distinct distributed-training-debugging niche, but the generic triggers ("'fix bug'", "'broken'", "'crash'") create minor overlap risk with general coding-debugging skills, so it is mostly distinct rather than minimal-conflict.

4 / 5

Total

17

/

20

Passed

Validation

100%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation16 / 16 Passed

Validation for skill structure

No warnings or errors.

Repository
ByteDance-Seed/VeOmni
Reviewed

Table of Contents

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.