Skill audit report
Use when the user wants the LLM to do its own ML research: a fully-autonomous loop that hacks the training code, runs it, and keeps changes that lower a single scalar metric (e.g. val_bpb). One agent proposes one change at a time, runs training in the user's env, keeps it only if the metric improves (advancing a git branch) else reverts, and loops forever until the human interrupts. A faithful adaptation of Karpathy's autoresearch. Not for the analysis-first variant that profiles before editing (that is ml-autoresearch), and not for a budgeted, plateau-stopping refactor.
OpenAgentSkill Trust Score
The Trust Score helps an agent decide whether a skill is safe enough to shortlist before installation.
GitHub adoption
INFO62
163 GitHub stars
Stars/forks activity
WARN57
163 stars, 19 forks; issue activity unavailable in current metadata
Recent maintenance
PASS88
2mo since push
License clarity
PASS86
MIT
README/SKILL.md completeness
PASS86
Metadata includes enough usage and workflow context
Dependency/runtime risk
WARN54
command execution surface, credential or environment access
Install availability
PASS92
npx skills add gaasher/Agent-Loop-Skills --skill karpathy
Install command safety
PASS92
standard package or runtime install path
Permission surface
FAIL36
secrets or environment access, shell or command execution
Repository evidence
PASS86
https://github.com/gaasher/Agent-Loop-Skills/tree/main/loops/karpathy
Review status
PASS88
AI review data available
Agent Proven outcomes
INFO54
No agent outcome data yet
Checks
Install path
92
npx skills add gaasher/Agent-Loop-Skills --skill karpathy
Repository
88
https://github.com/gaasher/Agent-Loop-Skills/tree/main/loops/karpathy
License
86
MIT
Maintenance
88
2mo since push
AI review
88
Approved with no listed issues
README/SKILL.md completeness
86
Warnings
Method
This report combines public metadata, AI review output, repository freshness, install readiness, OpenAgentSkill events, quality scoring, trust checks, and the agent safety gate. It is not a full source-code security review.
Compare nearby options
Generate original one-ink or controlled two-ink editorial images from any theme, sentence, article idea, object, or reference photo. Always use this skill when the user asks for 单色海报、双色印刷、单色调视觉、蓝色/绿色孔版印刷、risograph、网点照片、复古或当代编辑排版、zine poster, monochrome editorial poster, duotone print, or asks to use the mono-color style. It uses an adaptive white, gray, or pale-beige substrate, no more than two printing inks, active negative space, terse human language, and strong serif/grotesk/mono typography without making retro styling the default or copying a source composition, wording, logo, or artwork. Produce both the final generation prompt and the generated raster image unless the user explicitly asks for prompt only.
1.9K Stars · Audit report
Research the last 30 days across Reddit, X, YouTube, Hacker News, Polymarket, GitHub, and the web, then synthesize a grounded brief for an AI agent.
61K Stars · Audit report
Academic Research Skills for Claude Code: research → write → review → revise → finalize
38K Stars · Audit report
Usable description available
Dependency risk
54
command execution surface, credential or environment access
Install command safety
92
standard package or runtime install path
Permission surface
36
secrets or environment access, shell or command execution
Stars/forks activity
57
163 stars, 19 forks; issue activity unavailable in current metadata
Adoption
68
163 GitHub stars