Does that skill actually help your agent?
NVIDIA tests whether skills improve results, Univer gives agents editable spreadsheets, and Jane Street shares a tutor that makes you do the thinking.
11-18 November 2026. Edition 9. Star counts come from LLMRepos daily snapshots; GitHub counts may have changed.
Another skill installed. Another page of instructions for your agent. Has the work improved, or have you just given it more reading?
My pick is SkillEvaluator. NVIDIA's toolkit separates checks on the instructions from a live comparison of an agent working with and without them. Start with one skill you wrote and its keyless quality check. Then choose a task it should improve and compare the actual outputs. A neat file and a useful skill are different achievements.
Also this week: spreadsheets your agent can edit and you can inspect, code review with explicit file selection, a tutor that withholds the answer, and a practical refresher on what your retrieval system actually retrieves.
Moving fastest
NVIDIA/SkillEvaluator
SkillEvaluator checks agent skills and plugins for quality, security, overlap and compatibility, then offers live evaluations of their effect on tasks. It gained 2,749 stars (+409.1%) to reach 3,421.
For builders maintaining their own skills, skillevaluator quality-check ./my-skill is the small first step: no API key, GPU or Docker required. Read the findings before chasing the score. A separate harness check can flag whether a supported agent will load the skill at all.
The quickstart explicitly leaves the overall report INCOMPLETE after a passing quality check: security scanners and live agents have not run. Full evaluation needs extra setup, including a model provider and a sandbox for live runs; usage can cost money. Review the evaluation cases yourself. A generated test can share the blind spots of the instructions it tests.
dream-num/univer
An agent can explain the budget in chat. The budget still needs updating. Univer is an Office SDK for embedding editable spreadsheets and documents in your own app, with APIs that let your code or agent change the same content a person sees. It added 6,510 stars (+28.2%) to reach 29,606.
Try the README's quotation example: create a small sheet, change the number of seats, read back the recalculated total, then save and reopen it. That is a useful first loop for an agent that prepares quotes or budgets. The person reviewing the result gets cells and formulas to inspect.
The Apache-2.0 core covers Sheets and Docs editing and formula calculation. Your app supplies the model, tool orchestration and storage. Collaboration, Office file conversion and commercial AI extensions are separately distributed and licensed; saving a JSON snapshot is different from exporting an Excel file. Check which capabilities your workflow needs before treating the wider SDK demos as the free starter kit.
Demo thumbnail: Univer README, showing its document-editing surface. Wider SDK demos can include separately licensed features.
alibaba/open-code-review
A reviewer can sound thorough while quietly skipping half the diff. Open Code Review combines programmed file selection, grouping and rule matching with an agent that analyses the changes. It gained 4,247 stars (+8.5%) to reach 54,066, entering our largest band.
Try it on an old pull request with a known bug. Check which files it covered, whether it finds the defect and whether its comments point to the right lines. That gives you something more useful than a second opinion on code you already believe is fine.
The README distinguishes two modes. Native mode uses OCR's own model configuration; delegation mode uses your existing coding agent and its authentication. Both need the ocr CLI and Git 2.41+. Delegation supplies the review workflow but does not run the full native pipeline, so compare results in the mode you intend to use. File coverage alone does not prove defect coverage.
New arrival
janestreet/agent_skills
Reading a convincing explanation is an excellent way to feel briefly qualified. Jane Street's first published skill, socratic-tutor, asks questions instead: diagnose what you know, map the gaps, then work through them one at a time. Created 3 November, the repository reached 9,963 stars, up 5,698 (+133.6%) over the week.
Pick a topic you keep looking up, such as database isolation levels. Paste the skill into a chatbot or copy it into a compatible agent, then ask to be taught. Its instructions call for independent checks and later recall before raising your level. The useful test is whether you can solve a fresh example tomorrow.
This is a prompt-driven teaching method, with progress judged by the model. Its level map is not an independently validated assessment. You can ask it to save state for another session; it does not supply a separate learning-record service. Jane Street says it exports skills used internally and is not seeking changes to the published skill text, so keep your own adaptations.
Quiet but still loved
HandsOnLLM/Hands-On-Large-Language-Models
When a RAG answer is wrong, start with the passages it retrieved. This repository holds the companion notebooks for Jay Alammar and Maarten Grootendorst's Hands-On Large Language Models, covering embeddings, search, generation and fine-tuning. It had 30,203 stars and 208 days since the last commit in the 18 November snapshot, gaining 86 stars (+0.3%) during the week.
Start with Chapter 8's search notebook. It compares embedding search with keyword search and reranking before adding generation. Try three questions against a short document you know, then inspect the retrieved passages. You may learn more from the wrong result than from another polished answer.
These are teaching examples from a 2024 book. The notebook pins older package versions and uses Cohere for some examples, so check current API compatibility and usage costs before running those cells. A local-model example is included, but the setup recommends a GPU. The enduring value is understanding the retrieval steps; it is not a maintained production RAG stack.
Which skill would you test first, and what task would show whether it helps? If this reached your inbox, reply with one example.
Selection: three risers from distinct end-date star bands (500-9,999; 10,000-49,999; 50,000+), one project created within 90 days and one with at least 5,000 stars and 180 days without a commit. Growth compares paired observations from completed imports on 11 and 18 November; risers need at least 500 stars at both dates. Previous archive features, duplicate identities and permanent editorial exclusions are removed. Editorial choices weigh practical value alongside growth. Stars measure attention, not quality. Read this edition online.
LLMRepos tracks LLM and agent repositories using daily GitHub snapshots.
Get the next edition
Five open-source AI repos worth watching, every Tuesday.
Free weekly digest. Double opt-in, unsubscribe anytime.
