GitHub Trending Weekly, July 22, 2026: Engineering Skills Lead, AI Tools Chase Better Outputs
Data window: July 15–22, 2026 (rolling seven days) Sources: GitHub Trending weekly and monthly, GitHub Search and REST APIs, HN Algolia
TL;DR: mattpocock/skills led the week with 10,651 new stars, showing that agent skills are becoming composable engineering methods rather than prompt collections. Hallmark, OpenCut, img2threejs, and Wardrobe share another theme: AI tools are being judged by the quality of the artifact they leave behind. The new-repo chart also contains high-risk finance and privacy tools, so ranking is not endorsement.
📈 Fastest Growing, Top 10
| # | Repository | Stars/week | Total | Language | Created |
|---|---|---|---|---|---|
| 1 | mattpocock/skills | +10,651 | 181,017 | Shell | 2026-02-03 |
| 2 🔁 | Nutlope/hallmark | +8,948 | 15,132 | CSS | 2026-04-27 |
| 3 🔁 | OpenCut-app/OpenCut | +8,341 | 77,181 | TypeScript | 2025-06-22 |
| 4 🔁 | stablyai/orca | +5,733 | 25,319 | TypeScript | 2026-03-17 |
| 5 | Shubhamsaboo/awesome-llm-apps | +5,385 | 125,937 | Python | 2024-04-29 |
| 6 | tirth8205/code-review-graph | +4,791 | 24,827 | Python | 2026-02-26 |
| 7 🔁 | iOfficeAI/OfficeCLI | +4,047 | 20,755 | C# | 2026-03-15 |
| 8 | HKUDS/Vibe-Trading | +3,679 | 26,220 | Python | 2026-04-01 |
| 9 | earendil-works/pi | +3,569 | 74,980 | TypeScript | 2025-08-09 |
| 10 | HKUDS/DeepTutor | +2,908 | 28,925 | Python | 2025-12-28 |
🔁 means the repository also appeared in the monthly snapshot. All counts come from the saved July 22 dataset.
🆕 Top New Repositories, Top 10
| # | Repository | Stars | Language | Created |
|---|---|---|---|---|
| 1 | Fei-Away/Codex-Dream-Skin | 11,637 | JavaScript | 2026-07-15 |
| 2 | lopopolo/harness-engineering | 1,839 | Python | 2026-07-18 |
| 3 | hoainho/img2threejs | 1,599 | Python | 2026-07-15 |
| 4 | tandpfun/wardrobe | 1,292 | JavaScript | 2026-07-16 |
| 5 | pablostanley/yoinks | 979 | TypeScript | 2026-07-16 |
| 6 | nethical6/conversation-steganography | 919 | Go | 2026-07-17 |
| 7 | v-modal/vmodal_sdk_flutter | 800 | Dart | 2026-07-16 |
| 8 | MIgHTy-alIeN/MEV-Arbitrage-Bot | 739 | Solidity | 2026-07-17 |
| 9 | Blueturboguy07/cue | 706 | JavaScript | 2026-07-15 |
| 10 | Blaizzy/nativ | 660 | Swift | 2026-07-20 |
These results use GitHub Search for repositories created from July 15 through July 22. A young project’s star count is not a security review.
Fastest Growing Highlights
#1 mattpocock/skills: encode engineering judgment
The repository favors small, editable, composable skills over one process that takes control of the entire development cycle. Setup teaches an agent the team’s issue tracker, labels, and documentation layout before task-specific skills handle triage, diagnosis, TDD, or architecture. Try one recurring failure mode first and measure rework instead of installing everything blindly.
#2 Hallmark: audit generic AI design before shipping
Hallmark exposes four modes, build, audit, redesign, and study, plus 20 themes and 57 author-defined anti-slop gates. Audit is the lowest-risk entry because it produces a punch list without editing code. Its taste claims are not independent accessibility or conversion benchmarks, so real-device testing still matters.
#3 OpenCut: promising, but the main branch is a rewrite
OpenCut targets a free editor across browser, desktop, and mobile, with a future Rust core, plugin system, headless rendering, and MCP. Its README explicitly says the current repository is being rewritten and recommends OpenCut Classic for use today. Do not mistake roadmap features for production features.
#4 Orca: parallel agents with worktree boundaries
Orca places Codex, Claude Code, OpenCode, or Pi side by side, each in its own Git worktree, and adds desktop, mobile, and remote monitoring. It solves coordination visibility, not correctness. Start with two independent tasks because more parallel agents still produce conflicts and bad answers faster.
#5 awesome-llm-apps: a discovery catalog, not one product
The repository contains runnable examples for agents, skills, RAG, voice, and multiple model providers. Each folder has its own dependencies, API keys, costs, and maturity. The parent repository’s star count cannot certify every example.
#6 code-review-graph: retrieve structure before asking for review
This local-first tool uses Tree-sitter and SQLite to map callers, dependents, tests, and blast radius through MCP and CLI. Its documentation also discloses limits: graph-derived recall can be circular, small edits may consume more context, and conservative impact analysis creates false positives. Test it on the same historical PR against your current process.
#7 OfficeCLI: let agents render and inspect Office files
OfficeCLI reads and edits Word, Excel, and PowerPoint without Office installed. HTML and PNG rendering closes a render-look-fix loop instead of making agents manipulate XML blindly. A directly relevant HN thread reached 215 points and 62 comments, with practical concern around complex-document fidelity. Use your own templates as regression fixtures.
#8 Vibe-Trading: research software, not investment proof
Vibe-Trading combines data analysis, backtesting, agents, connectors, and MCP. Its recent fixes around look-ahead bias, data gaps, and trade safety are useful evidence that financial automation has many failure modes. Begin with paper trading, isolate credentials, and independently validate data, slippage, costs, and every live-order path.
#9 Pi: extensible harness with explicit security boundaries
Pi separates a unified model API, agent runtime, coding CLI, and terminal UI. Its README says the base process inherits the user’s filesystem, process, network, and credential permissions; isolation requires an external micro-VM, container, or sandbox choice. That disclosure should shape deployment.
#10 DeepTutor: persistent learning rather than one-off answers
DeepTutor combines a document knowledge base, guided learning, research and solving modes, partner agents, and memory. It offers a self-hostable system and paper, but personalized output is not proof of learning effectiveness. Formal-course answers still need checks against teachers or authoritative material.
New Repository Highlights and Risk Checks
Codex-Dream-Skin uses local CDP injection to theme the unofficial Codex Desktop experience. It says it does not modify the official package, but the repository has no detected license and promotes an affiliate API relay. Read installers before use and treat redistribution rights as unclear.
harness-engineering is Ryan Lopopolo’s CC-BY-4.0 field guide for making intent, nonfunctional requirements, tools, constraints, and verification recoverable by an agent. Its directly relevant HN discussion had 78 points and 33 comments.
img2threejs reconstructs a reference object as procedural Three.js code, not photogrammetry. It is useful for editable low-poly assets and web prototypes, while photorealistic scanning is outside its stated approach.
Wardrobe imports clothing photos and uses gpt-image to create item photos and lookbooks. It stores app data locally but still sends images to an external model API, so test with nonsensitive images and understand cost and retention policies.
yoinks wraps yt-dlp and FFmpeg in a terminal interface. Technical capability does not grant copyright permission; download only material you own, are licensed to use, or the platform permits.
conversation-steganography is a dual-use proof of concept. The README makes strong privacy claims while admitting detection methods and known problems. Do not describe it as undetectable or guaranteed private.
vmodal_sdk_flutter wraps a remote multimodal-video search gateway. Its provenance and service-side handling are not independently established by this roundup. Verify credentials, pricing, retention, and indexing location before adoption.
MEV-Arbitrage-Bot asks beginners to deploy a contract and fund it through a third-party page, has no recognized license, and lacks an audit. This roundup does not recommend running it or connecting a wallet; irreversible financial loss is possible.
cue captures screen, microphone, and meeting audio while trying to stay hidden from screen sharing. Its README responsibly says hiding is best-effort and warns about exams, interviews, platform rules, and consent laws. Limit use to legitimate notes, practice, or accessibility, with participant consent.
nativ is a SwiftUI workspace for running MLX models on Apple silicon, exposing OpenAI- and Anthropic-compatible local endpoints to coding tools. It requires macOS 26+, and model downloads plus initial build dependencies still use the network.
Monthly Signal and Takeaways
Hallmark, OpenCut, Orca, and OfficeCLI appeared in both weekly and monthly snapshots. Each has a visible artifact or interface, making value easier to test than a benchmark claim. Weekly discovery and monthly persistence are useful signals, but neither replaces hands-on validation.
Three themes matter. First, skills are becoming executable engineering methods. Second, AI products are closing the last mile by rendering documents, producing editable geometry, or managing real artifacts. Third, new-repo charts need risk labels for money, credentials, recording, privacy, and copyright. If I tested three projects, I would start with Hallmark audit, a controlled code-review-graph comparison, and Wardrobe using disposable images. I would not fund a trading bot or hide an assistant in a meeting because it happens to rank.
Was this article helpful?
