Selection rule: ranked by stars gained in the last 24 hours. Repos with large cumulative counts but flattening momentum are excluded. A repo featured yesterday returns only if today’s gain held or grew.
freestylefly/awesome-gpt-image-2 — +3,600 ⭐ today (21.8k total, 2.2k forks)
https://github.com/freestylefly/awesome-gpt-image-2
This is a prompt engine and template library built on the premise that image-generation prompts for GPT-Image2 should be treated as code rather than prose. Its core move is to reverse-engineer roughly 530 community results and decompose them into atomic parts — subject, lighting, material, layout — so they compose like function arguments. From that decomposition it ships more than 20 industrial-grade templates plus an Agent Skills package (agents/skills/gpt-image-2-style-library), so an agent or an automation pipeline can call a prompt instead of a human rewriting one from scratch each time. A companion site at gpt-image2.canghe.ai lets you filter the gallery by style or scenario, copy full prompts, and run a generation test after signing in. It hit #1 on GitHub Trending overall on August 23, collapsed 55% on August 26, and today posted its highest daily gain ever at +3,600 — this is a re-ignition, not an ember.
Practical angle: Instead of leaving product imagery and thumbnails to whoever has the best eye, register the templates as a skill and batch-produce assets with brand tone locked in.
Tags: #PromptAsCode #ImageGeneration #PromptEngineering #AgentSkills #TemplateLibrary #StyleTransfer
tt-a1i/archify — +1,500 ⭐ today (day 2 in a row, 1.6× yesterday’s +954; 22.8k total, 1.4k forks)
https://github.com/tt-a1i/archify
Archify is an agent skill that takes a plain-English description of a system or process and returns a verifiable technical diagram. It covers five diagram types — architecture, workflow, sequence, data-flow and lifecycle — and the output is a single self-contained HTML file with no external dependencies. That file is not just a picture: it carries its own dark/light theme toggle, motion, clipboard copy, and maximum-resolution export to PNG, JPEG, WebP and SVG, so it can be dropped straight into documentation or shared as-is. It installs as a skill across the major agent runtimes — Claude Code, Codex CLI, OpenCode, Cursor, Raven and DeepSeek Harness (for example, dsh plugin --profile web add @tt-a1i/archify-dsh@0.1.0) — and the signature use is handing it an entire repository and asking it to map the runtime architecture. It first appeared on August 9, went quiet, then hit #1 on GitHub Trending overall for the first time on August 26 and has grown its daily gain for two consecutive days since.
Practical angle: Generate the system diagrams you currently hand-draw for every client onboarding and handover directly from the repo, so the documentation updates when the code does.
Tags: #DiagramGeneration #ArchitectureVisualization #SelfContainedHTML #SequenceDiagram #AgentSkills #DocsAsCode
andrewyng/openworker — +772 ⭐ today (16.4k total, 2.3k forks)
https://github.com/andrewyng/openworker
Released by Andrew Ng in late July, OpenWorker is a local-first desktop AI coworker built around one claim: it should hand back finished work, not a chat transcript. The deliverables it targets are a polished document, a Slack reply with the numbers already in it, a code review with vulnerability fixes attached, a triaged inbox. Architecturally it is a Tauri 2 + React shell wrapping a local Python FastAPI agent server, and that server sits on aisuite, which unifies multiple model providers behind one interface. Nothing is locked to a single model — bring your own API key and run GPT-5.6 Sol, Claude Fable, Gemini 3.6, open-weight models like Kimi, GLM or DeepSeek, or Ollama entirely locally — and apart from the LLM provider and connectors you explicitly choose, your data never leaves the machine. It ships 40 connectors to tools teams already use, and every action an agent takes is governed by policy and written to a log. MIT licensed.
Practical angle: For organizations that cannot send data outside their perimeter, the local execution plus audit log combination is a far easier security review to pass than a cloud SaaS agent.
Tags: #DesktopAgent #LocalFirst #Tauri #BYOK #ToolConnectors #AuditLogging
JetBrains/go-modern-guidelines — +617 ⭐ today (1.7k total, 46 forks)
https://github.com/JetBrains/go-modern-guidelines
An official JetBrains skill set aimed squarely at the problem of AI coding agents writing outdated Go. The diagnosis is precise: older syntax dominates training data, so absent explicit instruction a model reaches for whatever is statistically most common rather than what is current. The fix is a set of explicit rules covering language features and standard-library additions from Go 1.0 through Go 1.27, filtered against the version declared in your project so guidance for newer releases is excluded outright. The interface is split into two levels so agents pull only what they need:
list— short, scannable guidelinesexplain— detailed examples, fetched only on demand
It works out of the box with Junie, Claude Code, Codex and Cursor, and with other agents via skills.sh. It needs only the Go toolchain, runs from a local cache, and does not modify your project. Published on the JetBrains Go blog on August 24, it spread through Chinese-speaking developer circles on X and entered the daily top 15 for the first time today — with no trending badge history at all, which marks it as a genuine new entrant rather than a returning name.
Practical angle: A Go microservices team that adds this one skill will stop seeing the “nobody writes it that way anymore” review comment, because it gets caught at generation time instead.
Tags: #Golang #VersionAwareGuidelines #StdlibModernization #AgentSkills #CodeReviewAutomation
pathwaycom/arc-task-gen — +611 ⭐ today (second day rising; 7.7k total, 46 forks)
https://github.com/pathwaycom/arc-task-gen
A generator that produces an unlimited supply of original tasks statistically distribution-matched to the public ARC-AGI-1 evaluation set. The problem it addresses is benchmark contamination: because ARC-AGI-1 is public, a high score cannot distinguish genuine few-shot rule induction from prior familiarity picked up during training. This tool builds a private evaluation set with the same properties as the public one but composed of problems no model has ever seen, so you can place public-benchmark performance next to fresh-task performance and read the gap as a contamination estimate. Output is a standard-format tasks.json, compatible with existing ARC evaluation harnesses, which makes adoption close to free. It is a small project — Python, MIT, one contributor, 46 forks — which is exactly why two consecutive days of rising gains is a meaningful signal.
Practical angle: When a vendor’s published benchmark scores are hard to trust during model procurement, re-measure on unreleased tasks from the same distribution and check for overfitting yourself.
Tags: #ARCAGI #BenchmarkContamination #SyntheticTaskGeneration #ReasoningEvaluation #EvalHarness
Notable Exclusions Today
- bilawalsidhu/gods-eye-view — +3.0k yesterday to +1.7k today (-43%). Still a large number, but the momentum broke.
- FlashML-org/FreeToken — grew its daily gain four days running, then dropped out of the daily top 25 entirely. Treated as a collapse.
- AgriciDaniel/claude-obsidian — +825 to +715 (-13%), only 34% of its weekly gain. Third-day decay.
- calesthio/OpenMontage (+970) and diegosouzapw/OmniRoute (+1.0k) — 52.3k and 56.2k cumulative stars with multiple GitHub Trending #1 finishes dating to June. Excluded as rank-holders.
- MadsLorentzen/ai-job-search, deepseek-ai/deepseek-harness, DietrichGebert/ponytail — permanently excluded.
Daily AI Brief 
Leave a comment