GitHub AI Weekly · Issue 05 (Sep 12, 2026) · AI Beyond

Top 10 projects

1. 01

   ayghri/i-have-adhd

   AI codingPython

   This week+13,164
   

   An agent skill that makes coding assistants lead with action, number steps, suppress tangents and remove long preambles and closers, with packages for Claude Code, Codex, Cursor, OpenCode and other hosts.

   Stars

   42,134

   Forks

   2,383

   Open issues

   60

   Last commit

   2026-09-10

   Its first appearance in this series goes straight to the top by addressing a broad irritation rather than a diagnosis: coding agents burying the next action under setup. Sep 10 commits also repaired installation and loading paths for Zed, AstronClaw and OpenCode.

   People who routinely hunt through long replies for the next step, or teams that want agents to share an action-first, short-list style, can trial it in one project and tune the rules around their own work.

   It changes presentation, not the correctness of analysis, code or facts. Its default suppression of side paths, list caps and missing recaps can remove necessary context in architecture decisions, security reviews and teaching, so easier reading is not a quality check.

2. 02

   DietrichGebert/ponytail

   AI codingJavaScript

   This week+11,054
   

   A coding-agent skill that resists over-implementation by asking whether a feature should exist, then preferring code already in the project, standard libraries, native platform capabilities and installed dependencies without cutting validation, security or accessibility.

   Stars

   135,918

   Forks

   7,285

   Open issues

   256

   Last commit

   2026-09-07

   It turns “write less code” from a prompt into an installable decision ladder and publishes a corrected agent experiment with four runs across 12 tasks in one open-source project instead of leaning on an earlier single-turn headline range. This week's commits themselves are mostly documentation and branding.

   Teams whose agents regularly add dependencies, wrap native controls or casually invent frameworks can use it as a YAGNI and reuse check before implementation, then accept the result through their existing tests.

   The reported average reductions in lines, cost and time come from one FastAPI plus React repository, one model and a small task set, so they do not generalize to every codebase. Always-on activation also uses trusted Node.js hooks, and future extensibility cannot be judged by line count alone.

   Deep dives

   Many prompts that ask an agent to be concise only remove explanation while leaving an overbuilt implementation intact. Ponytail's useful distinction is ordering: ask whether the feature needs to exist, look for an implementation already in the project, then try the standard library, a native platform feature and an installed dependency before writing the smallest new code. It aims to remove solution branches, not compress necessary error handling into one line.

   The repository also corrected its evidence. An earlier single-turn generation test reported 80% to 94% less code, but the maintainer acknowledges that the no-skill baseline contained more prose and alternatives. The published agentic test now runs 12 feature tasks in the same real repository four times each and reports average reductions of roughly 54% in lines, 20% in cost and 27% in time. The sample remains small, but it is more auditable than one before-and-after date picker.

   The safest use is not to let “lazy” replace engineering judgment, but to make it an adversarial check before implementation: why can this not be reused, why is a new dependency necessary, why is the platform primitive insufficient? Security boundaries, data-loss handling, accessibility and tests still deserve independent acceptance. If a team's real problem is unstable requirements or a confused domain model, deleting a few dozen lines will not repair it.

3. 03

   tt-a1i/archify

   AI codingJavaScript

   This week+11,006
   

   An architecture-diagram system for coding agents: the agent writes typed JSON intermediate data, then deterministic rendering and validation produce architecture, workflow, sequence, data-flow and lifecycle diagrams.

   Stars

   58,705

   Forks

   3,829

   Open issues

   154

   Last commit

   2026-09-12

   Its third consecutive week on the board comes with delivery work rather than momentum alone: support for evidence from Gitee and local-only repositories, embedded viewer fonts for a true single file, a modularized viewer source and reproducible Windows archives.

   Development teams that need Codex, Claude Code, Cursor or OpenCode to hand over a reviewable single-file system map, and want before-and-after architecture comparisons, can test the evidence chain on a real repository.

   The map still starts from agent-authored intermediate data, does not inspect live infrastructure and does not infer runtime impact or merge safety. It focuses on five system-diagram families rather than a general canvas, while 154 open issues make version pinning and regression checks prudent.

4. 04

   mattpocock/skills

   AI codingShell

   This week+10,571
   

   A composable set of engineering agent skills for requirements grilling, shared domain language, specs and ticket decomposition, TDD, difficult diagnosis, code review and architecture improvement, installable selectively or as a Claude Code plugin.

   Stars

   259,880

   Forks

   21,916

   Open issues

   490

   Last commit

   2026-09-04

   The 10,571 weekly stars look like concentrated circulation around its existing “grill, specify, implement, review” toolkit rather than a release: the main branch has no new commit after Sep 4, so the chart position is honest popularity inertia, not a feature event.

   Teams that do not want one framework to own their process, but do want a single discipline such as TDD, diagnosis or ticket decomposition inside an existing agent, can select individual skills and adapt them to local terminology and tools.

   Initial setup asks for an issue system, labels and documentation locations. Installing both the managed plugin and editable copies duplicates every skill, 490 open issues argue for taking only what is needed, and written discipline never replaces tests, review or domain expertise.

5. 05

   bilawalsidhu/gods-eye-view

   AI agentsJavaScript

   This week+8,916
   

   A browser-based 3D spatial-data interface that overlays public aircraft, vessel, satellite, earthquake, traffic and camera signals on a globe, with an optional realtime voice agent for controlling layers and viewpoints.

   Stars

   27,329

   Forks

   5,587

   Open issues

   200

   Last commit

   2026-09-12

   Its second consecutive week has no launch-sized feature. Instead, the project split place search from routing and aircraft from vessel providers, while making provider fixtures independent of calendar time — boundary cleanup for adding and maintaining live feeds.

   People building GIS, public-data, OSINT visualization or spatial-interaction prototypes can study how heterogeneous live signals, a Cesium scene and voice tool calls cooperate in one interface.

   Traffic includes simulated movement along real roads, camera orientation and rocket paths are coarse estimates, and voice needs a separate OpenAI API. GitHub still identifies no standard license, and 200 open issues mean the cinematic interface should not be mistaken for uniformly precise data.

6. 06

   affaan-m/ECC

   AI codingJavaScript

   This week+8,714
   

   An agent workflow repository spanning Claude Code, Codex, OpenCode and Cursor; its current README lists 68 specialized subagents, 292 skills, 94 commands, plus memory, rules, hooks, installation repair and orchestration scripts.

   Stars

   256,593

   Forks

   38,397

   Open issues

   186

   Last commit

   2026-09-12

   Late-August release 2.2.1 brought guided setup, doctor and repair, a memory vault and cross-platform adapters together. Main added Rails patterns and repaired Windows linter paths, ESLint 9, hook argument parsing and explicit consent this week; a search catalog entry added and then removed the same day also shows how quickly the ecosystem surface changes.

   Teams already running coding agents across repositories and hosts, and wanting one system for planning, TDD, review, memory and subagent roles, can start with a minimal profile and dry-run to see whether it fills an actual process gap.

   Its breadth is itself a cost: skills, commands, MCP, rules and shell hooks expand both context and execution surface, while some adapters lack Claude Code feature parity. Pin an exact release, inspect every planned file and hook, and enable only required modules instead of handing all 292 skills to a production repository.

   Deep dives

   ECC is not a point skill so much as an operating layer for an agent harness. It places planning, TDD, code review, security checks, worktree orchestration, session memory and model routing in one tree, then packages adapters for different hosts. That can reduce duplicated configuration for a team already maintaining Claude Code, Codex and OpenCode. For someone who only needs a diagnosis loop, it may weigh more than the problem.

   One sound boundary is that ECC does not bind the model provider. Workflows run through the host's existing configuration and can sit behind an official API, a compatible gateway or a self-hosted endpoint; the installer manages project-local skills, rules, MCP and hooks. Provider independence does not imply behavioral parity, however, and smaller models still need regression tests on real orchestration tasks.

   The adoption question is trust. A document skill can change a prompt, a hook can execute shell on tool events, and an MCP server can touch an external system. ECC is adding explicit hook consent, doctor, repair and owned uninstall state, which is the right direction. Users should still begin with the smallest profile, inspect the dry-run, pin releases and treat third-party catalog entries as supply-chain dependencies rather than treating stars as a security audit.

7. 07

   cathrynlavery/diagram-design

   AI codingHTML

   This week+7,776
   

   An editorial-diagram skill for Claude Code, Codex and other agents that turns prompts or draw.io, Mermaid and Excalidraw sources into self-contained HTML, SVG and PNG. The repository description says 38 types while the current README lists 39.

   Stars

   38,543

   Forks

   2,443

   Open issues

   41

   Last commit

   2026-09-10

   Back after Issues 01 and 02, it shipped several increments through 2.6.22 this week: GitHub Copilot marketplace installation, a doctor fallback when python3 cannot report its version, self-check rejection of external CSS assets, and plugin descriptions shortened to fit Cowork's limit.

   People starting with draw.io, Mermaid or Excalidraw and needing separate detail, size and audience choices for documents, slides and social cards can use its importer and fidelity ledger to track what changed.

   The agent still chooses hierarchy, deletion and emphasis and cannot verify system relationships. Simplified imports merge or drop nodes even though they disclose the ledger, so publication needs comparison with the source, while external fonts and CSS can be rejected by the single-file and safety constraints.

8. 08

   heygen-com/hyperframes

   MultimodalTypeScript

   This week+5,100
   

   An open framework that renders HTML, CSS, media and seekable animation into MP4, with a CLI, agent skills, browser preview and headless-Chrome plus FFmpeg paths for local, Docker and Lambda rendering.

   Stars

   49,051

   Forks

   4,483

   Open issues

   118

   Last commit

   2026-09-12

   It turns the web language coding agents already know into a video-production interface. This week's work reached v0.8.35 and repaired macOS multi-worker dispatch, pinned audio-carving behavior and clarified search-tier installation, matching popularity with dense maintenance.

   Developers who need repeatable product demos, data animation, course clips or regression-tested video templates, and already know HTML and CSS, can make one ten-second piece with the Core Skills before committing to a pipeline.

   This is not a no-code editor: users still own fonts, media rights, browser rendering, FFmpeg, audio mixing and compute. All 20 public skills need not be installed, development golden videos use about 240 MB of Git LFS, and changing external media breaks the premise of identical input.

   Deep dives

   Agents have lacked a good intermediate layer for video. Directly manipulating a traditional timeline is brittle, while generating a pile of FFmpeg arguments is hard to maintain. HyperFrames chooses HTML: the picture is a DOM, time and tracks live in data attributes, and motion can use GSAP, CSS, Lottie, Three.js or a custom adapter. The agent produces a human-readable web file before a renderer seeks and captures every frame.

   The useful claim is not simply that the web can make video, but that the handoff can be inspected. A browser previews it, the CLI lints it, CI can compare known frames, and the same composition can run locally, in Docker or on Lambda. Its distinction from Remotion is the authoring bet: plain HTML without a React requirement or build step is easier for a human and an agent to exchange.

   Determinism still has boundaries. Frames reproduce only when input media, fonts, browser versions, external assets and animation clocks are fixed; a web font or changing URL reintroduces drift. HyperFrames is a sensible way to automate an established visual system, not a license to make one prompt responsible for creative direction, asset rights and final quality control. Apache-2.0 covers the code, not the commercial rights to the media placed inside it.

9. 09

   microsoft/markitdown

   AI infrastructurePython

   This week+4,650
   

   A Python conversion library for LLM and text-analysis pipelines that turns PDFs, Office files, images, audio, HTML, archives, EPUB and selected web content into Markdown while preserving useful headings, lists, tables and links.

   Stars

   182,710

   Forks

   13,432

   Open issues

   634

   Last commit

   2026-09-11

   This week's work targeted real document edges: removing quadratic blank-row handling in CSV, avoiding UTF-8 splits during charset detection, repairing Japanese Outlook MSG decoding and DOCX style namespaces, and falling back to HTML when YouTube extraction finds no video content.

   Teams preparing common Markdown inputs for RAG, agents or text analysis can install only the format extras they need and first benchmark structural retention on their own Office, PDF and mail samples.

   It explicitly is not a high-fidelity document converter for human consumption. The permissive convert() can read local files, remote URIs and streams with current-process privileges, so servers must restrict paths, schemes and private destinations; OCR, transcription and Azure Content Understanding also add dependencies, model calls or billing.

10. 10

    THU-MAIC/OpenMAIC

    AI agentsTypeScript

    This week+4,509
    

    An open multi-agent classroom generator that can plan and revise lessons, quizzes, interactives, images, narration and video exports from prompts, documents, audio, video and web material.

    Stars

    35,934

    Forks

    5,705

    Open issues

    252

    Last commit

    2026-09-11

    Its second consecutive week shifts from “can generate a course” to making import and export survive: fixing PPTX hangs outside browsers, converting Equation.3 formulas with degradation telemetry, checking the render queue before export, reusing compiled ZIPs on retry, and persisting generated media in the asset pool.

    Teachers, training teams and education-product developers with existing source material can test imports and answer quality in the hosted demo before deciding whether the editable classroom prototype merits self-hosting.

    Facts, answer keys and media rights still need page-by-page human review. Self-hosting needs a model provider or local Lemonade, while persistence, extraction and video rendering add Postgres, FFmpeg and queue services, and the example authentication is not ready for a public production deployment.