GitHub Trending Top 10: design-taste skills take ranks 1 and 2, a single CLAUDE.md holds 227,997 stars with no licence, and T3 Code ships 39 releases in a week (Oct. 25–Nov. 1, 2026)

Rank 1 and rank 2 this week sell the same fix. Both stop a coding agent from building the page coding agents always build: cream background, thin serif, purple-to-blue gradient, cards nested in cards. Nutlope/hallmark took 11,566 stars in seven days; pbakaus/impeccable took 12,017. 1

Rank 9 sits at the other end of the same idea. Its README calls it a single CLAUDE.md file, and it holds 227,997 stars with no licence, no release ever, and no commit since April 20. 2

Rank 6 holds the largest total on the board at 310,637. It committed nothing and released nothing inside the window. 3

Attention and engineering ran on separate clocks again, and this week the gap is easiest to see in the star column itself. The ranking below is GitHub's own weekly board, captured November 1, 2026 at 8:00 p.m. ET. GitHub orders that board by its own ranking, and this week the order does not follow stars gained: rank 1 took fewer than rank 2. 1

"Stars this week" is the gain GitHub attributes to the trailing seven days; "stars captured" is the running total at that moment. Those two numbers measure attention. Whether a repository is ready to adopt is a separate question, and each entry closes on the boundary that decides the star.

Current signal: Rank 1 with 11,566 stars this week and 44,882 captured. Twelve commits, no release in its life, MIT, created April 27, 2026, 72 open issues. 4 It led this chart for July 12–19 at 13,374, when it shipped twenty themes and fifty-seven anti-slop gates. It now ships fifty themes and credits 755 photographs.

Problem solved: An agent asked for a landing page returns the same landing page every time, because that is what the average of its training data looks like.

Implementation: A skill directory installed with npx skills add nutlope/hallmark -g. A build reads one theme file of at most 8 KB, draws the page's anatomy from the job the page does, and renders the hero alone before the rest exists. scripts/sloplint.mjs --render then checks the result at eight widths. Browser checks need puppeteer-core and Chrome; the static checks run without them. 5

Two decisions carry the quality. The photo library lists photographs hosted on Wikimedia Commons instead of Unsplash or Pexels, because those terms bar AI tools from picking their images. sloplint --kept also fails a redesign if a link, field, handler or heading went missing.

Differentiation: A component kit gives an agent the same components everyone else has. Hallmark gives it fifty materials and a checker that argues with the output.

Verdict: Star it, and read the sloplint rules before the themes, because the rules are the transferable part. Boundaries: 72 open issues, no version to pin, and the photo credits live on the page it generates.

Current signal: Rank 2 with 12,017 stars this week, the largest gain on the board, and 99,333 captured. 51 commits and three releases on October 26: engine-v0.1.18, skill-v4.8.0 and cli-v4.8.0. Apache-2.0, created November 16, 2025, GitHub labels it Rust. 6 Third appearance here, after June 1–8 at 35,556 and September 27–October 4.

Problem solved: "Make it pop" tells an agent nothing. Someone wants to name a design move, point it at one component, and get a judgement back.

Implementation: npx impeccable install writes the skill into whichever of nineteen harnesses it detects, from Claude Code to Veto. For Claude Code it is also a plugin. A self-contained engine binary downloads once into ~/.impeccable/bin/. 7

The substance is 24 named commands and a detector with 62 deterministic rules and no model call. critique scores a screen against usability heuristics; audit ranks findings P0 to P3; live opens the dev server and hot-swaps variants in place.

Differentiation: Hallmark generates a page from a theme. Impeccable reviews pages that already exist, in the agent, in CI as npx impeccable detect --github, and in a Chrome extension over any live site. The two are complements, which is why they occupy the top two slots.

Verdict: Star it for the detector, which runs without an agent at all. Boundaries: the newest release added Devin and Replit Agent support plus GitHub Actions annotations, so the surface keeps widening. 8 Live mode edits a local checkout and runs a script from your package.json.

Current signal: Rank 3 with 5,034 stars this week and 25,542 captured. No commit inside the window and no release inside it; the newest tag is v2.5.0, published October 22. MIT, created February 7, 2026, 78 open issues. 9

Problem solved: A site wants the audit a consultant would run — technical health, content quality, schema, AI-search citation odds — and wants it on Monday, not next quarter.

Implementation: A Claude Code plugin of 26 sub-skills and 19 specialist agents behind one /seo command, with up to 17 agents working a site at once. The README puts a full audit at 10 to 15 minutes. 10

The agents are split by discipline rather than by page. seo-schema writes and validates structured data. seo-geo scores how likely AI answers are to cite the page. seo-drift watches for changes between audits. Nine optional extensions pull live data from Google APIs and commercial SEO providers.

Differentiation: An SEO crawler reports a list of issues. This bundles the consultant's roles as separate agents with separate models, and returns one plan ranked by what to fix first.

Verdict: Star it to read how the 19 roles divide the work, which is the reusable pattern. Boundaries: the headline results — 10.8K clicks and a 99 PageSpeed score — are the project's own site auditing itself. The nine extensions are where the live data comes from, and those cost money. 11

Current signal: Rank 4 with 5,852 stars this week and 109,292 captured. 103 commits and seven releases between October 26 and November 1, v15.3.0 through v15.8.1. Apache-2.0, created August 31, 2025, 173 open issues. 12 It last appeared here for October 4–11, running at v13.35.0.

Problem solved: An agent starts every session from zero. The decisions behind last week's code live in a transcript nobody reads again.

Implementation: An npm package on Node 20 or later. Lifecycle hooks call claude-mem hook claude-code <event> at session start, prompt submit, tool use and session end. A Bun-managed worker holds SQLite for sessions, observations and summaries, plus a Chroma vector database for hybrid keyword and semantic search. 13

The week's releases are the story. All seven build out "My Work", the live-requests surface the worker serves: deliverables on a card, focus areas, one store for the interface's state, a cloud viewer host. v15.8.1 is a repository cleanup that moves the root from 60 entries to 36, and says outright that it changes nothing about how claude-mem runs. 14

Differentiation: Claude Code's own memory belongs to one tool on one machine. claude-mem puts a compression pass and a searchable store between sessions, across harnesses.

Verdict: Star it, and note that the engine has held still while the interface took seven releases. Boundaries: two major versions in three weeks, so pin one and read the migration notes. 15 The value depends on a model call per session, on your key.

Current signal: Rank 5 with 10,069 stars this week and 56,382 captured. Zero commits inside the window; the last push to main was September 14. No release ever. MIT, created June 18, 2026, and the backlog grew from 96 to 110 open issues in a week. 16 It led this chart for October 18–25 at 15,072 and 46,917. It was rank 4 for September 20–27 at 4,805 and 22,338. In six weeks it has taken 34,000 stars without a commit since mid-September.

Problem solved: One agent asked to "find vulnerabilities" returns claims nobody can check, repeat, or rank.

Implementation: Prompt files plus two zero-dependency Node validators, run by a coding agent in six phases. Reconnaissance writes architecture.md and coverage-ledger.json; hunters work one ledger unit at a time; a fresh verifier tries to disprove each candidate. Survivors become records in findings.json, checked against report-schema.json. 17

Verdicts stay distinct: confirmed carries a complete source trace, needs_validation an unresolved fact and no severity, rejected a disproved candidate. Repeat runs are additive and aim at earlier gaps.

Differentiation: A scanner matches patterns and reports hits. This writes down what it did not cover alongside what it found.

Verdict: Star it for the ledger, the verdict states and the validators, which transfer to any review an agent runs. Boundaries: the repository is a mid-September snapshot, and a confirmed finding is evidence a reader still has to reproduce. Read the Cloudflare post it credits as its origin first. 18

Current signal: Rank 6 with 5,896 stars this week and 310,637 captured, the largest total on the board. No commit inside the window and no release inside it; v7.0.1 landed October 21. MIT, created October 9, 2025, 377 open issues, 27,649 forks. 3 It appeared here for May 18–25 at 205,000 stars, so it added 105,000 in five and a half months while barely moving.

Problem solved: An agent takes a request and starts typing. Nobody asked what the request meant, so the code answers the wrong question.

Implementation: A methodology expressed as skills, installed per harness through sixteen separate paths — the official Claude plugin marketplace, Antigravity's agy plugin install, or a marketplace per tool. The workflow is spec first. Tease out what the user actually wants, show the spec in readable chunks, and write a plan tight enough for a junior engineer to follow. Then run subagent-driven development against that plan. 19

The skills trigger themselves, so there is nothing to invoke. The README also lists commercial support through an email address at primeradiant.com.

Differentiation: A prompt pack tells the agent how to behave in one project. Superpowers ships a full development method and keeps it identical across sixteen harnesses, which is what 310,637 stars are voting for.

Verdict: Star it as the reference implementation of a method-as-skills, and skim the "when something goes wrong" section before trusting an unattended run. Boundaries: a repository this large and this quiet is in maintenance, and v7.0.1 is seven weeks old. 20

Current signal: Rank 7 with 5,760 stars this week against 6,971 captured. That gain is 83 percent of everything the repository has ever collected, the most lopsided on the board. 93 commits and nine releases, from v0.3.0-nightly.7 to v0.3.1 on November 1. Apache-2.0, created July 8, 2026, 324 open issues against under 7,000 stars. 21 First appearance here.

Problem solved: An agent working in your checkout touches your checkout. Run three and the branches, terminals and edits collide.

Implementation: A desktop IDE built around the pi coding agent. Every task gets a git worktree with its own branch, terminal and editor tabs, and finished workspaces settle onto a per-project shelf when their pull request merges or closes. 22

Two pieces are worth lifting. Living specs: a graph of SPEC.md files the agent reads, searches and maintains through dedicated tools, so intent outlives the session. And the agent asks structured multiple-choice questions in the chat instead of guessing. Both ship as standalone pi packages for anyone not using the IDE.

Differentiation: Orca and T3 Code supervise agents you already installed. ThinkRail replaces the editor around them, and carries JetBrains' incubator badge.

Verdict: Star it if worktrees have become your unmanaged mess. Boundaries: 324 open issues on 6,971 stars is a backlog out of proportion to the audience, and v0.3.1 added Russian localisation one day after v0.3.0. 2324 The README also asks for a star.

Current signal: Rank 8 with 6,926 stars this week and 38,016 captured. 1,346 commits and 39 releases in seven days, the busiest repository on the board by a wide margin. MIT, created February 8, 2026, 3,922 open issues and pull requests, and GitHub shows no description at all. 25 Second appearance, after September 27–October 4.

Problem solved: Someone runs four agents on a desktop machine and wants to check on them from a phone. Right now the only way to see what they are doing is to walk back to the desk.

Implementation: T3 Code calls itself an agent harness control surface. A server on your machine drives Claude Code, Codex, Cursor, Grok Build, OpenCode, Antigravity and fx. The clients are an iOS app, an Android app, a web app and an Electron desktop app. curl -fsSL https://t3.codes/install.sh | sh starts the server; t3 service install keeps it running. 26

The cadence is nightly. v0.1.6 on October 31 folded in 521 changes from 137 contributors, and the newest nightly added organization-wide policies for IT teams. 2728

Differentiation: Claude Desktop and Codex's own app each control their vendor's agent. T3 Code points one interface at seven of them, which is the part worth copying.

Verdict: Star it for the remote-control design. Boundaries: the README says the project is very early, expects bugs, and is mostly not accepting contributions. 39 releases in a week means you pin a tag or you follow nightlies.

Current signal: Rank 9 with 4,459 stars this week and 227,997 captured. No commit since April 20, no release ever, no licence anywhere, and no primary language because the repository is Markdown. Created January 27, 2026, 138 open issues, 22,742 forks. 2 First appearance here.

Problem solved: An agent invents an assumption, runs with it, overbuilds the result, and edits code it never understood. That is the pattern Andrej Karpathy described in a post the README quotes at length. 29

Implementation: Four principles in one file: think before coding, simplicity first, surgical changes, goal-driven execution. Each principle carries a test. Surgical changes: every changed line should trace to the user's request. Simplicity first: if 200 lines could be 50, rewrite it. 30

It installs as a Claude Code plugin or as a per-project CLAUDE.md.

Differentiation: Superpowers ships a process with skills that fire on their own. This is 200 lines of prose you can read in four minutes and paste by hand, which is a large part of why it collected 227,997 stars in nine months.

Verdict: Star it to read the four tests, which are the useful part. Boundaries decide the adoption. There is no licence, so the default is all rights reserved. The install instructions still name forrestchang/andrej-karpathy-skills, while the repository now lives under multica-ai. 30 And nothing has changed in six months.

Current signal: Rank 10 with 1,607 stars this week, the smallest gain on the board, and 10,080 captured. No commit inside the window, no release ever, MIT, created July 10, 2026, 11 open issues. 31 First appearance here.

Problem solved: A developer who cares about craft can see what is wrong with an interface: the type scale, the focus ring, the empty state. Naming it precisely enough for an agent to fix is the harder half.

Implementation: Fourteen skills, each a SKILL.md, installed with npx skills add jakubkrehel/skills or as a Claude Code plugin. better-interface reviews a screen across accessibility, layout, writing, typography, colour and polish, then returns one ranked report. break renders a component under every scenario that can reach it. state-machine renders every state on a throwaway page with a switcher. 32

Differentiation: Impeccable automates its judgements with 62 deterministic rules. These are written positions with exact values — border radius, optical alignment, animation — that an agent applies.

Verdict: Star it as the smallest complete example of a design skill set that covers a whole discipline rather than one move. Boundaries: 1,607 stars in a week with no commit and no release means this arrived on attention alone. Five of the fourteen skills are user-invoked, so read which ones fire on their own.

Weekly gain as a share of each repository's own total separates this week's ten more sharply than rank does. ThinkRail took 83 percent of its lifetime stars in seven days. Hallmark took 26 percent, Claude SEO 20, T3 Code 18, the security-audit skill 18, jakubkrehel/skills 16, Impeccable 12. Superpowers and the Karpathy file each took 2. Those last two hold 538,634 stars between them and barely moved.

Release behaviour split four and six. Impeccable, claude-mem, ThinkRail and T3 Code published 3, 7, 9 and 39 releases inside the window. The other six published none. Four of those six have never published a release at all: Hallmark, the security-audit skill, the Karpathy file and jakubkrehel/skills. So six of the ten have no release to pin from this week, and one of the four that do shipped three of them on a single day.

Look at what GitHub puts in the language column. CSS for Hallmark, Markdown for jakubkrehel/skills, Shell for Superpowers, nothing at all for the Karpathy file. Four of ten are labelled by a file format because the product is prose, and four of ten ship no release because prose has no build.

Where the ten sit is the more useful split. Eight are files or glue for an agent somebody else built: a skill directory, a CLAUDE.md, a detector binary, a plugin bundle, a memory hook, a coverage ledger. Two are applications that host agents — ThinkRail wraps the pi agent in an IDE, T3 Code puts one control surface over seven of them.

That is the week's real symmetry. Ranks 1 and 2 are two answers to one complaint about how agents design, and both won by constraining the model with files a person wrote. Meanwhile the two entries that did the most ordinary software engineering this week — 1,346 commits and 39 releases at rank 8, 93 commits and 9 releases at rank 7 — sit below them.

A star records that a repository moved someone. Whether it belongs in your toolchain turns on the licence, the version you can pin, and the server or key it asks you to run. Those three facts sit in the boundaries above, and no star count measures any of them.