Blog

Guides and notes on the agent-skill ecosystem.

GEO Assumes Google. Half the Web Isn't.

fire-your-seo-agency (8.1) covers Naver and NEO in Korean-English. Nothing in 765 skills covers Baidu or WeChat — multilingual SEO skills are a vacuum.

2026-10-28

Core Web Vitals Reports Tell You What. Not How.

performance-monitor (8.9) reads Core Web Vitals and technical-seo-checker (9.3) locates the structural cause. The bottleneck is translating a number into a fix — exactly what agents do well.

2026-10-28

Your README Is a Landing Page Now

agentic-seo-skill (7.8) audits GitHub repos — README structure, topics, social preview cards. AI coding assistants treat the README as the landing page when they pick libraries.

2026-10-28

The Three-Piece Stack for AI Search Visibility: llms.txt, JSON-LD, and Citability

AI search visibility isn't one thing — it's a three-link chain: discovery (llms.txt), parsing (JSON-LD), and citation (quotable writing). Miss any link and the answer engine is blind to you at a different point.

2026-10-28

One Author vs. the SEO Catalogues: Who You're Actually Buying From

One vendor ships 15+ SEO skills. One solo author ships 5, including the category's 9.3 peak. The six-dimension data says the real gap is in engineering — and which you should buy depends on your SEO maturity.

2026-10-28

Programmatic SEO Came Back. This Time It Has to Earn It.

page-play-builder (8.9) plus keyword clustering and content briefs make agent-era programmatic SEO possible. It only survives when every page is backed by real data.

2026-10-28

Security Isn't the Boring Dimension: Why It Veto-Kills SEO Skills

SEO skills crawl the web, call third-party APIs, and batch-edit your files. That's three attack surfaces, and in our 765-skill reviews, security is the one dimension where a low score actually kills the skill.

2026-10-28

Read-Only vs. Implementer: The Two Tribes of SEO Skills, Scored

We scored both tribes of SEO skills — the auditors that never touch your site and the implementers that rewrite it. The risk asymmetry is the whole story.

2026-10-28

11,302 Stars, No Verdict

The most-starred SEO repo on GitHub isn't in our directory, and the highest-scoring SEO skill we've reviewed has almost none. Stars are a marketing outcome; review scores are an engineering outcome.

2026-10-28

What a GEO Audit Score Actually Measures (and Why It Isn't an SEO Score)

GEO tools hand you a 0–100 score. Here's what's inside that number — robots.txt AI crawler rules, llms.txt, JSON-LD, citability — and why it's orthogonal to your SEO score.

2026-10-28

Generate vs. Edit — The Two Kinds of Video Skills, and Why You Probably Need Both

Half our video category makes footage from prompts; the other half reshapes footage that already exists. The two disciplines have opposite requirements — and using generation to solve an editing problem is the most expensive mistake in the category.

2026-10-23

Why Developer Tools Score 9.21 and Documents Score 8.26 — A Category Gap Autopsy

A full point of average difference between categories, and it isn't author skill — it's how verifiable the domain is. The categories that wrap stable, documented ground truth score highest, and the fix for mushy domains is writing checkable contracts.

2026-10-23

Skills That Touch Money — What 76 Finance Skills Taught Us About Letting Agents Near Your Wallet

From free market-data wrappers to Solana DEX execution and prediction-market orders: the finance spectrum, the confirmation-gate engineering that separates the top execution skills, and where we draw our security line.

2026-10-23

Productivity Is Our Biggest Category — And It Isn't About Getting More Done

173 skills, second-highest average score, and almost no todo lists. The category is really three jobs: checking your own work, thinking critically, and simulating stakeholders before decisions.

2026-10-23

You Can Now Build Your Own Skill on Skill123 — Scored Before You Download It

Our guided skill builder is live: four plain-language questions in, a scored SKILL.md out. The evaluation step — same six-dimension rubric as our 745-skill directory — is what makes it more than a file generator.

2026-10-23

A Quarter of Skills State What They Won't Do — Just Not Where It Matters

25.6% of SKILL.md bodies contain a "not for" boundary. Only 0.5% of taglines do. The boundary exists in the corpus; it is just buried where the router never reads it.

2026-10-23

A Ten-Minute Audit for Any Skill

A time-boxed checklist based on what 765 SKILL.md files got right and wrong: routing, workflow, output, verification, and the three security questions that matter.

2026-10-23

All Three Skills That Mention rm -rf Are Guarded

A scan of 765 SKILL.md files found "rm -rf" in three of them. Reading the context, none is a careless command — each is an example of how destructive operations should be written.

2026-10-23

Almost a Quarter of Skills Have No Numbered Steps

183 of 765 SKILL.md files contain no numbered list items at all. The median file has six. Step count correlates weakly with score (r=0.17) — but the absence of steps is a different problem than the number of them.

2026-10-23

Bash Appears in 1% of Skills. Read and Write Appear in a Quarter.

Among Claude Code tool names, Read (24.4%) and Write (22.6%) dominate the corpus, while Bash appears in 1.0%. The skills ecosystem is mostly about producing documents, not running commands.

2026-10-23

Fewer Than Half of Skills Check Their Own Work

42.9% of SKILL.md files mention verification and 7.8% describe what to do when a step fails. Confirmation before irreversible actions is rarer still, at 1.8%.

2026-10-23

42 Skills Point at example.com. Which Ones Mean It?

42 of 765 skills contain example.com in their instructions, including 7 that use api.example.com as a base URL. Some are illustrations; some are placeholders an agent might actually call. The difference is how the URL is labelled.

2026-10-23

How to Write a "When Not to Use" Clause

The boundary belongs in the description, not the body. A practical guide to writing the clause, with patterns that work, patterns that fail, and a three-request test you can run in five minutes.

2026-10-23

21.8% of the Corpus Is Chinese — and Most of It Is One Author

167 of 765 SKILL.md files contain Chinese text. 119 of them come from a single catalogue, which turns what looks like a bilingual ecosystem into one author's translation decision.

2026-10-23

Only 1.8% of Agent Skills Ask for Confirmation Before Acting

Across 765 SKILL.md files, exactly 14 contain an explicit instruction to confirm before the agent acts. It is the rarest structural feature in the corpus — and the one most associated with a higher score.

2026-10-23

Only 22.7% of Skills Say What the Output Looks Like

174 of 765 SKILL.md files contain an output section. Without one, the last instruction in the workflow becomes the de facto contract — and the agent decides what "done" means.

2026-10-23

Progressive Disclosure Is Mostly Aspirational

Only 4.2% of the 765 SKILL.md files reference an assets/ directory, 22.6% reference scripts/, and 53.5% reference references/. The pattern everyone recommends is the one almost nobody ships.

2026-10-23

Every Structural Feature We Measured Fits Inside the Evaluator's Noise

We correlated ten structural features across 765 SKILL.md files with their scores. The largest gap is 0.55 points. Our own re-runs move scores by up to a full point. That makes the whole table directional at best.

2026-10-23

30 Skills Call an API. Only 15 Document the Rate Limit.

Thirty skills in the corpus reference an api.* endpoint. 28 document auth, but just 15 mention rate limits and 14 describe environment variables. The missing third is where agents get stuck.

2026-10-23

The Two Skills That Say "Ignore Previous Instructions" Are Defending Against It

A naive scan of 765 SKILL.md files found the classic injection phrase in two of them. Both occurrences are inside security rules telling the agent to treat such text as data. String matching cannot tell a warning from an attack.

2026-10-23

Two "Standards" That Are Really One Author's Template

The "Skill Contract" heading appears in 119 files and "Data Sources" in 125. Both look like emerging conventions. 118 of the 125 and all 119 come from the same catalogue.

2026-10-23

What 765 SKILL.md Files Actually Contain

Overview in 828 sections, Source in all 765, quick start in 162, instructions in 148. A frequency table of every heading in the corpus, and what the shape says about how skills are written.

2026-10-23

What an LLM Judge Can and Cannot Tell You About 765 Skills

Our six-dimension rubric ranks skills consistently, but run-to-run movement of about a point is larger than every structural effect we measured. Here is how to read the scores — and what they are not.

2026-10-23

What Changed After We Re-Evaluated 174 Skills

174 skills carried zeroes in three dimensions because an older evaluator never wrote those fields. Re-scoring them changed the bottom of the distribution, not the top — and exposed how a missing field can masquerade as a measurement.

2026-10-23

What We Learned From Scoring 765 Skills

After 20 posts of corpus analysis, three themes hold up: routing is decided in the description, safety is mostly absent, and most "structural" effects are smaller than the measurement noise.

2026-10-23

Score Predicts Downloads. Star Count Predicts Nothing.

Across 591 evaluated skills, our score correlates with downloads at r=0.47 while GitHub stars correlate with quality at r=0.02. But the relationship is a cliff, not a slope — the effect lives almost entirely above 9.5.

2026-10-21

The Dimension That Predicts Downloads Is Not the One We Rank Highest

Across 591 evaluated skills, structure (r=0.56) and content (r=0.52) correlate with downloads far more than trigger quality (r=0.15) — even though trigger quality is the dimension we say decides whether a skill ever fires.

2026-10-21

The Long Tail Isn't Winner-Take-All

In a directory of 765 skills, the top 10% account for only 31% of downloads and the top 50% for 79%. Twenty percent have never been downloaded at all. That is a much flatter distribution than app stores, and it changes how a new skill should think about competition.

2026-10-21

Six Authors Supply Three Quarters of the Directory

765 skills come from 65 GitHub authors — but six accounts supply 535 of them, and the top ten account for 76%. The "ecosystem" is closer to a dozen publishers with large catalogues.

2026-10-21

What 765 Taglines Reveal About How Skills Describe Themselves

23% of skills open their description with "Use when", but only 0.5% say what the skill should not do. The negative space we keep recommending is almost entirely absent from the corpus.

2026-10-21

Longer SKILL.md Files Score Higher — Which Sits Awkwardly With Our Own Advice

Across 591 evaluated skills, SKILL.md length correlates with score at r=0.24, and the longest quartile averages 8.82 versus 8.14 for the shortest. Our rubric says to keep files lean. Both can be true, but only if you read "lean" correctly.

2026-10-21

The 152 Skills Nobody Has Downloaded

19.9% of the directory has zero downloads — and those skills average a higher score (9.03) than the ones that have been downloaded (8.52). The number is measuring exposure, not quality.

2026-10-21

Single-Skill Repos Beat Mega-Repos on Score — Barely

29 skills come from repositories that publish exactly one skill, and they average 8.69 with 10.5 downloads each. The 440 skills inside mega-repos average 8.49 and 7.5 downloads. The gap exists, but it is smaller than the narrative suggests.

2026-10-21

A Category-by-Category Look at the Six Dimensions

Marketing skills lead on trigger quality (9.30) but trail badly on engineering (5.61). Finance is the mirror image. Documents score highest overall. The variation is mostly about who writes in each category, not what the category requires.

2026-10-21

The 18 Skills That Scored 9.5 or Higher

Out of 591 fully-evaluated skills, 18 clear 9.5 — and 12 of them come from just two authors. What the elite set has in common, and why the list is narrower than it looks.

2026-10-21

The 18 Skills That Scored Below 7.0

Two categories — design and finance — account for 10 of the 18 lowest-scoring skills, and five of those come from a single family. The common failure is not bad code; it is a file that is a knowledge dump instead of a workflow.

2026-10-21

What the September and October Batches Look Like

555 fully-evaluated skills arrived in September, 36 in October — and the October batch averages 0.5 points higher but a third of the downloads. Little of that is about quality; almost all of it is about time and discovery.

2026-10-21

Views and Downloads Move Together — But Neither Measures What You Think

Across 591 evaluated skills, views and downloads correlate at r=0.46. Median views is 2 and median downloads is 5 — which tells you more about what the counters measure than about what users want.

2026-10-21

The Verified Badge Doesn't Predict Quality — and That's Fine

577 verified skills average 8.53; 14 unverified average 8.21. The gap is small, and the badge doesn't mean what readers assume. Here's what it actually says.

2026-10-21

Twelve Skills Have a Placeholder Tagline — a Sign of What Breaks in Bulk Import

Twelve skills in the directory carry a tagline like "A skill from the GitHub repository X" instead of a real description. It is a small number, and a useful illustration of how bulk import fails silently.

2026-10-21

Why 174 Skills Show Zeros in Three Dimensions

22.7% of the directory displays a score around 8 while structure, content, and engineering are all zero. They were imported but never evaluated. Here is the defect, its blast radius, and what we are doing about it.

2026-10-21

Reading 765 Skills: What the Ecosystem Looks Like Now

765 skills, 65 authors, 66 repositories, a median score of 8.5 and a median of two views per skill. A compact snapshot of what a skill directory looks like after two months.

2026-10-21

The Shape of a Skill Name

37 skills have single-word names. 16 end in -builder, 14 in -expert, 10 in -pro. Naming patterns cluster by category and by author, and the most common suffix is not the one that predicts a better score.

2026-10-21

Skill Families and How They Grow

37 name-prefix families cover 214 of 591 evaluated skills. The largest — baoyu with 21 skills — averages 9.11. The widest spread inside a family is nearly three points. What that says about how catalogues age.

2026-10-21

What a Placeholder Row Does to Your Averages

Adding 174 rows with zeroed dimensions moves the site-wide structure average from 8.37 to 6.47 — a 1.9-point distortion — while barely moving the overall score. A worked example of why completeness checks come before analysis.

2026-10-21

Skills With a Built-In Verdict — The Ones That Can Say No

Most skills can only say yes — produce, generate, output. A small class encodes a standard strong enough to reject its own work: validators, auditors, reviewers. They score a full point above the directory mean, and they're the safest class we host.

2026-10-21

One Skill, Two Languages — The Hidden Tax on Bilingual SKILL.md Files

Chinese-native skills for Chinese platforms score 9.1–9.4 — the language is load-bearing. Skills that merge two languages in one file pay a measurable tax on trigger and content quality. The data, and the fork-don't-merge pattern.

2026-10-21

Super-Authors — The Handful of Names Behind a Third of Our Top 50

Sort 600+ evaluated skills by score and the author column collapses: eleven skills from one name, ten from another, eight from a third. What the top three multi-skill authors share — systems, not files.

2026-10-21

The 21-Skill Personal Toolkit — One Author, Every Publishing Job, 8.9 to 9.5

The largest single-author skill family in our directory covers the entire bilingual content pipeline — translate, fetch, summarize, diagram, generate images, build slides, post to WeChat. Twenty-one skills, zero overlap, a masterclass in narrow triggers.

2026-10-21

One Repo, 92,647 Stars, Four Skills — What Megarepos Actually Ship

The highest-starred repos in our directory ship few skills — a constraint methodology, an interviewer, a diagram tool. With the hype variable maxed out, what remains is the actual skill engineering, and it's not wrappers.

2026-10-21

Skill Interference — Why Your Tenth Install Is Weaker Than Your First

Every skill is evaluated in isolation, but users install them in sets — and sets fail differently: trigger collisions, capability overlap, contradictory defaults. Four interference patterns and the loadout rules that avoid them.

2026-10-20

The Skills That Interview You — And Why the Best Prompts Are Questions

A pattern hiding at the top of our directory: skills whose core mechanism is asking the user one sharp question instead of guessing. From academy-guide (9.6) to discernment-nudge to interview-me (9.4).

2026-10-20

Agent Skills That Work on Your Machine, Not Just in Chat

A smaller class of skill operates the machine around the agent — driving a real logged-in Chrome, extracting from Electron apps, managing Obsidian vaults. They score well precisely because they define their side effects honestly.

2026-10-20

Skills That Answer in HTML — The Highest-Scoring Pattern We Hadn't Named

A quiet pattern runs through our top scores: skills that turn an agent's answer into a one-page visual artifact — HTML explainers, isometric SVG figures, motion-graphics videos. Five skills from 9.3 to 9.6.

2026-10-20

The HyperFrames Family — Ten Skills, One Video Pipeline, Scores From 5.7 to 9.7

One author shipped ten skills covering a complete agent video pipeline. Nine scored 8.7 or above — and the one that didn't shows exactly what a missing system looks like.

2026-10-20

The Best Frontend and React Skills for Claude Code

Six frontend and React skills for Claude Code, scored 8.5 to 9.2 — what each one is actually good at, where they overlap, and which combination covers a real app.

2026-10-17

The Best Security and Authentication Skills for Claude Code

Three security skills for Claude Code scored 8.5 to 9.3, what each covers, and why the security dimension's scores say less than the capability manifest does.

2026-10-17

The Best Architecture and System Design Skills for Claude Code

Six architecture skills for Claude Code scored 9.0 to 9.4 — system design, cloud architecture, RAG, microservices, and the framework-specific architects that fill the gaps.

2026-10-17

The Best Deploy and CI/CD Skills for Claude Code

Seven deploy and CI/CD skills for Claude Code scored 8.4 to 9.4 — infrastructure as code, containers, Kubernetes, Vercel deployment, and the one that ships an app end to end.

2026-10-17

The Best LinkedIn Skills for Claude Code

Eight LinkedIn skills for Claude Code scored 8.5 to 9.5 — posting, commenting, replying, planning, hook extraction, and de-AI-ing drafts. What each does and which four cover a real workflow.

2026-10-17

The Best Skills for Publishing to WeChat and Xiaohongshu

Eight skills scored 9.1 to 9.4 for Chinese social publishing — WeChat Official Accounts, Xiaohongshu image cards, Weibo, and the account-level strategy skills that surround them.

2026-10-17

The Best User Research Skills for Claude Code

Eight user research skills scored 7.9 to 9.4 — interviewing yourself and users, simulated buyer panels, pricing, offers, and the interview scripts that produce better material.

2026-10-17

The Best SEO Skills for Claude Code

Seven SEO skills for Claude Code scored 8.6 to 9.4 — technical audits, on-page audits, SERP markup, programmatic pages, keyword mining, and SERP analysis.

2026-10-17

The Best Slide Deck Skills for Claude Code

Six slide deck skills for Claude Code scored 7.3 to 9.7 — native PPTX editing, image-based decks, article-to-deck conversion, and the narrative skills that decide what the slides should say.

2026-10-17

The Best Meeting and Notes Skills for Claude Code

Four meeting and notes skills for Claude Code scored 7.9 to 9.3 — transcript capture, structured meeting summaries, interview notes, and release notes built from the same raw material.

2026-10-17

The Best Database Skills for Claude Code

Four database skills for Claude Code scored 7.6 to 9.4 — query optimization, schema design, PostgreSQL operations, and natural-language SQL. Which one fits which problem.

2026-10-17

The Best Code Review Skills for Claude Code

Six code review skills for Claude Code scored 8.7 to 9.4 — bug and security review, over-engineering review, diff comprehension, and the quality workflows that run before a merge.

2026-10-17

The Best Image Generation Skills for Claude Code

Five image generation skills for Claude Code scored 8.2 to 9.3 — multi-provider generation, article covers, website assets, and the prompt guides that make the models behave.

2026-10-17

The Best Diagram and Visualization Skills for Claude Code

Four diagram and visualization skills for Claude Code scored 8.2 to 9.4 — SVG diagrams, Mermaid rendering, 3D explainers, and d3 charts. Which output format fits which audience.

2026-10-17

The Best Django and FastAPI Skills for Claude Code

Four Python web skills for Claude Code scored 9.0 to 9.4 — Django applications, Django storage configuration, FastAPI APIs, and the modern-Python baseline skill underneath them.

2026-10-17

The Best Crypto and Trading Skills for Claude Code

Seven crypto and trading skills for Claude Code scored 8.6 to 9.3 — Solana DEX execution, prediction markets, DeFi analytics, DEX pair data, and AMM pool mechanics.

2026-10-17

The Best Landing Page and Copywriting Skills for Claude Code

Five conversion skills for Claude Code scored 8.4 to 9.0 — landing page optimization, offer construction, conversion tracking QA, value mapping, and pre-launch experience checks.

2026-10-17

The Best Testing and TDD Skills for Claude Code

Five software testing skills for Claude Code scored 9.0 to 9.1 — TDD workflow, E2E with Playwright, test generation, and full-flow click testing. Plus why "testing" skills for marketing are a different category.

2026-10-17

The Best Data Visualization and Chart Skills for Claude Code

Five data visualization skills for Claude Code scored 7.9 to 9.3 — infographics, d3 charts, motion graphics, 3D explainers, and trading-specific charts.

2026-10-17

The Best Podcast and Audio Skills for Claude Code

Seven audio skills for Claude Code scored 8.4 to 9.4 — text-to-speech in Chinese and English, music generation, audio-to-video, transcripts, and voice consistency.

2026-10-17

Security Almost Never Separates a Good Agent Skill From an Average One

Across 628 scored agent skills, security is the least discriminating of our six dimensions — 92% score 9 or above, and it explains just 0.09 of the 1.97-point gap between top and median skills. Workflow design, trigger quality, and structure are what actually separate them.

2026-10-17

How to Read a SKILL.md in 60 Seconds

A six-check reading routine for people choosing a skill, ordered by the dimensions that actually separate good skills from average ones — file structure first, security last, and the one question almost every description omits.

2026-10-17

"Which Skill Category Is Best?" Is the Wrong Question

Across nine categories and 591 evaluated skills, the gap between the best and worst category is 0.59 points — while the gap inside a single category runs up to 4.1. Category averages mostly measure who happened to write the code.

2026-10-10

The 20-Point Field Most Agent Skills Get Wrong

Across 591 evaluated skills, trigger quality is the weakest of our six dimensions at 7.74/10 — and it decides whether the skill ever fires. The most common failure is a missing "when NOT to use this" clause.

2026-10-10

MCP Server or Skill? A Decision Rule That Actually Decides.

Both make your agent more capable and they fail in different ways. If you can write the capability as a function signature it is a server; if describing it takes conditions and exceptions it is a skill. Includes the five MCP-related skills we scored.

2026-10-10

We Ran the Same Evaluation Twice. Seven of Eleven Scores Moved.

Same skills, same SKILL.md, same rubric, same model — re-run two hours later. Seven of eleven scores changed, the largest by a full point. What that means for reading any automated score, ours included.

2026-10-10

We Scored 591 Agent Skills. Here's the Shape of the Data.

The score distribution across 591 evaluated skills and the six-dimension averages behind the headline number — including why our highest-scoring dimension is partly an artifact of how we score.

2026-10-10

We Scored the Skills Inside the Most-Starred Agent Repos. Star Count Is Not a Score.

We pulled live star counts for the top agent-skill repos and compared them with our per-skill evaluations. Inside a single repo scores span 6.7 to 9.8 — starring a repo does not tell you which of its skills to install.

2026-10-10

AI Agents Now Outnumber Human Users on Our Skill Directory

30 days of install data: agentsaccount for 87% of all skill installs. What that means for how you distribute agent skills.

2026-10-08

What 600 Skill Evaluations Taught Us About Writing Good Agent Skills

Patterns from scoring 600+ skills: the top 1% verify their own output, trigger boundaries are the weakest link, and popularity correlates with nothing.

2026-10-08

How We Dynamically Test Agent Skills (Not Just Read Them)

Static review catches what a skill says. Dynamic testing catches what it does. Our probe-based methodology, explained end to end.

2026-10-08

The Best MCP Skills for Claude Code (Tested and Scored)

MCP skills teach your agent to build and use Model Context Protocol servers. The ones worth installing, evaluated.

2026-10-08

The Best Email Skills for Claude Code (Deliverability, Campaigns, Compliance)

Email skills that survive contact with real inboxes — deliverability QA, rendering, and the consent gap in the category.

2026-10-08

We Ran 384 Adversarial Probes Against 48 Agent Skills. Injection Wasn't the Problem.

Prompt injection hidden in artifacts failed 1.6% of the time; probes that simply pushed past a skill's declared scope failed 30%. Full per-probe dataset published under CC BY 4.0 — including the two ways our own grading flatters the skills.

2026-09-29

We Injection-Tested the Legal Skills: What Actually Broke

We ran the legal-category skills through live adversarial probes — prompts hiding instructions in intake files, telling them to skip their own disclaimers. The readers held. The generators cracked. Full results inside.

2026-09-28

The Security-Conscious Skill Loadout: What to Install, How to Contain It

After evaluating 600+ skills — including one credential harvester we delisted — the containment-first loadout, and how to read the new dynamic-injection badges on every skill page.

2026-09-28

The Solo Founder Skill Stack: Product, Launch, Legal, Without Hiring Anyone

The cross-category loadout a one-person company actually needs — spec the product, build the funnel, paper the deals. Seven skills spanning four categories, tested and scored.

2026-09-28

Ship a Side Project in a Weekend: The Deploy-It-Yourself Skill Loadout

You don't need to be an engineer to ship anymore — you need an agent with the right four skills. Spec, scaffold, deploy, polish. Tested scores and the one gotcha per skill.

2026-09-28

The Content Repurposing Pipeline: One Idea, Every Platform, One Agent

Turn one YouTube video into an article, a slide deck, a LinkedIn post, and a WeChat piece — without writing any of them twice. The baoyu family tested and sequenced.

2026-09-28

The Meeting-to-Deliverable Pipeline: Recording In, Documents Out

Meetings produce recordings; recordings produce nothing. The skill chain that turns a call into a recap, a spec, and a deck — with injection results for the document generators.

2026-09-28

The Product Manager Skill Stack: Spec, Deck, Ticket, Numbers

The PM's actual week — PRD, review deck, tickets, retention numbers — covered by six tested skills across four categories. Includes injection-test results for the document generators.

2026-09-28

The Research-to-Publish Workflow: Sources, Citations, Drafts, Done

For the grad student, the paralegal, the analyst: the skill chain from a pile of sources to a citation-clean, formatted document. Five skills, four categories, injection results included.

2026-09-28

Skills That Call Other AIs: claude-api, dashscope, and the API-Reference Family

A quiet category of skills teaches your agent to call other models and APIs — image generators, market data, on-chain data, even Claude itself. Tested scores, plus why these deserve the strictest security review.

2026-09-28

The Bilingual Publishing Stack: One Workflow, Both Internets

Publishing in both Chinese and English isn't translation — it's two platforms with different physics. The baoyu family, dashscope, and video-translate, sequenced for creators working both markets.

2026-09-28

The Muse Skills: Session Memory for AI Pair Programmers (Evaluated)

Your AI coding agent forgets everything when the session ends. The MUSE family — /resume, layered context, strategic compaction, semantic search — is a complete memory system in plain Markdown. All six core skills evaluated.

2026-09-27

The Jev Skill Family: Five Workflow Primitives for Coding Agents (Evaluated)

Most agent skills are single prompts. The Jev family arrived as a system — five skills that sort, find, judge and act, the first family to score 10/10 on security in every member.

2026-09-24

🛡️ Security Report

Monthly Security Report: September 2026

611 skills evaluated, one credential harvester delisted, and what our capability manifests found in the agent-skill ecosystem.

2026-09-24

The Complete Guide to AI Agent Skills

What AI agent skills are, where to find them, and the six-dimension framework for telling great skills from prompt-wrappers before you install.

2026-09-10

I Found an Agent Skill That Harvests Browser Credentials

What checking the scripts — not the README — revealed about a popular research skill, and the three-step security check I now run before installing anything.

2026-09-16

The 10 Claude Skills Worth Installing (And What Most People Get Wrong)

I evaluated 500+ agent skills on a six-dimension rubric. These are the ten I would actually install, each with the one mistake people make when using it.

2026-09-17

The 10 Best Video-Generation Skills for AI Agents (Tested and Scored)

Video is where agent skills went from nice trick to production tool. I evaluated the full video category — 60 skills and counting. Here are the ten that hold up.

2026-09-19

Document Skills for Claude Code, Tested: Excel, Word, PowerPoint, PDF

I tested every document skill in the ecosystem — Excel workbooks, Word documents, PowerPoint decks, and PDF workflows. Here are the seven worth installing.

2026-09-22

Developer Tool Skills for Claude Code, Tested: Debugging, TDD, Architecture

The skills that teach your agent real engineering discipline — systematic debugging, test-driven development, architecture diagrams, and the API reference that scores 9.8.

2026-09-22

Data & Analytics Skills for Claude Code, Tested: pandas, SQL, ML Pipelines

I evaluated every data skill in the ecosystem — pandas workflows, PostgreSQL optimization, Spark, ML pipelines, and A/B test statistics. Seven worth installing.

2026-09-22

Design & Creative Skills for Claude Code, Tested: Diagrams, Images, Infographics

The creative category has the widest quality spread — for every skill that produces professional output, three produce generic AI slop. Here are the seven that cleared the bar.

2026-09-22

Writing Skills for Claude Code, Tested: Content, Editing, Translation

Writing skills are the most crowded and most mediocre category — half are prompt-wrappers. Here are the seven that actually improve your agent's writing.

2026-09-22

Marketing Skills for Claude Code, Tested: SEO, Ads, Email, Landing Pages

The marketing category is dominated by a single coherent skill family — a full growth stack covering SEO, ad testing, email, and audience research. Ten skills, tested and scored.

2026-09-22

Legal Skills for Claude Code, Tested: Contracts, Compliance, Legal Research

Legal skills are the newest and most cautious category — CCPA compliance, Bluebook citations, NDAs, patent analysis, and complaint drafting. Six skills, evaluated.

2026-09-22

Finance & Trading Skills for Claude Code, Tested: Crypto, DeFi, Market Data

The finance category leans into crypto infrastructure — Solana DEX execution, MEV protection, and free market-data APIs, plus systematic trading methodology. Ten skills, tested and scored.

2026-09-22

Video Creation Skills for Claude Code, Tested: Explainers, Motion Graphics, 3D

Video creation is the fastest-growing skill category — code-drawn frames, music-synced visuals, 3D scenes, and mathematical animation. The full collection, evaluated.

2026-09-22