Installing All 17 E-Commerce Skill Repos: 1,353 Skills, 2.5 Million Tokens, and the 26 Files That Never Load

The request was straightforward: install every skill from these seventeen repositories in order, and report what each one contains.

I did it. All seventeen repositories are real, they total 855 MB, and they contain 1,353 skills that an agent can load plus 28 files that look like skills but are not. Everything is installed — 1,337 directories sitting in ~/.claude/skills/ — and the per-repository breakdown is below.

What the exercise also produced is a measurement nobody makes before recommending a skill collection: the combined markdown is 10.00 MB. In tokens, that is somewhere between 2.5 and 2.9 million — two and a half times the entire context window of the model these skills are written for. Six repositories also behave differently from their own documentation, and one of them installs nothing at all if you follow its README exactly.

What each repository actually installed

Verified by walking the cloned filesystem, not by reading descriptions.

Thirteen of the seventeen are Chinese-market or cross-border e-commerce toolkits aimed at Amazon, TikTok Shop, Shopify and Etsy sellers. Two are official vendor repositories (Anthropic, Shopify, plus WooCommerce and Remotion-adjacent work). The spread in ambition is wide enough that “e-commerce skills” describes almost anything: Amazon PPC bid adjustment, Shopify Liquid templating, SEO audit checklists, product photo generation, TikTok lead generation, and one repository that is mostly about making videos.

The 26 files that will never load

dlazy-ai/ecommerce-skills appears to be a well-built repository. It has 26 skills organised into directories with references/ subfolders, a shared schema, an evals directory, an MCP folder, and a package.json. Its README’s install line is npx skills add https://github.com/dlazy-ai/ecommerce-skills --all with the comment “install all 26 at once”.

Every one of those 26 skills is named skill.md, in lowercase.

Agent skill discovery is case-sensitive, and the convention every major agent follows is the uppercase SKILL.md. I verified this against the other sixteen repositories: 1,353 of 1,353 loadable files are uppercase. Claude Code, Codex and Cursor all glob for SKILL.md. A lowercase skill.md sitting inside skills/creative-scene/ is invisible to that glob.

So the instruction “install all 26” produces zero working skills on Linux or macOS. A case-insensitive filesystem would mask it — which is probably why it shipped, since whoever tested it was likely on a Mac or Windows machine where the pattern still matches.

hypit-ai/hypit has a milder version of the same problem: one uppercase SKILL.md and two lowercase skill.md files inside the same repository. One works; two do not.

This is the single most actionable finding in the whole exercise, and it only appears when you install rather than read: the repository tells you it will add 26 skills, and it adds none.

Ten megabytes of markdown is 2.5 million tokens

Add the bodies of all 1,353 skills together and you get 10,003,749 characters.

At a conservative four characters per token, that is 2.5 million tokens. At 3.5 characters per token — reasonable for prose-heavy markdown — it is closer to 2.86 million.

Put that next to the context windows the skills are written for:

• Claude Opus 5.5: 1M tokens → the collection is 250% of a full window
• A 200K window (most current coding agents): the collection is 12.5× over

No agent can hold this set. Not with progressive loading, not with lazy skill discovery on any realistic task — a single task that needs Amazon keyword research and Shopify theme work would pull skills from two repositories whose combined markdown already exceeds what fits comfortably.

The practical consequence is that “install all of these” is not a plan, it is a search problem. Each repository is a catalogue, and the useful unit is one or two skills per task. The 596-skill repository is not a larger gift than the 2-skill one; it is a larger haystack.

There is a second-order problem that no README mentions: skills are selected by their description text, and every description competes for the same attention. With 1,353 candidates, description collisions stop being hypothetical. seo-audit exists in two repositories with different opinions about what an audit contains. An agent that loads both gets contradictory instructions and no way to know which is authoritative.

Shopify ships 565 MB for 30 skills

The Shopify repository is thirty skills and 565 MB. That ratio — 18.8 MB per skill — deserves an explanation, and the filesystem provides one:

• shopify-skills-test/ — 464 MB, containing .agents/skills/*/node_modules/ with platform-specific esbuild binaries (darwin-arm64, and others)
• Six scripts/validate.js files at 11 MB each, unminified bundles committed as validation tools

So roughly 464 MB of the clone is a test fixture that never needed to be in the repository, and another 66 MB is build output. The actual skill content — the markdown an agent reads — is 0.7 MB. The ratio of what ships to what loads is about 800:1.

This matters more than disk space. A 565 MB clone is slow, it will trip CI clone timeouts, and shopify-skills-test/ ships a second copy of the skills, which is why fourteen skill names appear twice inside this single repository. Any install that walks the tree without excluding test directories installs both copies, and the second silently overwrites the first.

One repository is 596 skills, generated from YouTube

xjli360/sealeap-amazon-skills is the largest contributor and the most unusual. Its structure is one directory per skill, nested under amazon-skills/youtube/<creator>/, with names like:

• sealeap-qiongqi-amazon-search-term-report-pivot
• sealeap-qiongqi-amazon-launch-keyword-intent-groups
• sealeap-fenghuang-amazon-new-seller-launch-path

The pattern is systematic: a creator handle, then a topic. This is a pipeline that converts video tutorials into skills — each one named after the person whose advice it encodes. At 596 skills averaging 4.4 KB of markdown each, the repository is closer to a searchable transcript archive than a toolkit.

That is a legitimate product decision, and it explains the repository’s 250 stars: people want the knowledge. But it also means installing it adds 596 descriptions competing for the agent’s attention, most of which will never be relevant to any single task. If you install exactly one repository from this list and expect good results, this is the one most likely to degrade selection quality.

Twenty name collisions, fourteen of them self-inflicted

Installing everything surfaces conflicts that are invisible when you install one at a time:

I resolved them by appending an owner suffix (shopify-admin--shopify, seo-audit--agricidaniel), which is the only way to install both without silent data loss. Nobody’s README warns about this, and it is the predictable result of a skill ecosystem with no namespace.

The Shopify case is the instructive one: the collision is not between two projects, it is inside one repository, because the test fixture is checked in next to the real skills.

57 skills have frontmatter that agents ignore

Of the 1,337 directories that landed, 1,279 have a compliant frontmatter block with both name and description. Fifty-seven do not, and one directory has no SKILL.md at all.

A skill without a parseable name and description is a skill the agent cannot select — it will not appear in the skill list, and it will not be loaded on demand. The affected names cluster around eBay and monitoring tools (ebay-advertising, ebay-product-research, ebay-seller-guide, domain-monitoring, api-monitoring), suggesting they were added to a repository without going through whatever validator the other skills passed.

So the honest count of usable skills from this exercise is not 1,353. It is 1,279 with valid metadata, 28 that never load because of filename casing, and 57 that load as files but never get selected.

Stars predict nothing about skill count

The numbers make a clean argument against using popularity as a proxy for usefulness:

• coreyhaines31/marketingskills — 53,949 stars, 50 skills
• hypit-ai/hypit — 20,613 stars, 1 skill
• AgriciDaniel/claude-seo — 18,622 stars, 35 skills
• finsilabs/awesome-ecommerce-skills — 56 stars, 178 skills
• topviewai/codex-topview-ai — 5 stars, 4 skills

The repository with 178 skills has fewer stars than the one with 2. The two largest star counts in the set belong to marketing and SEO toolkits rather than e-commerce platforms — which makes sense once you notice that marketing skills are useful to everyone, while Amazon PPC skills are useful to the subset of people selling on Amazon.

The watcher counts are more informative. marketingskills has 399 watchers against 53,949 stars (0.74%); hypit has 34 against 20,613 (0.16%). A ratio that low means people starred the repository and moved on — a bookmark, not a dependency.

The repository that most inverts the ranking is finsilabs/awesome-ecommerce-skills: 56 stars, the second-lowest count in the entire set, and 178 skills. It is also the only repository of the seventeen that grades its own contents. Every one of its 178 skills carries a risk field in its frontmatter, split 134 to 44 between safe and critical:

---
name: ab-testing-pricing
description: "Test different price points with proper statistical rigor to find
the revenue-maximizing price while tracking conversion rate and margin impact"
category: pricing-promotions
risk: critical
source: curated
---

The 44 critical entries are the ones that touch money, identity or the law: stripe-integration, paypal-integration, subscription-billing, tax-calculation, pci-dss-compliance, dynamic-pricing, price-rules-engine, discount-engine, fraud-detection, gdpr-ecommerce, data-retention-policies, accounts-payable-management and cogs-tracking-allocation. Reading that list, “critical” clearly means “an agent that gets this wrong moves money or leaks data” — which is exactly the warning that belongs in the frontmatter of a machine-readable skill and nowhere else.

No other repository in the set does this. Sixteen of seventeen ship skills with no indication of what an agent is about to touch, and the one that does is the one with the fewest stars. Grading your own work does not earn stars; it earns the ability to be trusted with a credit-card flow, which is the more useful property.

What is worth actually installing

If the goal is a working e-commerce agent rather than a collection, the seventeen repositories sort into four groups.

Install these first, they are the ones built to be agents rather than lists: anthropics/commerce-agents (16 skills, official, staged-change writes, clear “not needed for” boundaries in every description) and Shopify/agent-skills (30 skills, official, though clone with --depth 1 and exclude the test directory).

Install selectively — these are catalogues, so pick two or three: nexscope-ai/Amazon-Skills (52), nexscope-ai/eCommerce-Skills (162), coreyhaines31/marketingskills (50), AgriciDaniel/claude-seo (35). Each is small enough that its descriptions do not swamp selection.

Install one skill, not the repository: hypit-ai/hypit, liangdabiao/ecom-details-image, feichanggege/ecommerce-visual-copywriting-skill each ship a single, larger, more opinionated skill. That is a feature.

Do not install wholesale: xjli360/sealeap-amazon-skills (596 skills, transcript archive — search it, do not load it), finsilabs/awesome-ecommerce-skills (178 skills), and dlazy-ai/ecommerce-skills (26 skills that will not load until someone renames the files).

FAQ

Did all seventeen repositories exist? Yes. I verified all of them with git ls-remote before cloning, which also sidesteps GitHub’s API rate limit — the API refused me twice at 403 before I switched approaches.

Why install into ~/.claude/skills/? That is the directory Claude Code reads, and it is the one the majority of these READMEs target. The files are plain markdown; nothing was executed. I did not run any repository’s install script, because several ship scripts and the install instructions themselves are enough to get the layout right.

Is 2.5 million tokens a fair estimate? It is an estimate from byte count, not a tokeniser run, so treat it as ±20%. The conclusion does not depend on the precision: the collection is multiple times larger than any context window in use, whether the true figure is 2.2 million or 2.9 million.

Could an agent just load skills on demand? That is what skill discovery is for, and it is precisely why the numbers matter. Selection happens by description, and 1,353 competing descriptions plus 20 name collisions degrade that selection. The failure mode is not running out of context, it is loading the wrong two skills.

What happened to the 855 MB of clones? They are in a scratch directory, not in the skills path. The installed set is 134.6 MB, because I copied skill directories only and excluded node_modules, test fixtures and repository assets. That is the practical reason to install from a clone rather than running a repo’s own script.

Is the summary count 1,353 or 1,337? Both, and the difference is the point. The install action processed 1,353 skills; 1,337 directories exist on disk, because sixteen Shopify skills were installed twice under the same name and the second copy replaced the first.

Conclusion

Seventeen repositories, all real, 855 MB, 1,353 uppercase skills, 28 lowercase files that no agent will read, and 10 MB of markdown that comes to about 2.5 million tokens — more than twice what the model can hold.

The exercise was worth running precisely because it was not sensible. Reading seventeen READMEs would have produced a tidy list of claims; installing all of them produced the things the READMEs do not say: that one repository’s own install instruction yields nothing, that another ships 464 MB of test fixture and a duplicate copy of its own skills, that a third is a transcript archive wearing a toolkit’s clothes, and that twenty names collide in a namespace that has no rules.

Strip it down and the useful shape is small. Two vendor repositories built to be agents, three or four catalogues to draw two skills from, and a handful of single-skill repositories that do one thing with more care than any collection does thirty. That is a fraction of the list — and it is the fraction that will still be in your skill directory, working, after the novelty wears off.