full-stack-skills: a catalog of 788 installable Agent Skills

A monorepo that became a catalog

The shape of this repository is the result of a decision, not an accident. The README describes Full Stack Skills as a navigation site for a collection of Agent Skills aimed at AI coding agents, and explains that it was originally a monorepo with 460 or more stars, now split into 51 independently installable packages under the full-stack-skills GitHub organisation. A note records that the migration finished in June 2026, that the 788 skills now all live in separate repositories, and that this repository is the catalog and directory, with the existing stars and history kept where they are.

What a package contains is described in one line: domain knowledge, workflows and reference material that an agent loads on demand, which keeps context consumption small while still supplying deep domain expertise when it is needed. So the unit of value is not code you import but instructions an agent reads.

The counting is the part to interrogate. The README states a statistics date of 2026-10-06, and that the figures come from counting the skills/*/SKILL.md entries on each package's main branch, giving 51 packages and 788 skill entries, not deduplicated by name. Community recommended entries are not counted, and three packages, skills-toolchain, teaching-skills and social-skills, keep their original scope and are not included in the catalog for now. The inventory itself is published as docs/catalog-inventory.json alongside a skill index in SKILLS_INDEX.md, so the numbers are auditable rather than decorative. An entry count that includes same-named skills in different packages is an index of instructions, not a count of distinct techniques.

Context pressure, not lack of skills, is the problem being solved

The README sets out four gaps in a table, and the reasoning behind the package split is clearer once you read them.

The first is context overflow: loading every skill at once exceeds the token limit. The second is domain expertise, the claim that a general model lacks deep framework knowledge, answered with 788 professional skills across 15 domains. The third is discoverability, since skills scattered across repositories are hard to find, answered by a single README with links to every package. The fourth is version locking, since a monorepo forces coordinated releases, answered by independent version management and publishing per package.

The first and the fourth are the technical reasons the repository looks the way it does. If the catalog is a directory and the packages are the products, then installing is a per-package operation and upgrading one package does not force a coordinated release across the other fifty.

There is a second layer of structure around the catalog. The ecosystem table lists four directions under two organisations: Full Stack Skills for software development, architecture, testing and operations, Full AIGC Skills for image, video and audio content creation, and matching plugin collections for each, named Full Stack Plugins and Full AIGC Plugins. The distinction between a skill and a plugin is not spelled out in the README, so a reader has to infer it: skills are instructions an agent reads, plugins are tool integrations, and the split lets an agent take domain knowledge without also taking an integration surface.

npx skills add is the whole installation

Installation starts in a project directory with Node.js and npm present, and the README is explicit that this repository is a navigation catalog: you install from the individual skill package repositories, and the CLI interactively selects the skills and the host.

The first command to know is the one that shows what a package contains before committing to it.

npx skills add full-stack-skills/vue-skills --list

Adding a single skill to a specific host is the other common case, and the flag names are part of the CLI contract rather than something to guess.

npx skills add full-stack-skills/vue-skills --skill vue3 --agent codex

The same command with the global flag installs for the user instead of the project, and the README shows it commented out in its own example, which is a reasonable default posture for a file that lands in someone's repository.

# npx skills add full-stack-skills/vue-skills --skill vue3 --agent codex --global

If you want a package rather than a named skill, the command is the same without the skill flag.

npx skills add full-stack-skills/vue-skills --skill vue3

Once something is installed, the README says you can list what is there with npx skills list, and points to QUICKSTART.md, PLATFORM_GUIDE.md and the Skills CLI documentation for more. The practical point for a team is that installation state is per project by default, which means it travels with the repository, and that the choice of host is made at install time rather than at load time.

Domain totals hide packages with a single skill

The catalog is organised by domain, and the domain totals are large enough to be misleading if you read them as coverage. AI design tools listed as MCP hold 170 skills across three packages, with t2ui-skills at 97, stitch-skills at 45 and pencil-skills at 28. Backend frameworks come to 184 skills. Frontend frameworks total 67, split across vue-skills at 7, react-skills at 7, angular-skills at 1, svelte-skills at 16, uniapp-skills at 13, swift-skills at 7, agent-skills at 16 and pixijs-skills at 1. Frontend UI libraries add 13 across four packages, with uview-skills at 2. Build and tooling account for 26, split into build-skills at 7, nvm-skills at 15 and vscode-skills at 4. Charts and ASCII art cover 15, with ascii-skills at 13 and chart-skills at 2.

Read the per-package numbers rather than the domain totals and the picture changes. A package with one skill, such as angular-skills or pixijs-skills, is a thin entry whatever its domain total says, while a package with 97 entries is a substantial body of instructions that will pull context when it loads. A team that installs on demand, which is the whole premise, needs the per-package view to know what a given install will cost in context.

The catalog also lists one recommended package, baoyu-skills, at 21 skills, alongside a table of related resources: the Agent Skills specification at agentskills.io, the Skills CLI in the vercel-labs/skills repository, the skills.sh directory, and other authors' collections including addyosmani/agent-skills, browserbase/skills and google/skills.

Host compatibility is a first class concern, not a footnote

Every install command in the README carries an agent flag, and that is because the same skill file has to land in different hosts with different conventions. The topic list for the repository names agent-skills, claude-code, codebuddy, cursor, qoder and skills, which tells you the range of tools this catalog is expected to serve.

The repository backs that up with a PLATFORM_GUIDE.md, a QUICKSTART.md, a ROLE_DEFINITIONS.md and a PLANNING_SKILL_CATEGORIES.md, and with host-specific README files named README_agentskills.md, README_anthropic_skills.md, and Chinese-language variants of each. The presence of the agentskills.io badge and a compatibility claim against that specification suggests the skills follow a portable format rather than a host-private one, which is the property that makes a single catalog usable across several tools.

There is also an evaluation artefact, EVALUATION_REPORT.md, and a PROVENANCE.md, plus THIRD-PARTY-NOTICES.md for third-party attribution and a CLAUDE.md and AGENTS.md family of files at the root for agents working in the repository itself. Whether the evaluation report covers quality of the individual skills or only the catalog process is not stated in the README, so a reader who wants scored quality per skill should open that file rather than infer it from the 788 figure.

Where this catalog sits against writing your own skills

There are two reasonable alternatives, and they fail in different ways.

The first is to skip curated skills and write your own. That gives you exactly the conventions and the house rules you want, and it keeps your internal APIs and deployment steps inside the agent's instructions instead of in a public repository. What it costs is the domain depth. The catalog's argument is that a general model lacks the deep framework knowledge that 788 hand-written instruction files encode, and that argument is strongest for a team that has no time to write a Vue or Spring or Tauri playbook from scratch.

The second is to take skills from elsewhere in the ecosystem rather than from this catalog. The related resources table points at the Agent Skills specification, the skills.sh directory, the vercel-labs Skills CLI, and collections maintained by other authors, and the catalog itself lists a recommended package that is hosted outside the full-stack-skills organisation. The difference in approach is curation and consistency: this catalog gives one table with per-package counts, a published inventory record and a stated counting rule, while a directory gives you a search over many independently published collections with no unified count. Neither is wrong, but the catalog is the one you can audit against SKILLS_INDEX.md.

Apache 2.0 across the packages, and no release for the catalog itself

The licence for this repository is Apache 2.0, with a LICENSE file at the root, and the table entries show packages published by several different owners, which is a reminder that the licence of the catalog and the licence of each skill package are separate questions. THIRD-PARTY-NOTICES.md exists for that reason.

Maintenance signals for the catalog are mixed by design. The repository is not archived and the last push was on 2026-10-06, matching the statistics date in the README, so the index itself is current. But the repository has no GitHub releases, which is consistent with a directory that has no version to release: the versions live in the 51 packages, each with its own version management and publishing, and the catalog is refreshed by pointing at their main branches and recounting.

The upgrade cost for an adopter follows from that. Upgrading a skill means reinstalling or updating one package, and there is no catalog-wide release to pin, so a team that wants reproducibility should pin per package or vendor the skill files it relies on. The count of 788 is a description of a moving target, and the inventory file and index are the two documents that make it checkable rather than a marketing number.

Editorial conclusion

full-stack-skills fits a developer who wants curated, per-package domain instructions for a coding agent without maintaining a personal monorepo of skills, and it does not fit someone who needs a single pinned skill set, a measured quality score, or a release number for the catalog itself. Before installing, run the list command against one package and read SKILLS_INDEX.md and docs/catalog-inventory.json to see exactly which skills the 788 count includes.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. partme-ai/full-stack-skills on GitHub
4. README