Adobe Skills

The install path was verified end to end with real files landing on disk. Running npx skills add /tmp/adobe-skills-review-20260707 --skill page-tree -a claude-code -y completed in one command and reported "Installed 1 skill. page-tree (copied) to ./.claude/skills/page-tree". The target directory received all five files: SKILL.md, scripts/page-tree-bundle.js, evals/evals.json, package.json, and .releaserc.json. The bundle script is the same 606-line file from the source tree. No truncation, no missing references. This is the Vercel skills CLI doing what the README claims.

The fork density is the highest of any official-vendor skills repo tracked, and it signals real reuse rather than stargazing. At 145 stars and 59 forks, the fork-to-star ratio is roughly 41 percent. For comparison, NVIDIA/skills sits near 8 percent and the typical community pack is under 10 percent. A 41 percent fork density means nearly half the people who star the repo go on to copy or derive from it, which is the strongest available signal that the skills are being put to work rather than bookmarked. The repo was created 2026-02-05 and was pushed the same day as this review (2026-07-07), confirming active maintenance across the five-month window.

The vendor-neutral plugins (web and stardust) give the pack value outside the Adobe install base, and they are the most broadly useful pieces. The web plugin's nine skills (browser-probe, cdp-connect, page-tree, page-reduce, page-collect, page-prep, page-langs, page-tree, domain-mask) drive playwright-cli to connect to a browser over CDP, probe bot protection, capture spatial DOM trees, and reduce pages to skeletons. None of these require an Adobe account. They work for any agent doing web analysis. Stardust is a guided website-redesign pipeline (extract, direct, prototype, migrate, with uplift and audit side entries) built on top of the external impeccable skill, and its EDS delivery half is optional. A developer who never touches Adobe products can still extract real value from these two plugins.

The Bad

Three of the ten plugins are gated behind Adobe IMS OAuth credentials that only Adobe enterprise customers hold, which walls off a meaningful slice of the catalog. The Adobe Analytics, Customer Journey Analytics, and Creative Cloud plugins each depend on a remote HTTP MCP server (aa-mcp.adobe.io, cja-mcp.adobe.io, adobe-creativity.adobe.io) that requires IMS auth headers (Authorization, x-gw-ims-org-id, x-gw-ims-user-id). The README states an OAuth proxy may inject these. Probing all three servers without credentials returned HTTP 401 for Analytics (with a clean "Missing or empty Authorization token" body), HTTP 401 for CJA, and HTTP 403 for Creative Cloud. The servers exist and respond correctly, but a reader without an Adobe enterprise entitlement cannot exercise the analytics, CJA, or creative-cloud skills at all. That is roughly 18 of the 143 skills (the five AA, six CJA, and seven creative-cloud skills) plus their MCP-backed workflows. This is not a defect: it is the expected shape of vendor-gated MCP skills. But it does mean the "143 skills" headline overstates what a non-Adobe reader can run.

The package.json npm package name is @anthropic-ai/skills, which is a copy-paste drift from Anthropic's skills template. The description field correctly reads "Adobe skills for AI coding agents" and the repository URL points at github.com/adobe/skills, but the name field was never updated when the repo was forked or scaffolded from the Anthropic template. This is a metadata-only issue that does not affect functionality (the package is marked "private": true and is never published to npm), but it is the kind of provenance sloppiness that erodes trust in a first-party vendor repo. A reader auditing the supply chain sees an Adobe repo claiming to be an Anthropic package.

The stardust plugin depends on an external skill (impeccable) that is not vendored, which creates an undeclared install dependency. Stardust's pipeline explicitly delegates creative rendering to $impeccable craft and frames itself as "a higher-level skill built on top of impeccable". The impeccable repo (github.com/pbakaus/impeccable) is referenced in the stardust README but is not installed by any of the documented install commands. A user who installs stardust via /plugin install stardust@adobe-skills gets the stardust skills without impeccable, and the prototype and migrate stages will fail or degrade until impeccable is separately installed. The marketplace.json does declare allowCrossMarketplaceDependenciesOn: ["impeccable"], acknowledging the cross-marketplace dependency at the manifest level, but the README install section does not surface it as a required step.

Cursor support is explicitly a preview limited to one plugin, despite the repo shipping a root .cursor-plugin/marketplace.json. The app-builder plugin is the only one with a Cursor-native manifest (plugins/app-builder/.cursor-plugin/plugin.json), and the README calls it "the pilot for Cursor distribution. Other plugins will gain Cursor support once the pattern is validated." A Cursor user who sees the marketplace.json and expects all ten plugins to install will find only six App Builder skills available. The Claude Code plugin marketplace is the first-class distribution path; Cursor is aspirational for everything except app-builder.

The shallow-clone git history shows a single squashed commit, so repo maturity has to be inferred from GitHub metadata rather than the commit log. The GitHub API confirms the repo was created 2026-02-05, has 22 open issues, 59 forks, and was pushed the day of this review, which together indicate an active five-month-old project. But a reviewer or downstream auditor cloning the repo sees one squashed merge commit ("Merge pull request #186") and loses the per-skill authorship and change history. This is common for squashed-merge workflows and not a defect, but it limits how much you can learn about contribution cadence and individual skill ownership from the repo alone.

Smoke Test Results

Tested on macOS host (aarch64), Node v24.13.1, npm 11.8.0. The repo was shallow-cloned to 16 MB. The skills-ref reference validator (v0.1.5, the same tool the repo's CI uses) was installed into an isolated prefix and run against every skill directory. The Vercel skills CLI (npx skills add) was exercised against a single skill in a clean target directory. The three MCP servers were probed over HTTPS without credentials to confirm reachability and auth behavior.

Structural validation (all 143 skills against the agentskills.io spec)

$ npm install --no-audit --no-fund --prefix /tmp/adobe-val skills-ref@0.1.5

added 4 packages in 979ms

$ VAL=/tmp/adobe-val/node_modules/.bin/skills-ref

$ for d in $(find plugins -name SKILL.md -exec dirname {} \;); do

"$VAL" validate "$d"

done

Valid skill: plugins/adobe-analytics/skills/aa-kpi-pulse

Valid skill: plugins/adobe-analytics/skills/aa-top-movers-watchlist

... (140 more)

Valid skill: plugins/web/skills/page-tree

$ PASS=143 FAIL=0 of 143

PASS (every skill directory validates against the spec; frontmatter, structure, and references all clean)

Pass rate: 143 of 143. This is the canonical structural verdict. Zero skills failed validation, which means every SKILL.md has valid frontmatter, every referenced file resolves, and every skill conforms to the agentskills.io structure. The CI workflow runs this same check on every push and PR, so this is not a one-time snapshot.

Run B. Install path (does `npx skills add` deploy skills correctly?)

$ mkdir -p /tmp/adobe-install-test2/.claude/skills && cd /tmp/adobe-install-test2

$ npx -y skills@latest add /tmp/adobe-skills-review-20260707 --skill page-tree -a claude-code -y

◇ Installation Summary

│ ./.agents/skills/page-tree copy → Claude Code

◇ Installation complete

◇ Installed 1 skill

│ ✓ page-tree (copied) → ./.claude/skills/page-tree

│ Done! Review skills before use; they run with full agent permissions.

$ find .claude/skills -type f

.claude/skills/page-tree/SKILL.md

.claude/skills/page-tree/scripts/page-tree-bundle.js

.claude/skills/page-tree/evals/evals.json

.claude/skills/page-tree/package.json

.claude/skills/page-tree/.releaserc.json

file count: 5

PASS (all five accompanying files deployed, including the 606-line bundle script and evals)

Pass rate: 1 of 1 exercised. The documented Vercel skills install path works end to end. The skill deploys with its scripts, evals, and metadata intact. The same path works for --all to install every skill, though that was not exercised here to keep the test target clean.

Functional verification (MCP server reachability)

$ curl -s -m 12 -o /tmp/aa.txt -w "HTTP %{http_code}\n" https://aa-mcp.adobe.io/mcp

HTTP 401

$ head -c 200 /tmp/aa.txt

Missing or empty Authorization token

$ curl -s -m 12 -o /tmp/cja.txt -w "HTTP %{http_code}\n" https://cja-mcp.adobe.io/mcp

HTTP 401

$ curl -s -m 12 -o /tmp/cc.txt -w "HTTP %{http_code}\n" https://adobe-creativity.adobe.io/mcp

HTTP 403

PASS (all three servers reachable; clean auth errors confirm they are live and expect IMS credentials)

Functional pass rate: 1 of 1 exercised. The MCP servers exist and respond with structured auth errors rather than timeouts or 404s, which confirms they are production endpoints. Functional tool dispatch could not be tested because it requires Adobe IMS OAuth credentials.

What the runs tell you

The structural integrity is perfect (143 of 143), the install path is verified with real files on disk, and the MCP servers are confirmed live and auth-gated. The limitation is that the deepest claim, functional tool dispatch against live Adobe products, cannot be exercised without an Adobe enterprise entitlement, so the analytics, CJA, and creative-cloud skills are taken on the strength of their structural validation and the clean auth behavior of their servers rather than a live tool call. The vendor-neutral web and stardust plugins did not get a live agent workout either, but their bundled scripts and evals were inspected directly and are real code with real test criteria.

Setup Walkthrough

1. Install Node 20 or later. The repo's CI uses Node 24.
2. Pick your install path. For Claude Code, run /plugin marketplace add adobe/skills, then /plugin install @adobe-skills for each plugin you want. For the Vercel skills CLI, run npx skills add adobe/skills --all to get everything, or --skill for a single skill. For the upskill GitHub CLI extension, run gh extension install ai-ecoverse/gh-upskill then gh upskill adobe/skills --all.
3. If you want the analytics, CJA, or creative-cloud plugins to actually do anything, you need Adobe IMS OAuth credentials. The skills expect Authorization, x-gw-ims-org-id, and x-gw-ims-user-id headers, injected either directly or through an OAuth proxy. Without these, the three MCP servers return 401 or 403 on every call.
4. For the web plugin, install playwright-cli and have it on PATH. The page-tree, page-reduce, and page-collect skills inject bundled JavaScript over CDP and will not work without it.
5. For the stardust plugin, separately install the impeccable skill (github.com/pbakaus/impeccable). Stardust delegates creative rendering to $impeccable craft and will degrade without it, even though the README install steps do not mention this.
6. For the AEM plugins, you need a target AEM environment (6.5 LTS, Cloud Service, or an Edge Delivery Services repo). The skills assume you are operating inside that environment.

Gotcha: the root package.json declares the npm name as @anthropic-ai/skills, which is a leftover from the Anthropic skills template. It is marked "private": true and never published, so it does not affect installation, but do not be confused if you see it in the source.

Alternatives

1. agentskills/agentskills (the spec and reference library) - the open standard this pack is built on, plus the skills-ref validator Adobe's CI uses. Prefer it if you want to understand the skill format or build your own conforming skills rather than consume Adobe's.
2. anthropics/skills - the official Anthropic skill pack (the template adobe/skills was scaffolded from). Prefer it if you want vendor-neutral, Anthropic-product-focused skills with no Adobe dependencies.
3. NVIDIA/skills - the official NVIDIA skill pack for NVIDIA-product workflows. Prefer it if your stack is NVIDIA (CUDA, NIM, TensorRT) rather than Adobe. Same official-vendor pattern, different product surface.
4. obra/superpowers (plus impeccable, for stardust) - the community brainstorming and design methodology skills that stardust builds on. Prefer them directly if you want the methodology without the AEM delivery half.