The Claude Code Skills ecosystem is in its gold-rush phase. As of May 2026, there are 15,134 indexed repos across the official Anthropic marketplace, ClaudePluginHub, Skills.sh, Agensi, LobeHub, and the awesome-agent-skills GitHub list. The official Anthropic frontend-design skill alone has crossed 277,000 installs. Most weeks ship at least one viral Karpathy-flavored CLAUDE.md drop that hits the front page of every AI newsletter.
But here's the dirty secret: most skills don't survive Opus 4.7. They were written for 4.5 or 4.6 — before the new tokenizer, before file-system memory, before the 1M context window. They worked, but their prompts assume a model that's been retired. On 4.7 they either misfire or just sit there bloating your context budget.
We run AgentHive's 36 shipped AI apps on Claude Code. Every skill on this list has been audited against 4.7 in the last 30 days. Every "skip" note is something we actually tripped on.
How we ranked them
Five criteria, no fluff:
- Frequency of invocation. Does it fire often enough to justify the context cost? Skills that load on every session but only help on 2% of tasks are net-negative.
- Output durability. Does the work survive review by a senior engineer, or does it produce LLM-flavored noise that you have to rewrite anyway?
- Opus 4.7 alignment. Does the skill exploit 4.7's new capabilities — file-system memory, 1M context, the new tokenizer's token economics — or fight them?
- Maintenance velocity. Has the author shipped an update in the last 90 days? Stale skills age fast.
- Composition behavior. Does it play nicely when stacked with other skills? Anthropic's playbook draws the line at 8–12 skills before performance degradation shows up.
The 12 skills that survived
1. code-reviewer
The most-installed skill on Agensi (116 installs and climbing). Reviews the current diff and surfaces correctness bugs at low/medium/high effort levels. On Opus 4.7's expanded context, it can hold the entire diff plus 4–5 dependency files at once, which is where it shines vs. earlier models.
Install when: you ship more than one PR per week. Skip if: your team already has a serious human reviewer cadence — this duplicates effort without adding much. Pair it with --effort high only when stakes are high; default medium is the sweet spot.
2. git-commit-writer
65 installs. Reads git diff and drafts a commit message that follows your repo's existing style (it learns from the last 20 commits). On Opus 4.7, the output is notably better at picking the right verb — "add" for genuinely new features, "fix" for bugs, "refactor" only when behavior is preserved — which earlier versions confused constantly.
Install when: you commit daily and care about a clean history. Skip if: you squash everything to a single PR-level commit on merge — the upstream review tool's title will do the job.
3. pr-description-writer
36 installs. Reads the branch diff and drafts a PR description with summary + test plan. The skill that pays for itself the fastest, because writing PR descriptions is the universally hated step.
Install when: you open more than two PRs a week. Skip if: your repo has a strict PR template that diverges from the skill's default — you'll spend more time editing than writing.
4. env-doctor
30 installs. Diagnoses broken local environments — missing env vars, mismatched node/python versions, port conflicts, broken docker compose graphs. On Opus 4.7's file-system memory, it remembers the fixes between sessions, so the second time you hit "ELIFECYCLE error" it surfaces the prior solution instantly.
Install when: you maintain more than one codebase. Skip if: you only work in one tightly-controlled container; you'll never trigger it.
5. changelog-generator
27 installs. Generates a Keep-a-Changelog-formatted entry from the commit range between two tags. Worth its weight in gold on release days. On Opus 4.7, it now groups changes by user-facing impact rather than file path, which is the correct framing.
Install when: you ship versioned releases. Skip if: you continuously deploy without versioning — no changelog to generate.
6. seo-optimizer
21 installs. Audits a webpage's HTML and proposes title, meta description, heading structure, and schema.org markup improvements. The skill that prompted us to write this very article — and the home page you're reading from.
Install when: you ship marketing content or product pages. Skip if: your product is a Slack-distributed B2B tool — your traffic doesn't come from organic search.
7. readme-generator
49 installs. Auto-generates a README for a new repo from its file tree, package.json, and example test files. The Opus 4.7 version is notably better at distinguishing "what this does" from "how to install" — earlier models conflated them constantly.
Install when: you start new repos often (consultants, agency teams). Skip if: you only maintain one or two long-lived projects.
8. test-generator
Anthropic's official, ~85k installs. Generates unit tests from a target function. The Opus 4.7 version now writes tests against the function's contract rather than its implementation, which is the difference between tests that survive refactors and tests that lock the code in place.
Install when: you ship anything to production. Skip if: the code under test has heavy mocked dependencies — generated tests will be brittle. Hand-write those.
9. security-scanner
Community skill that wraps the patterns from Anthropic's new Claude Security public beta and applies them to non-Enterprise tiers. Flags hardcoded secrets, prompt injection vectors, dangerous shell exec patterns, and broken access control.
Install when: you handle user input or run user code. Skip if: your entire stack is statically generated marketing HTML; you'll get noise.
10. doc-writer
Generates JSDoc / TSDoc / docstrings for a target file. The Opus 4.7 update is the first version where the docs are tight enough that we stopped deleting them. Earlier versions wrote multi-paragraph essays for a five-line helper.
Install when: you maintain a library or SDK that other teams consume. Skip if: the code is internal-only — docs rot faster than they help.
11. dependency-auditor
Reads package.json / requirements.txt / Cargo.toml and flags outdated, deprecated, or known-vulnerable dependencies, sorted by risk. The Opus 4.7 file-system memory means it now remembers which deps you've already chosen to keep pinned, so you stop seeing the same "upgrade lodash" suggestion every week.
Install when: you have any production system. Skip if: you already run Dependabot or Renovate religiously — overlap is total.
12. perf-profiler
Reads CI traces, runtime profiles, or N+1 query logs and proposes targeted fixes. The least-installed skill on this list (sub-1k installs), but the one with the highest hit rate when it does fire. Pairs especially well with the 1M context window because it can hold an entire flamegraph plus the call sites in one prompt.
Install when: you have a production system with measurable latency. Skip if: your stack is "static site + serverless functions" — you're not the customer.
The 3 most overhyped
Hit the front page of HN, lit up the AI Twitter feed, and then quietly stopped getting used. Sharing because the install counts mislead.
1. The viral "Karpathy CLAUDE.md" repos
When Karpathy posted his frustrations with LLM coding pitfalls in January 2026, developer Forrest Chang turned them into a 4-rule CLAUDE.md. The repo hit 144k stars in weeks. The rules are fine — but they encode Karpathy's specific 2025-era complaints, several of which Opus 4.7 actually fixed. The rule "don't generate fake APIs" is now redundant; 4.7's grounding is materially better. Install it as inspiration, then prune.
2. Mega-bundles ("177 production-ready skills!")
The 177-skill bundles look impressive in a tweet. In practice, 165 of the 177 skills never fire in your project, and the 12 that do compete with each other for invocation. Anthropic's own guidance is 8–12 skills max. The bundle vendors know this — they're selling FOMO.
3. "humanize-writing"-type skills
Skills that try to make Claude's output less "AI-sounding" almost always make it worse — they bias toward verbose, mealy-mouthed prose. If your concern is detection, you have a process problem, not a prompt problem. Skip.
Installation & management
The cleanest install path as of May 2026:
# From inside your project
claude /plugin install code-reviewer
claude /plugin install git-commit-writer
# ...etc
# Or drop the whole kit at once
unzip claudefarm-skill-toolkit.zip -d ~/.claude/skills/
# Verify
claude /plugin listThree management rules we follow:
- Audit quarterly. Remove any skill that hasn't fired in 30 days. Context budget is real money on a 1M-token model.
- Pin versions. Skills update silently when authors push to the marketplace. Pin specific versions in production setups.
- Watch the security checklist. Agensi scans every paid skill against an 8-point checklist — prompt injection, secret exfiltration, dangerous shell, etc. Apply the same checklist to free skills before installing.
The takeaway
Pick 8 skills. Pin them. Audit quarterly. Buy the bundle that already did the curation work instead of installing 47 and tuning your way out of context collisions.
If you're still piecing it together by hand, our Skill Farmer's Toolkit ships the 12 skills above as a single zip, pre-configured for Opus 4.7's tokenizer and memory behavior, with lifetime updates whenever Anthropic ships a model change. One payment. No subscription. The same kit our team runs.