The Best Claude Skills to Install in 2026

This is the skills-only companion to our mixed starter kit of agents, skills, and commands. Everything here is a SKILL.md from the AgentsCamp library: a procedure that sits dormant until your task matches its description, then loads and runs like a checklist Claude actually follows. New to the format? Start with What Are Claude Skills?

What earns a slot on this list: the skill encodes a procedure (not trivia the model already knows), the job recurs, and getting it wrong by hand is expensive. Here's the list, by the job each one does.

The git loop (install these first)

The highest-frequency wins, because you commit more often than you do anything else:

• conventional-commits — reads staged changes and writes a correct Conventional Commits message. The gateway skill: one job, fires on "commit this," never needs babysitting.
• pr-description — turns a branch diff into a PR description reviewers actually read.
• branch-rebaser — rebases onto main and walks each conflict by understanding both sides instead of clobbering one.
• commit-splitter — splits a mixed working tree into atomic commits that each build and pass tests.
• git-blame-investigator — reconstructs why a line exists from history before you delete it. Cheap insurance against re-introducing the bug the line was fixing.

Testing and review

• test-scaffolder — scaffolds test files that match your project's framework and conventions instead of a generic template.
• coverage-gap-finder — runs your coverage tool and names the highest-value untested paths, not just the percentage.
• mutation-test-runner — answers the question coverage can't: would the suite actually catch a bug?
• integration-test-designer — designs tests against real collaborators at a deliberate seam, for the bugs that slip past mocked-everything unit suites.

Databases and SQL

The specialist tier: rarely triggered, disproportionately valuable when they are.

• migration-writer — expand-contract migrations that stay reversible and never lock a live table. The single most "I should not free-hand this" job in the list.
• query-plan-analyzer — reads EXPLAIN output and turns it into a concrete fix.
• postgres-index-strategist — picks the right index type for the query shape instead of defaulting to btree-everything.
• deadlock-diagnoser — reconstructs the lock cycle from the engine's own deadlock report and fixes the ordering.

Performance

• memory-leak-hunter — heap-snapshot diffing to name the growing object and cut its retention path.
• flamegraph-analyzer — reads a CPU profile correctly (width is time, depth isn't) and fixes the widest leaf.
• web-vitals-optimizer — treats field data at p75 as truth and applies one targeted fix per metric.
• bundle-analyzer — ranks the JS bundle's biggest size wins instead of guessing.

Security

• secret-scanner — scans for committed keys and tokens before they reach a public remote. Pair it with a pre-push hook.
• dependency-audit — turns noisy scanner output into a triaged, prioritized upgrade plan.
• threat-model-builder — builds a STRIDE model from the actual architecture, not a template.

LLM and RAG work

If you build AI features, these encode the eval-first discipline that separates working pipelines from demos:

• prompt-optimizer — diagnoses why a prompt underperforms before rewriting it, and tells you what to measure.
• llm-eval-suite-scaffolder — stands up a real eval suite (dataset, metrics, baseline, CI gate) for a feature that has none.
• chunking-strategy-optimizer — sweeps chunk configurations against a fixed eval set instead of guessing at 512-with-overlap.
• hallucination-evaluator — decomposes answers into atomic claims and checks each for grounding.

Skills that extend Claude Code itself

The compounding tier — each one makes every future session better:

• hook-writer — turns "always format after edits" into an enforced hook with the right event, matcher, and exit-code plumbing.
• claude-settings-auditor — audits every settings layer before you trust a new repo's checked-in config.
• plugin-scaffolder — packages your accumulated skills, agents, and commands into one installable plugin for the team.
• agent-memory-designer — writes a CLAUDE.md from the repo's real commands and gotchas, sized to stay in context.

How to choose from the other sixty

The full library is at /skills — 90+ skills across twelve categories. The selection rule that holds up: install for friction you've felt twice, and check the new skill's description doesn't overlap one you already have. Overlapping descriptions are the main failure mode — Claude routes by description, and two skills claiming the same job means the wrong one fires half the time. When you outgrow the library versions, edit them — they're plain Markdown, and the frontmatter description is the tuning knob.