10 Claude Code Skills That Save Developers Hours Every Week
Ten SKILL.md recipes for the work you already repeat: reviewing diffs, writing tests, shipping PRs, auditing pages, and drafting content. Each one shows when to invoke it and how the description should read.
Most developers don’t need more AI. They need the same five or six procedures to stop living in their clipboard.
That’s what a skill is for. The Claude Code docs say to create one when you keep pasting the same instructions, checklist, or multi-step procedure into chat — or when a section of CLAUDE.md has grown into a procedure rather than a fact. Unlike always-on instructions, a skill’s body loads only when it’s used.
This post is a shortlist of ten skills that earn their keep in a real week. Each one is a folder with a SKILL.md, following the Agent Skills shape (name, description, Markdown body). They work the same way in Claude Code, and the same files load in Cursor and Codex with only the install path changing. If you’re still writing your first file, start with How to Write Your First SKILL.md. If a skill you already wrote never fires, use the troubleshooting checklist.
I’m not going to invent “hours saved” numbers. I’ll describe the workflow each skill replaces, show a description that actually matches how people ask, and note the frontmatter tricks that matter.
How to use this list
- Put project skills in
.claude/skills/<name>/SKILL.md(Claude Code),.cursor/skills/or.agents/skills/(Cursor), or.agents/skills/(Codex). See the install guide for the full table. - Keep the description specific. At startup, tools load only each skill’s name and description. The specification frames this as progressive disclosure: metadata first, full body when activated, supporting files on demand.
- Use
/skill-name(Claude Code / Cursor) or$skill-name(Codex) when you want the skill every time, not only when the agent guesses right. - Side-effect workflows (commit, deploy, send a message) should set
disable-model-invocation: trueso the agent waits for you. Claude Code and Cursor both support that flag.
1. Summarize the uncommitted diff
Workflow it replaces: Opening a terminal, running git diff, pasting half of it into chat, asking “is this safe to commit?”
Claude Code’s own getting-started example does this with dynamic context injection: a !git diff HEAD`` line runs before the skill content reaches the model, so the summary is grounded in the live tree rather than guesswork from open files.
---
name: summarize-changes
description: Summarizes uncommitted changes and flags anything risky. Use when the user asks what changed, wants a commit message, or asks to review their diff.
---
Keep the body short: summarize in two or three bullets, then list risks (missing error handling, hardcoded secrets, tests that need updating). Invoke it with /summarize-changes, or let the description catch “what did I change?”
2. Review a pull request like a teammate
Workflow it replaces: Opening the GitHub diff in a browser tab, skimming files out of order, forgetting the test plan.
Claude Code ships a bundled /code-review skill. For a custom version tuned to your repo, fork the idea: pull the live PR into the prompt, then apply your team’s checklist.
---
name: pr-review
description: Reviews the open pull request against our checklist (tests, migrations, logging, secrets). Use when the user asks to review a PR, check a diff before merge, or wants a ship/no-ship call.
context: fork
agent: Explore
allowed-tools: Bash(gh *)
---
The body can inject !gh pr diff, `!`gh pr view --comments, and !gh pr diff –name-only`` the way the Claude Code docs show for a PR-summary skill. context: fork runs that work in an isolated subagent so your main chat stays clean. Add your real checklist under the injections — naming conventions, “no new any in TypeScript,” “migrations must be reversible,” whatever your team actually enforces.
3. Write tests for the files you just touched
Workflow it replaces: Finishing a feature, staring at the test folder, asking the agent for “some tests” without telling it which files matter.
---
name: write-tests-for-diff
description: Writes unit or integration tests for files changed on this branch. Use when the user asks for tests for their changes, coverage for a PR, or to fill gaps after a feature.
---
In the body: list the changed paths (!git diff –name-only origin/main…HEAD``), tell the agent which test runner the repo uses, require one failing-then-passing cycle per case, and forbid inventing helpers that don’t exist. Scope with paths if you only want this skill near src/** or *.test.ts — both Claude Code and Cursor support a paths frontmatter field for that.
4. Draft the PR title and description
Workflow it replaces: Squashing a day of commits into a blank GitHub form at 6pm.
---
name: pr-description
description: Writes a pull request title and description from the current branch's commits and diff. Use when the user asks for a PR description, PR summary, or help opening a pull request.
disable-model-invocation: true
---
disable-model-invocation: true keeps the agent from volunteering a PR write-up mid-refactor. You type /pr-description when you’re ready. Body recipe: summary, test plan, risk notes, and a “how to verify” section. Pull commits and the diff with injected shell commands so the text matches the branch, not the conversation.
5. Ship release notes from merged PRs
Workflow it replaces: Scrolling merged PRs since the last tag and rewriting each title by hand.
---
name: release-notes
description: Writes release notes from merged pull requests since the last git tag. Use when the user asks for release notes, a changelog entry, or a ship summary.
disable-model-invocation: true
---
Have the skill list tags (!git tag –sort=-creatordate | head -5``), then the merged PRs since the previous tag via gh. Group by feature / fix / chore. Don’t invent version numbers the repo hasn’t tagged.
6. Commit only when you say so
Workflow it replaces: The agent “helpfully” staging everything, including a .env you didn’t mean to touch.
---
name: commit
description: Stages and commits the current changes with a conventional message. Use only when the user explicitly asks to commit.
disable-model-invocation: true
allowed-tools: Bash(git add *) Bash(git commit *) Bash(git status *) Bash(git diff *)
---
Claude Code’s docs use almost this exact pattern for a commit skill: slash-only invocation, plus allowed-tools so git commands don’t prompt on every file during that turn. The grant clears on your next message. Put the conventional-commit rules in the body, and tell the agent to stop if git status shows secrets or generated junk.
If you keep a project verify skill, Claude Code’s commit instructions (v2.1.286+) tell Claude to run it before each commit when that skill is present and invocable — useful once you’ve recorded how to build and smoke-test the app.
7. Audit page meta before a content deploy
Workflow it replaces: Publishing a blog post, then discovering the title is wrong, the canonical points at staging, or noindex disappeared.
---
name: page-meta-audit
description: Checks title, description, canonical, robots/noindex, and Open Graph tags on a built HTML page or live URL. Use when the user asks to audit page meta, check noindex, or verify SEO tags before deploy.
---
Point the body at your build output (dist/blog/<slug>/index.html) and your known invariants. On skillaisle that means confirming the page still carries a noindex robots meta — the site is content-only through the end of 2026, so flipping indexability is a bug, not a feature. For the same checklist packaged as a ready-made skill, start from Page Meta Check in the Skillaisle catalog, then adapt the thresholds to your stack.
8. Draft a blog post from a calendar brief
Workflow it replaces: Turning a Notion row (“keyword, angle, related posts”) into a blank editor and losing an afternoon to structure.
---
name: blog-post-draft
description: Drafts an English technical blog post from a title, angle, keyword, and related links. Use when the user asks to write a blog post, stage a calendar article, or turn a content brief into Markdown.
disable-model-invocation: true
---
Put your house rules in the body: frontmatter fields your Astro collection expects, target length, “cite sources with links — never invent data,” soft CTA only, no pricing. Reference sibling posts by path (/blog/write-first-skill-md/) so interlinks stay real. This is the skill behind posts like this one; the human still rewrites for voice.
9. Scaffold a travel itinerary page
Workflow it replaces: Copy-pasting yesterday’s Zion outline into a new San Francisco draft and fixing the day labels by hand.
---
name: itinerary-page
description: Scaffolds a multi-day travel itinerary page with overview, day sections, and logistics. Use when the user asks for an itinerary draft, a day-by-day trip page, or a travel outline for a park or city.
---
Body: require sources for hours, fees, and shuttle facts; leave blanks instead of guessing; keep day sections parallel (morning / afternoon / evening); link the overview to day posts. Supporting files under references/ can hold park-specific notes so the main SKILL.md stays under the ~500-line guidance Claude Code recommends.
10. Turn a finished article into a short video script
Workflow it replaces: Re-reading a 1,500-word post to produce a 60-second voiceover, then realizing the hook was buried in paragraph four.
---
name: video-script-from-post
description: Turns a published or drafted blog post into a short vertical-video script with hook, beats, and on-screen text. Use when the user asks for a Shorts script, Reel script, or voiceover from an article.
disable-model-invocation: true
---
Require the agent to quote the source post, keep claims that already have citations, and flag anything that would need a new source. Output: hook (first 3 seconds), three beats, closing CTA that matches the post (catalog or waitlist — not a paid plan). Pair it with your studio’s shot list format in assets/ if you have one.
Patterns worth stealing across all ten
Write the description for the request, not the architecture. Codex’s docs say the initial skills list uses at most 2% of the model context (or 8,000 characters when the window is unknown), and long descriptions get shortened first. Front-load trigger phrases so a truncated description still matches.
Prefer supporting files over a novel-length SKILL.md. The specification and Claude Code docs both recommend keeping the main file focused and moving reference material into references/, scripts/, and assets/. Those load on demand.
Separate “knowledge” skills from “action” skills. A conventions skill can stay auto-invocable. A deploy or commit skill should not. That’s the disable-model-invocation / Codex allow_implicit_invocation: false split — same idea, different field names.
Test in a fresh session. Leftover context from writing the skill hides gaps in the written instructions. Claude Code’s skill-creator plugin and claude plugin eval exist for exactly this comparison loop.
Start with one, not ten
Pick the workflow you pasted into chat twice this week. Write the folder, write a specific description, invoke it by hand once, then see if free-form requests pick it up. If they don’t, fix the description before you write skill number two — that’s the failure mode covered in Why Your AI Agent Ignores Your Skill.
When you want a known-good starting point instead of a blank SKILL.md, browse the Skillaisle catalog or join the waitlist for new packs. The formats above stay free to copy into your own repos.
Sources
- Claude Code docs: Extend Claude with skills
- Agent Skills: Specification
- Cursor docs: Agent Skills
- OpenAI docs: Build skills
Products change quickly. If something here no longer matches what you see, the official page above is the reference.
Related on Skillaisle
Want the setup steps? Browse the Learn guides or walk the skills aisle.