Skills I Use Claude Code reference Checked against source, 11 November 2026

Skills I Use

Five Claude Code skills, how to install each, what each is good for, and one prompt that shows it working.

A skill is a folder of instructions and scripts that Claude Code loads when a task calls for it. Personal skills live in ~/.claude/skills/, project skills in .claude/skills/, and skills that ship as plugins are managed with /plugin.

Quick install

The main route for each skill comes first, then alternatives and setup. Skills can run code on your machine, so read this before installing.

Impeccable
  • CLInpx impeccable installShell
Karpathy Guidelines
  • Plugin 1/plugin marketplace add forrestchang/andrej-karpathy-skillsClaude Code
  • Plugin 2/plugin install andrej-karpathy-skills@karpathy-skillsClaude Code
Superpowers
  • Official/plugin install superpowers@claude-plugins-officialClaude Code
Taste Skill
  • CLInpx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"Shell
Trail of Bits Skills
  • Step 1/plugin marketplace add trailofbits/skillsClaude Code
  • Step 2/plugin menuClaude Code

Other routes, setup and updates

Impeccable
  • Plugin 1/plugin marketplace add pbakaus/impeccableClaude Code
  • Plugin 2/pluginchoose Discover, then ImpeccableClaude Code
  • Setup/impeccable initClaude Code
  • Setup/impeccable hooks onClaude Code
  • Updatenpx impeccable updateShell
Karpathy Guidelines
  • Filecurl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.mdReplaces any existing CLAUDE.md. Use Append if you have one.Shell
  • Appendecho "" >> CLAUDE.md curl https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md >> CLAUDE.mdShell
Superpowers
  • Author 1/plugin marketplace add obra/superpowers-marketplaceClaude Code
  • Author 2/plugin install superpowers@superpowers-marketplaceClaude Code
Taste Skill
  • Plugin 1/plugin marketplace add Leonxlnx/taste-skillClaude Code
  • Plugin 2/plugin install taste-skill@taste-skillClaude Code
  • Classicnpx skills add https://github.com/Leonxlnx/taste-skill-classics --skill "high-end-visual-design"Shell

Impeccable

impeccable.style

Gives the agent a designer's vocabulary and a set of anti-pattern filters, so the UI it produces reflects decisions instead of defaults.

Install

  • CLInpx impeccable installShell
  • Plugin 1/plugin marketplace add pbakaus/impeccableClaude Code
  • Plugin 2/pluginchoose Discover, then ImpeccableClaude Code
  • Setup/impeccable initClaude Code
  • Setup/impeccable hooks onClaude Code
  • Updatenpx impeccable updateShell

Reload Claude Code after installing. init interviews you once and writes PRODUCT.md. hooks on runs a detector on UI files after each edit.

Good for

  • One kind of change at a time. Each command names a design job, so you can ask for typography or layout without reopening everything else.
  • Stated context. Later commands read who the product is for from PRODUCT.md instead of guessing.
  • Review before rework. critique scores a page against UX heuristics. audit checks accessibility, performance and responsive behaviour.
  • Checks while you edit. With hooks on, problems such as low-contrast text are reported right after the file is written.

Commands by group

GroupCommands
Createshape
Evaluatecritique audit
Refinetypeset layout colorize animate delight bolder quieter overdrive
Simplifydistill clarify adapt
Hardenpolish optimize harden onboard
Systeminit extract document live generate

Example

/impeccable critique the pricing page
What happens

In a Next.js app it finds app/pricing/page.tsx and the components it renders, then reviews the page as a designer would. You get scores per heuristic and a prioritised list: for example, three plans with equal visual weight so no plan reads as the recommended one, or a comparison table that hides the differences people choose on. Nothing is changed yet. Follow up with /impeccable layout or /impeccable polish on the items you agree with.

Karpathy Guidelines

github.com/multica-ai/andrej-karpathy-skills

Four working rules, drawn from Andrej Karpathy's observations on how LLMs get coding wrong, that keep the agent from guessing, over-building and wandering.

Install

  • Plugin 1/plugin marketplace add forrestchang/andrej-karpathy-skillsClaude Code
  • Plugin 2/plugin install andrej-karpathy-skills@karpathy-skillsClaude Code
  • Filecurl -o CLAUDE.md https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.mdReplaces any existing CLAUDE.md. Use Append if you have one.Shell
  • Appendecho "" >> CLAUDE.md curl https://raw.githubusercontent.com/forrestchang/andrej-karpathy-skills/main/CLAUDE.md >> CLAUDE.mdShell

Pick one route. The plugin applies across all projects. The curl route writes the same rules into one project's CLAUDE.md; use the append form if the file already exists. The repository now lives under multica-ai. Its README still uses the forrestchang path, which GitHub redirects.

Good for

  • Wrong guesses surface early. Assumptions are written down before the code, so you correct a sentence and not a diff.
  • Smaller diffs. No speculative abstractions, no options nobody asked for.
  • Changes stay in scope. Neighbouring code, comments and formatting are left alone. Unrelated dead code gets mentioned, not deleted.
  • Done means checked. Work is tied to something that can pass or fail, often a test written first.

The four principles

PrincipleIn practice
Think before codingState assumptions, surface tradeoffs, ask when the request is unclear.
Simplicity firstWrite the minimum code that solves the problem. Nothing speculative.
Surgical changesTouch only what the task needs and match the existing style.
Goal-driven executionDefine a verifiable goal, write the failing test first, iterate until it passes.

Example

Orders with a 100% discount code return a 500 from OrderSerializer.validate. Fix it, failing test first.
What happens

In a Django REST project it reads the serializer and states what it believes the cause is before editing. If it is unclear whether a zero total should be allowed, it asks. It then adds one failing test to the existing orders test module, makes the smallest change in validate that turns it green, and reports the result. It does not reformat the file or refactor the discount model along the way.

Superpowers

github.com/obra/superpowers

A full engineering workflow for the agent: clarify, plan, test first, review, finish. It replaces jumping straight to code.

Install

  • Official/plugin install superpowers@claude-plugins-officialClaude Code
  • Author 1/plugin marketplace add obra/superpowers-marketplaceClaude Code
  • Author 2/plugin install superpowers@superpowers-marketplaceClaude Code

Official uses Anthropic's marketplace. Author 1 and 2 install the same plugin from the author's own marketplace. Pick one route.

Good for

  • Questions before code. Brainstorming asks what you want and shows a design to confirm.
  • Plans you can read. Work is split into small tasks, each with file paths and a verification step.
  • Tests first. A failing test comes before the code that makes it pass.
  • Isolated work. Changes happen in a git worktree, with a review between tasks. Critical issues block progress.
  • A clean ending. The finish step checks the tests, then offers merge, pull request, keep or discard.

Key skills

SkillWhat it does
brainstormingAsks what you want and presents a design for approval.
using-git-worktreesCreates an isolated branch and checks the tests pass before any change.
writing-plansBreaks the design into small tasks with paths, code and checks.
executing-plansCarries out the plan inline in one session, with a final review.
subagent-driven-developmentCarries out the plan with a fresh subagent per task and a review after each.
test-driven-developmentFailing test first, then the minimal code to pass it.
systematic-debuggingFinds the root cause before proposing a fix.
verification-before-completionRuns the checks before claiming the work is done.
requesting-code-reviewReviews the work against the plan between tasks.
finishing-a-development-branchVerifies tests, offers merge or pull request, cleans up the worktree.

Example

Add team invitations to this Laravel app. An owner invites by email, the invite expires after 7 days, and accepting it joins the team.
What happens
  1. Brainstorming starts before any code. Expect questions such as what happens when the email already has an account, and whether invites can be resent.
  2. You approve a short design. It creates a worktree and confirms the existing tests pass.
  3. The plan lists tasks: migration, Invitation model, mailable, controller, policy. Each one is built with a failing feature test first.
  4. A review runs between tasks. At the end it verifies the suite and offers a merge or a pull request.

Taste Skill

github.com/Leonxlnx/taste-skill

Anti-slop rules for frontend work: one consistent design system, real imagery, and a browser check before the agent calls a page done.

Install

  • CLInpx skills add https://github.com/Leonxlnx/taste-skill --skill "design-taste-frontend"Shell
  • Plugin 1/plugin marketplace add Leonxlnx/taste-skillClaude Code
  • Plugin 2/plugin install taste-skill@taste-skillClaude Code
  • Classicnpx skills add https://github.com/Leonxlnx/taste-skill-classics --skill "high-end-visual-design"Shell

The main repository now holds one skill, design-taste-frontend. The older style skills, including high-end-visual-design, minimalist-ui and industrial-brutalist-ui, moved to Leonxlnx/taste-skill-classics and kept their install names. The last row shows how to install one of them.

Good for

  • A direction up front. The agent commits to one design system and applies it across the page.
  • Real material. It calls for real photography and brand icons where a placeholder would otherwise go.
  • Redesigns that start by looking. An existing page is audited in a real browser before anything changes.
  • A checker with a verdict. scripts/consistency.mjs renders the page in headless Chrome at 1440, 1024, 820, 768 and 390 pixels and prints PASS, WARN or FAIL per rule.
  • Pairs with Impeccable. Taste sets the direction. Impeccable refines and audits the result.

Example

Redesign the marketing homepage in app/(marketing)/page.tsx. Audit the current page in the browser first. Keep the copy and our Tailwind tokens.
What happens
  1. It loads the running Next.js page in headless Chrome and reports what is inconsistent today: spacing, type sizes, mixed icon styles.
  2. It proposes one direction and rebuilds the page inside your existing tokens.
  3. It reruns the checker at all five widths and fixes what fails.
  4. From here, /impeccable audit is a good second opinion on accessibility and responsive behaviour.

Trail of Bits Skills

github.com/trailofbits/skills

Security review skills written by a professional audit firm, run from inside Claude Code.

Install

  • Step 1/plugin marketplace add trailofbits/skillsClaude Code
  • Step 2/plugin menuClaude Code

The marketplace is a collection. Use the menu to pick only the plugins you need.

Good for

  • Diff review with history. Changes are read the way an auditor reads them, using git history for context.
  • Fail-open defaults. Settings that quietly turn protection off are found, and a verifier tries to refute each finding.
  • Dependency risk beyond CVEs. Abandoned upstreams and install scripts are flagged alongside matched advisories.
  • One bug becomes a search. Variant analysis looks for the same mistake elsewhere in the codebase.

Plugins for web and backend work

PluginWhat it does
differential-reviewSecurity-focused review of code changes, using git history.
insecure-defaultsParallel audit for fail-open defaults, with a refuting verifier per file.
sharp-edgesFlags error-prone APIs, dangerous configurations and footgun designs.
supply-chain-risk-auditorAudits npm, PyPI and Go dependencies for advisories, abandoned upstreams and install scripts.
static-analysisSemgrep and CodeQL, plus SARIF parsing.
variant-analysisFinds similar vulnerabilities across a codebase by pattern.
property-based-testingWrites and reviews property tests for Hypothesis, fast-check and proptest.
modern-pythonCurrent Python tooling and practice with uv, ruff and pytest.

The collection also covers smart contracts, reverse engineering, and C, C++ and Rust review.

Example

Audit this checkout API diff for insecure defaults: git diff main...feature/checkout-api
What happens

With insecure-defaults installed, it reads the changed files in a Django project and looks for places where a missing value turns a protection off: a payment webhook that skips signature checks when the secret is unset, a SECRET_KEY with a hard-coded fallback, DEBUG that defaults to true. A verifier then tries to refute each finding. What survives comes back with file, line and a suggested fix.

How they stack

The five do not compete. Two shape how code gets written, two shape how the UI looks, and one checks what ships.

How the five skills stack A route from task to shipped with three stops. Superpowers and Karpathy Guidelines serve the first stop, write the code. Taste Skill and Impeccable serve the second, shape the UI. Trail of Bits Skills serves the third, check what ships. TaskShipped Write the code Shape the UI Check what ships 3Superpowers 2Karpathy Guidelines 4Taste Skill 1Impeccable 5Trail of Bits Skills How the five skills stack: write the code with Superpowers and Karpathy Guidelines, shape the UI with Taste Skill and Impeccable, check what ships with Trail of Bits Skills. TaskShipped Write the code Shape the UI Check what ships 3Superpowers 2Karpathy Guidelines 4Taste Skill 1Impeccable 5Trail of Bits Skills
  1. Write the codeSuperpowers supplies the process: clarify, plan, test first, review. Karpathy Guidelines supplies the restraint inside each step: small diffs, stated assumptions.
  2. Shape the UITaste Skill sets the direction and checks it in a browser. Impeccable then critiques, refines and audits what was built.
  3. Check what shipsTrail of Bits Skills reviews the diff, the defaults and the dependencies before the branch merges.

Before you install

A skill is not only text. It can ship scripts, and a plugin can add hooks that run on their own. All of it runs with your permissions, on your machine, inside your repositories.

  • Read the skill's SKILL.md and every script it ships before installing it.
  • For plugins, check which hooks and MCP servers they register.
  • Read the changes again when you update. An update is new code.
  • Try a new skill in one project's .claude/skills/ before installing it globally.