The skills I use every day
Matheus Cardosoupdated
Agent skills I use every day in Claude Code. Half interface, half engineering, each with the exact values and the definition of done that a prompt from memory never has.
Installation
Two doors. The plugin installs the set as a managed bundle that updates when I ship. The clone puts the files where you can edit them. Pick one: installing both leaves you with every skill twice.
# plugin: atualiza junto com o repositório
/plugin marketplace add matheuscarddoso/skills
/plugin install mcardoso-skills@mcardoso
# ou clone, se você quer editar as skills
git clone https://github.com/matheuscarddoso/skills.git ~/Projects/skills
~/Projects/skills/scripts/installThen /ask and one sentence describing your situation. It decides which skill fits and calls it.
If you only adopt three, adopt /grill, /review and /designer. One stops you building the wrong thing, one stops you shipping it wrong, and the third stops you shipping it ugly.
Interface
Six specialists, one per discipline, that the agent reaches for while it builds. One review that runs through all of them. Five tools that produce rather than judge.
/designer
you call itReviews a screen, a flow, a branch or a pull request through every interface skill and returns one ranked verdict, with a file and line on every finding.
Looks good! A few suggestions:
- improve the contrast
- adjust some spacing
- maybe review the copy
/designer review the checkout, including the empty and error states/polish
the agent calls itFinish with exact values: concentric radius, shadows instead of borders, image outlines, optical alignment, a 0.96 press, icon swaps, enters, exits and gestures.
radius 12 inside and out · solid border · no image outline
Polish this card: the radius, the shadow and the button press/typography
the agent calls itHow text renders: choosing and loading fonts, the type scale, line height, balance and pretty wrapping, tabular numbers, truncation and punctuation.
The small details that make an interface feel right
text-wrap: wrap · proportional-nums
The heading leaves one word alone on the last line and the price jitters when it changes/color
the agent calls itThe color system: OKLCH ramps, role-based tokens, format conversion, gamut, contrast measured on the pair that actually renders, and dark mode.
Build the ramp for the brand color #0ea5e9 in OKLCH, with tokens for both themes/a11y
the agent calls itKeyboard and focus, accessible names, semantics, forms, screen readers, hit areas and reduced motion, against WCAG 2.2.
Does this sign-up form work with only a keyboard and a screen reader?/layout
the agent calls itGrouping, alignment, reading order, breakpoints that come from the content and room for translated text, so a layout holds up when it is resized, translated or mirrored.
The settings read like a shopping list. Group them without dividers everywhere/microcopy
the agent calls itInterface copy: labels that name the action, errors that say what to do, empty states that teach and confirmations that name the consequence.
Rewrite the delete-site dialog and the empty projects state/variant
you call itThree genuinely different versions of a component, on an axis you name, behind a switcher on the real page. You flip between them and promote the winner.
/variant the plan card, varying the density/teardown
you call itExplains how an interface, effect or animation on another site was built, separating what it measured from what it inferred, and ends with the recipe that carries over.
/teardown how is the copy button animation on interfaces.dev built?/break
you call itRenders a component under every scenario that can reach it in production, huge text, zero items, 320px, RTL, on a temporary page, and marks what broke.
/break the problem card/states
you call itA workbench with every state of a component, loading, empty, error, plan and permission, with mock data and a keyboard switcher, to work on one at a time.
/states the problems list/handoff
the agent calls itBuilds from a Figma link or a screenshot, with the project's own tokens and components, and compares side by side before calling it done.
Build this frame: figma.com/design/…?node-id=12-345Engineering
Align before building, build with tests, diagnose, audit and ship. None of them assumes a language: what is stack-specific lives in references.
/grill
you call itA hard interview before building, one question at a time, until no decision is left open.
1. Who can cancel: the customer, support, or both? I'd say both, with a required reason for support.
/grill I want to add subscriptions with recurring billing/kickoff
the agent calls itA new project: a stack with a written rationale, security from day one, reproducible infrastructure and the feedback loop before the first commit.
Stack Next.js + Postgres, because… Security headers, secrets out of the repo, CI with audit Loop lint + types + tests running before the 1st commit
Start a billing dashboard from scratch/tdd
the agent calls itRed, green, refactor, one vertical slice at a time.
✗ prorates a mid-cycle upgrade red, for the right reason ✓ prorates a mid-cycle upgrade green, the minimum ↺ extract daysLeft() refactor, still green
Build the proration math with TDD/investigate
the agent calls itHard bugs and performance regressions, in six phases with a gate between each: reproduce, isolate, minimize, explain, fix, lock.
Phase 1 Reproduced: 2 deliveries 40ms apart Phase 3 Minimal: no lock on the idempotency key Gate a red test before the fix, green after
The payment webhook sometimes runs twice/review
the agent calls itThe diff on two axes that do not contaminate each other, standards and spec, in parallel subagents.
Standards 2 findings bare catch in billing.ts:88 Spec 1 finding the partial refund in the issue is missing
Review the diff before the PR/engineer
you call itAn audit across every layer, security first, starting from a non-negotiable baseline of ten items, with the specific fix on every finding.
[CRITICAL] unbounded query in payouts.ts:41 [HIGH] retry without backoff on the gateway Verdict fix first
/engineer the payouts module/qa
you call itFive required checks, real end-to-end flows, edge cases, and what should be tested and isn't.
✓ build, types, lint, tests, happy path ✗ coupon expiring mid-checkout: 500 Gap: no test covers currency changes
/qa the checkout flow/security
you call itAn audit in three levels, and the third chains medium findings until they become critical.
Chain: sequential IDs + no ownership check + email in the response
= anyone can list every customer [CRITICAL]/security before opening the public API/deploy
you call itA security gate, reproducibility from a clean clone, and the checks for each way of shipping.
✓ a clean clone builds and boots ✗ .env.example is missing STRIPE_WEBHOOK_SECRET ✓ reversible migrations, a real health check
/deploy the first production release/ask
you call itThe router over all the others. Describe the situation and it tells you which skill solves it, and why.
/break: renders the card with the worst content on one page and names the skill that owns each break.
/ask the card breaks with a long name and I don't know where to startMore
All of it, with each skill's references and the token budget, is at matheuscarddoso/skills. The interface suite builds on Jakub Krehel's skills and Emil Kowalski's principles, both MIT licensed.
Every rule in these skills is a lesson, with a demo and an exercise, in craft.ocardoso.com.