claude-kit

CryptoGnome's curated, minimal, opinionated Claude Code skills. Install once, use on every project, keep in sync, share with anyone.

Not a kitchen-sink mega-pack — a small, ruthlessly-curated set written our way: the best parts of the skills worth borrowing, stripped to their essence, portable across projects, and free of machine-specific defaults and secrets.

Quickstart

1. Add the marketplace + install (in Claude Code):

   /plugin marketplace add CryptoGnome/claude-kit
   /plugin install claude-kit@claude-kit
   

2. Restart your session — the skills load and the always-on lazy-surgical hook activates.
3. Just work. Skills fire automatically when your task matches; or run any explicitly with /<name> (e.g. /security-audit).

No config. Works with any model.

Other ways to install

skills CLI (copy-and-own):

npx skills add CryptoGnome/claude-kit                    # the whole kit
npx skills add CryptoGnome/claude-kit --skill research   # just one
npx skills add CryptoGnome/claude-kit --list             # list what's inside

Manual:

git clone https://github.com/CryptoGnome/claude-kit.git
cp -r claude-kit/skills/research ~/.claude/skills/       # personal (all projects)

Update

/plugin marketplace update claude-kit (or enable auto-update in /plugin) pulls new versions and skills. Manual/CLI installs: git pull and re-run npx skills add. See CHANGELOG.md.

Why these exist

Each skill removes a specific pain of coding with an AI agent:

• The agent over-builds or drifts → lazy-surgical (least code, surgical diffs) + grill (align before building) + research (get the facts first).
• The output looks like generic AI → anti-slop-frontend / redesign-existing-projects for UI, marketing-copy / seo-geo-aeo for words.
• Security & quality slip through → security-audit (agent-powered vuln audit), a11y-audit, react-best-practices.
• Crypto code is easy to get wrong → ethereum and solana (chain-specific security + tooling).
• Hard bugs, murky design, no plan → diagnosing-bugs (feedback-loop-first debugging), codebase-design (deep modules), prototype (throwaway code to settle a design question), plan-to-issues (PRD → vertical slices).
• The workflow glue → semver, handoff, caveman, image-gen, remotion.

Skills

Twenty-one skills, grouped by what they're for. All are model-invoked — Claude reaches for them automatically when your task matches, or you name them with /<name> (handoff is invoke-only).

Coding discipline & craft

Skill	What it does	Secrets
lazy-surgical	The least code that fully solves it — reuse first, surgical diffs, simple, verifiable (lite/full/ultra). Always-on via hook	none
codebase-design	Design deep modules — lots of behaviour behind a small interface at a clean seam, testable through it (precise shared vocabulary)	none
diagnosing-bugs	Feedback-loop-first debugging for hard/flaky bugs & perf regressions — reproduce → minimise → hypothesise → instrument → fix	none
security-audit	Deep vulnerability audit via parallel subagents — triage → source→sink → adversarial FP cull → ranked findings (no paid tools)	none
caveman	Terse-output mode — strips filler, keeps every fact/command exact (lite/full/ultra)	none
semver	The correct next version via SemVer — major/minor/patch, the 0.x rule, changelog→tag flow	none

Planning & workflow

Skill	What it does	Secrets
grill	Interrogate the plan (one question at a time, with recommended answers) until aligned, before any code	none
research	Delegate reading legwork to a background agent — investigate PRIMARY sources → a cited Markdown note to work against	none
prototype	Throwaway prototype to settle ONE design question — a terminal app for state/logic, or toggleable UI variants	none
plan-to-issues	Synthesize a PRD (no interview) then cut it into end-to-end vertical-slice issues — via gh or markdown	none
handoff	Compact the session into a standalone handoff doc for a fresh agent (invoke-only)	none

Frontend & design

Skill	What it does	Secrets
anti-slop-frontend	Anti-slop guardrails for NEW UI — design read, dials, countable layout rules, em-dash ban	none
redesign-existing-projects	Upgrade an EXISTING site without breaking it — audit + prioritized fixes, any stack	none
a11y-audit	Audit UI code for accessibility/UX violations → terse file:line findings	none
react-best-practices	Fix React/Next.js perf + component architecture in impact order (waterfalls, bundles, RSC)	none

Content & visuals

Skill	What it does	Secrets
marketing-copy	High-converting copy — positioning, headline/CTA formulas, page structure, offer & objection frameworks	none
seo-geo-aeo	Search & AI-answer readiness audit — SEO + AEO + GEO checklists, schema, a 1–10 scored report	none
image-gen	Generate any image via OpenRouter (Gemini / GPT / FLUX / Seedream / Grok) + prompt/style help	OpenRouter key
remotion	Programmatic video in React (Remotion) — frame-driven animation, deterministic rendering, sequencing	none

Web3 / crypto

Skill	What it does	Secrets
ethereum	Production EVM — Solidity security, Foundry testing, gas/L2 choices, Scaffold-ETH 2 / viem / wagmi	none
solana	Production Solana — @solana/kit, Anchor & Pinocchio, PDAs/CPIs, Token-2022, program security, testing	none

Always-on discipline

Note

You don't have to type slash commands — most skills auto-activate by description. The one exception by design is lazy-surgical: it's a coding temperament for every edit, so the kit ships a SessionStart hook (hooks/hooks.json → hooks/session-start.js) that injects it into every session. Turn it off by disabling the plugin in /plugin or deleting hooks/hooks.json. (Needs Node.js; skipped harmlessly if absent. The file is auto-loaded, so it's intentionally not declared in plugin.json.)

Philosophy

• Subtraction over accumulation. A skill earns its place or it's cut.
• One skill = one job. Split anything that does two things.
• Rules as countable checks, not vibes. "≤ 4 elements", "trace every line to the request" — not "use good judgment".
• Trigger-engineered descriptions. Each skill says exactly when to fire and when not to.
• Portable & safe by default. No machine-specific paths, no committed secrets, ever.

Adding a skill

Every skill clears the bar in GOVERNANCE.md before it lands: one job per skill, a trigger-engineered description, countable rules over vibes, and a hard no committed secrets / no machine-specific paths audit. Versioning follows the semver skill; minor & milestone releases get a GitHub Release. New skills are distilled in our own way from the best sources — credited, never blindly vendored.

License

MIT © CryptoGnome