LLM coding skills and agents

Skills and agents I use with AI coding tools. Mostly Laravel/Livewire, but some are general-purpose.

---

Heads up (19th August 2026): The marp-* skills have moved to their own project in marp-themes.

---

Heads up (11th July 2026): The plan-to-ait pipeline has been restructured for stronger main agents. The issue-crafting knowledge (epic/issue templates, testing modes, the amnesia test) now lives in a new ait-crafting skill, so the main agent can create issues straight from a feature conversation; the plan-to-ait agent is now a thin shell that loads that skill. There's also a new ait-amnesia-check agent — a fresh-context checker that demonstrates what it would build from newly created issues (never a verdict), so the caller can diff its reading against the real intent. The conversation-to-ait skill has been absorbed into ait-crafting and removed — if you'd previously synced it, delete it from your tool config (e.g. rm -rf ~/.claude/skills/conversation-to-ait).

---

Heads up (3rd July 2026): quality-gate has grown into a two-tier review stack: free deterministic checks first (a section-order checker for Eloquent models and Livewire components in skills/quality-gate/scripts/, plus a drop-in Pest architecture test in skills/quality-gate/arch/), then the reviewer agents, all briefed via skills/quality-gate/briefing.md with whole-file reading rules. There's a new laravel-conventions-reviewer agent, and test-quality-checker now also hunts tautological tests, unit-tests-that-should-be-feature-tests, and test proliferation. The section order, arch rules, and conventions checklist encode our house style — read through and tweak them to yours before adopting.

---

Heads up (5th April 2026): The Livewire/Flux skills and agents have been reorganised. frontend-fluxui is now flux-ui, livewire-principles is now modern-livewire, and the livewire-flux-simplifier agent has been replaced by livewire-flux-reviewer. There are also three new skills: feature-workflow (ait-driven feature development), ui-to-flux (migration helper), and quality-gate (post-feature review).

If you've been using ./sync, run ./migrate to clean up the old names and pull in the new ones. If you copied files manually, have a read through the migrate script to see what's changed — it's short and commented.

---

Skills

• ait -- Local-first issue tracker for coding agents. Tracks tasks, models dependencies, and helps agents pick up where they left off after session loss.
• ait-crafting -- Turns a settled feature discussion or plan document into consultant-ready ait issues: an initiative/epic vision document plus implementation specs a fresh agent could pick up cold. Includes a "does this capture it?" review gate for conversation-sourced work, and pairs with the ait-amnesia-check agent for a fresh-eyes conveyance check. Also the knowledge base for the plan-to-ait agent.
• ait-recap -- Generates a friendly markdown recap of recent ait activity for the current project or across a directory of projects. Handy for weekly status updates, "what shipped recently?" questions, and quick memory refreshers.
• ant -- Local-first notebook for the why behind project work. Captures decisions, alternatives rejected, pivots, and foundation context as a sibling to ait's task tracking.
• audience-credibility -- Defensive review pass for audience-facing deliverables. Verifies named places, institutions, personas, and other local details before writing, then checks finished copy and visuals for recognisable AI-generated tells.
• changelog -- Analyses git tags and diffs to draft CHANGELOG.md entries in Keep a Changelog format. Proposes changes for review rather than writing them directly.
• conversation-to-html -- Turns Claude Code or Codex session logs into single-file HTML transcripts you can share or present.
• cruft-or-keep -- Conversational second pair of eyes for reachable-but-dormant code. Builds evidence-grounded case files for suspect jobs, mailables, listeners, routes, commands, and asks the developer cruft or keep? rather than deleting anything.
• expand-issue -- Fleshes out a terse "note to self" ait issue into a proper plan. Explores the codebase to work out what past-you meant, makes reasonable assumptions (and lists them for correction), and produces a plan document ready for plan-to-ait. Handy for drive-by brain dumps you want to pick up properly the next day.
• feature-workflow -- Streamlined workflow for building features with ait issue tracking. Handles orientation, planning (via plan mode), issue creation (via plan-to-ait), implementation with acceptance criteria checking, and optional quality gate at the end.
• flux-ui -- Flux UI v2 component reference for Laravel/Livewire. Covers syntax, patterns, common mistakes, and modal/form/table patterns.
• github-create -- Creates a new GitHub repo from the current project, pushes code, and sorts out the README, licence, and metadata. Uses the safe git/gh wrappers from agent-commit if they're installed, falling back to plain git/gh otherwise.
• golang -- Conventions and patterns for Go CLI/TUI projects. Covers project layout, SQLite storage, Bubble Tea TUIs, embedded web UIs, flag parsing, migrations, error handling, and testing.
• grounded-recommendation -- Makes the agent investigate before recommending. Present findings first, then a recommendation with explicit uncertainty.
• homebrew-spring-clean -- Walks developers through tidying /opt/homebrew (or /usr/local on Intel Macs). Surveys Homebrew disk usage, ranks cleanup targets, and handles cleanup, leaves, casks, stale logs, and autoremove without auto-deleting anything.
• house-style -- Teaches a personal editor agent the user's own editing taste by comparing an AI draft with their hand-edited version. Turns the differences into concrete editing procedures, then verifies them with controlled re-runs rather than trusting the agent's report.
• i-impeccable -- Distinctive frontend/design implementation skill. Establishes design context, commits to a bold visual direction, and builds polished interfaces that avoid generic AI aesthetics. Supports teach for context setup, craft for shape-then-build, and extract for pulling reusable design tokens and components into the system.
• i-shape -- UX/UI shaping skill for planning a feature before code. Runs a structured discovery interview, then produces a design brief covering direction, layout strategy, key states, interactions, content, and the most relevant i-impeccable references for implementation.
• improve -- Senior-advisor audit and handoff-planning skill. Surveys codebases for bugs, security, performance, tests, tech debt, dependencies, DX, docs, and direction, then writes self-contained plans for other agents to execute while staying read-only on source.
• larastan -- Installs and runs PHPStan + Larastan, separates framework false positives from real defects, and iterates through analysis levels.
• livewire-v4-upgrade -- Audit-first workflow for upgrading production Laravel apps from Livewire v3 to v4. Covers app-shape triage, wire:model changes, layout/config traps, legacy model binding remediation, test baselines, browser smoke tests, and deploy pipeline gotchas.
• mac-disk-hunter -- Tracks down where disk space has gone on macOS, including misleading System Data, Time Machine local snapshots, developer caches, Docker, Xcode, and local ML models. Surveys and ranks the likely culprits first, then suggests appropriately cautious cleanup commands without deleting anything until you confirm.
• laravel-cloud -- Deploy, update, and tear down demo apps on Laravel Cloud using the cloud CLI.
• modern-livewire -- How we write Livewire components: principles, patterns, form state, testing. Covers everything from findOrNew+fill+save to named modals and wire:model behaviour.
• plan-to-html -- Converts markdown plans, design docs, and issue writeups into self-contained HTML documents for sharing with non-technical stakeholders. Includes a stylesheet you can customise, optional author/title metadata, print-friendly output, and an optional Background preamble for context.
• practical-laravel-api -- Conventions for practical, consumer-friendly Laravel JSON APIs with Sanctum. Covers self-describing response shapes, loud query failures, slugs, date windows, filtering, Scramble docs, and tests for API contracts.
• quality-gate -- One-stop post-feature review: free deterministic checks first (model/Livewire section-order script, drop-in Pest arch test), then fresh-eyes review agents. Two scopes — recent work, or a whole legacy codebase chunked by model orbit and delegated to a sub-orchestrator. Confirms before spending agent tokens. (Note: it likes to ask you to commit a lot - if you don't trust your agent to do git stuff, have a look at agent-commit).
• ui-to-flux -- Migration skill for converting older Laravel apps (Bulma, Bootstrap, Tailwind) to Flux UI. References flux-ui and modern-livewire for target patterns rather than duplicating them.
• ui-migration-screenshots -- Captures full-page reference screenshots of existing app UIs before or during frontend migrations. Uses Playwright for authenticated, role-gated, and interactive flows so rebuilt pages can keep a familiar 1:1 mapping.
• readme -- Generates a README from the actual codebase so claims stay grounded.
• technical-overview -- Generates a TECHNICAL_OVERVIEW.md covering stack, architecture, domain model, routes, and key logic. Useful for onboarding.
• uofg-design-system -- University of Glasgow web design system. Brand tokens, layout rules, and component patterns for HTML, CSS, React, Tailwind, and FluxUI.
• shell-function -- Creates a shell function (bash, zsh, fish, etc.) that does what you describe.
• sop-creator -- Writes Standard Operating Procedure documents from user outlines. Asks minimal clarifying questions, then generates a clean Markdown SOP with purpose, prerequisites, procedure, and verification steps.
• swift -- Conventions and patterns for native macOS Swift apps. Young sibling of golang: XcodeGen workflow, sandbox-first permissions posture, fixture-backed testing, headless verification tricks, and a growing list of hard-won AppKit/SwiftUI gotchas.
• wut -- Ask the model to rewrite it's response in a more human-friendly and less jargon-laden way.

Agents

• ait-amnesia-check.md -- Fresh-eyes amnesia test for newly created ait issues. With no conversation context it demonstrates what it would build from each spec — restated goal, files, first failing test, every guess and dead end marked — so the caller can diff its reading against the real intent. Never gives a verdict.
• ait-audit.md -- Reviews open ait issues against the codebase and flags work that looks done but hasn't been closed.
• fresh-eyes.md -- Fresh pair of eyes for when you're stuck or looping. Suggests one or two things to try rather than editing code directly.
• humaniser.md -- Removes common AI-writing patterns from text files like READMEs and docs.
• laravel-conventions-reviewer.md -- Reviews Laravel code against team conventions: readable model helpers over raw column checks, fat models, Eloquent over query-building, enums over strings, duplicate-purpose methods. Seeded with real before/after examples — swap in your own house style before adopting.
• laravel-owasp-reporter.md -- OWASP-aligned security sweep of a Laravel app. Reports actual exploitable findings with concrete fixes.
• livewire-flux-reviewer.md -- Reviews Livewire/Flux code for simplification opportunities. Quick wins, things worth considering, and what looks good. References flux-ui and modern-livewire skills for pattern details.
• phpmetrics-check.md -- Runs phpmetrics on a PHP/Laravel codebase, flags complexity hotspots, and compares against a saved baseline.
• plan-reconciler.md -- Reconciles a plan document against the conversation that produced it. Reports what was captured, missed, glossed over, or drifted without editing the plan.
• plan-to-ait.md -- Turns approved plan documents into ait epics and issues that a fresh coding agent can execute without context. Now a thin shell around the ait-crafting skill, which holds the templates and quality bar.
• test-debug.md -- Debugging assistant for stubborn failing tests. Uses dump() instrumentation rather than speculative rewrites.
• test-quality-checker.md -- Reviews Laravel tests for false-confidence patterns and coverage gaps: tautologies, weak assertions, missing control records, unit tests that should be feature tests.

Syncing to your tools

Different AI coding tools (Claude Code, Codex, Gemini, Cursor, etc.) each have their own config directories for skills, agents and commands. The sync script copies everything from this repo into the right places.

1. Copy the example config: cp sync.yaml.example sync.yaml
2. Enable the tools you use and adjust paths if needed (sync.yaml is gitignored).
3. Run ./sync (requires uv).

If a destination file is newer than the repo source and actually differs, the script will warn you and offer a diff before overwriting — handy if you've been editing skills in-place. Identical files are left alone whatever their timestamps say. Skipped files are listed at the end for easy copy-back.

To skip specific items entirely (e.g. you have your own readme skill), add an ignore block globally or per-target in sync.yaml:

# Never sync these to any target
ignore:
  skills: [readme]

targets:
  claude:
    enabled: true
    skills: ~/.claude/skills/
    # Also skip these for claude specifically
    ignore:
      skills: [uofg-design-system]

./sync             # sync with prompts for newer destinations
./sync --dry-run   # preview only
./sync --force     # overwrite everything without asking

Pulling in sibling repos with sources

Not everything lives in this repo — some tools ship as standalone repos that carry their own skills and agents alongside their code. Rather than remembering a separate install step for each, sync.yaml can list extra source directories and ./sync treats their contents as if they lived here:

sources:
  - ~/code/ux-agent/claude
  - ~/code/a11y-agent/claude

Each entry points at a directory that itself contains skills/, agents/ and/or commands/. Every mode sees the merged set — the full sync, --dry-run, pick, and preset bundles can all draw from sources. On a name collision the first provider wins (this repo beats sources, earlier sources beat later ones) and the shadowed copy is reported, so nothing disappears quietly. ignore blocks apply to sourced items like anything else.

Project-local skills with pick and preset

The default ./sync is a full reconcile against your global tool config. But quite often you want the opposite — a fresh project where only a handful of skills/agents are relevant, dropped into the project's own ./.claude/ directory. Two modes for that:

./sync pick --to ~/code/my-project/.claude/           # interactive fuzzy picker
./sync preset go --to ~/code/my-project/.claude/      # named bundle from sync.yaml

Both treat --to as a base path — items land in PATH/skills/<name>, PATH/agents/<name>, etc. Neither touches sync.yaml's targets block, so global syncs are unaffected.

pick — fuzzy picker with preview

Opens an fzf multi-select across every available skill, agent and command. Type to filter, Tab to toggle, Enter to copy. The preview pane on the right shows each skill's SKILL.md (or the file itself for agents/commands), so you can browse and pick in one motion instead of remembering what each name does.

Items are tagged so the filter doubles as a scope:

• [s] skills
• [a] agents
• [c] commands

Typing s golang finds the Go skill; typing [a] shows all agents.

Requires fzf on PATH. If you don't have it, install via the fzf docs or use preset mode below.

preset — named bundles for shell aliases

For the "every Go project gets these three skills" case, declare a bundle in sync.yaml:

presets:
  go:
    skills: [golang, github-create, changelog]
    agents: [test-debug]
  laravel:
    skills: [flux-ui, modern-livewire, larastan]
    agents: [livewire-flux-reviewer]

Then a one-line shell function gets you a zero-friction alias:

gopick() { /path/to/agentic-stuff/sync preset go --to "$PWD/.claude/"; }

cd into a fresh repo, type gopick, done.

--dry-run previews either mode without writing anything; --force skips the "destination is newer than repo source" prompts.

--dry-run and --force work for both modes.

You might also be interested in

Standalone repos and tools that pair well with this collection - the quality-gate skill knows how to call on the first two, and the sources: block above will sync their skills and agents alongside these:

• a11y-agent — a deterministic accessibility CLI (axe scan, keyboard tab-order walk, screen-reader transcript) with its own Claude Code skill and background checker agent. Makes the a11y audit nobody gets time for nearly free.
• ux-agent — skills that drive a real browser against your local app to produce bug-reproduction videos, user-guide videos that re-film themselves, and cold UX journey reports from a code-blind probe agent.
• playwright-cli - Microsoft's browser-automation CLI, a good add-on for any UI/browser work. It ships its own Claude Code skill (installed via playwright-cli install --skills), and the flux-ui skill here leans on it to verify rendered pages rather than trusting that compiling Blade means correct pixels.