zvadaadam/az-skills

az-skills

A curated set of skills I use with my agents. Feel free to grab what's useful.

Install  ·  Skills  ·  Update  ·  Feedback  ·  skills.sh

---

Each skill is a small package of instructions and code that gives an agent a new ability — like fixing failing CI pipelines, exploring problems from multiple angles, or cleaning up messy code. This repo is updated as I build and refine new skills.

Install

Use Vercel’s Skills CLI to select skills and agents:

npx skills add zvadaadam/az-skills

Install the whole collection globally for Claude Code and Codex:

npx skills add zvadaadam/az-skills --skill '*' --agent claude-code codex --global

Or list the available skills and install just one:

npx skills add zvadaadam/az-skills --list
npx skills add zvadaadam/az-skills --skill deslop

Each skill includes a standalone Hivenet feedback command. The optional az-skills-feedback skill adds detailed reporting guidance; it is included when you install the whole collection.

The Skills CLI installs from this public GitHub repository. The skills///SKILL.md layout is discovered automatically; no npm package or registry upload is needed. New installs use the default branch, so repository changes become available after they reach master. The skills.sh listing is driven by installations, as described in Vercel’s FAQ.

The Skills CLI has its own install-count telemetry. To disable it, prefix commands with DISABLE_TELEMETRY=1; see the CLI documentation. Hivenet commands in this repo disable automatic context collection separately.

Install from a local checkout

For development, link skills directly to a clone so local edits are available immediately:

git clone https://github.com/zvadaadam/az-skills.git
cd az-skills
./scripts/install.sh

This links the skills into ~/.claude/skills, ~/.agents/skills, and ~/.codex/skills and installs this repo’s Git hooks for updates.

Update

For Skills CLI installs:

npx skills update

For a linked local checkout, pull the latest:

cd az-skills
git pull

Moving from the old installer

The local install and uninstall scripts remove the retired feedback hook, this checkout’s old skill-feedback links, and the local installation ID. The Skills CLI copies individual skill directories and does not run this cleanup. Before switching from an earlier local install, run the cleanup from the updated checkout:

python3 scripts/remove_legacy_feedback.py

If you previously installed the retired skill-feedback package through the Skills CLI, remove it as well:

npx skills remove skill-feedback --global

Omit --global if that old package was installed in a project. Then install the current skills with npx skills add as above.

Uninstall

For Skills CLI installs, name the skill to remove:

npx skills remove deslop --global

Omit --global for project installs. For a linked local checkout:

./scripts/uninstall.sh

Feedback

Report broken, confusing, outdated, or notably good behavior through Hivenet:

DO_NOT_TRACK=1 npx --yes hivenet@latest --to az-skills --dsn https://[email protected]/az-skills --category skill --subject "" ""

The az-skills-feedback skill covers structured failed-task reports for the team’s eval corpus, attachments, and follow-up questions and replies. The bundled publishable DSN fixes the destination to az-skills; no user-provided key is needed. These examples disable automatic context collection. Keep reports specific and never include secrets or private data.

---

What’s inside

Engineering

• az-skills-feedback — Hivenet support channel for skill feedback, structured failed-task reports, and replies from the az-skills team
• call-advisor — Calls the premium Fable advisor through Claude Code CLI for hard judgment, architecture, and frontend/design critique, with resumable sessions and saved output artifacts
• call-worker — Calls the fast Codex GPT-5.5 worker through Codex CLI for clean code implementation, repo exploration, tests, and bounded engineering tasks
• code-review — Multi-lens code review with 3 parallel sub-agents (correctness, security, design) that validates and reports only high-signal findings
• devs-roundtable — 5 legendary engineers (Carmack, Hickey, Metz, Torvalds, Beck) debate your problem in parallel, then build consensus
• code-simplifier — Reviews code for clarity and maintainability, then cleans it up
• complexity-check — Audits a change for unnecessary additions, assumptions, spread, and duplication
• deslop — Detects and removes AI-generated code slop (unnecessary abstractions, over-engineering, verbose patterns)
• implementation-rehearsal — Walks through a proposed implementation without changing code, challenges likely regressions, and revises the plan before building
• pre-factor — Auto-fires before a feature or non-trivial change: maps the code the change will land in and surfaces the prep refactors that make the change easy (reshape the seam, add a safety net, kill duplication) — each one traced to the upcoming change, landed as its own commit first. The before-bookend to complexity-check
• repo-history-book — Builds an evidence-backed account of how a project evolved from its commits, PRs, releases, and docs
• setup-agent-workflow — Sets up a repository for effective AI agent work: learn the product, repair the development loop, encode recurring mistakes in code and checks, connect concise AGENTS.md guidance to task skills and feature knowledge, then demonstrate and maintain a complete change-and-verification loop. Reuses existing tools and adds only what the repository needs
• tour — Explores a codebase subsystem and produces a self-contained HTML tour
• verify-harness — Gives a repo a verification harness its agents actually use: a control- driver CLI that runs the real app in an isolated session and drives it the way a person and an agent do, plus a user-POV feature map (what exists, how to reach it, a replayable recipe of expects, what misleads) checked in CI for completeness. Covers setup, expanding the map with parallel mappers, verifying a change with proof, reproducing reports, and maintenance — with a “no silent scripts” gaps loop that turns every one-off probe into a driver command. Based on Lauren Tan’s (@poteto) verification skills in Cursor’s pstack

Design

• design-roundtable — 5 legendary designers (Rams, Ive, Vignelli, Fukasawa, Jongerius) debate your brief in parallel, then build consensus

Marketing

• brand-name-explore — Generates product/company names using multiple creative personas (Lexicon methodology, poet, linguist, culture hacker, futurist)
• ai-answer-audit — Reverse-engineers an AI “best X” answer back to the searches, sources, and assumptions behind it: an evidence ledger, a model-layer vs content-layer split, the multi-hop search path, and which claims are unsupported model guesswork. User-run and read-only — it never alters the answer
• geo-optimize — Turns an ai-answer-audit into a prioritized plan to get a brand cited in AI answers (ChatGPT, Perplexity, Google AI Overviews): authority gap, four levers (get-cited / fix-open-territory / open-a-lane / upgrade-evidence), per-engine moves, and a Fast Wins / Roadmap / Backlog roadmap

Product

• product-rehearsal — Imagines a proposed feature has already shipped, walks a real user or agent through finding and using it, surfaces where it will be bad, and revises the proposal before building. The behavior-level sibling of implementation-rehearsal

DevOps

• greenlight-pr — Takes a PR, fixes CI failures, addresses review comments, and iterates until everything passes

Productivity

• interview-me — Interviews you about a plan or design until it has all the context to build the right thing
• noah-zender-it — Selects relevant mental models from Noah Zender’s idea library and explains how to apply them
• plan-for-goal — Turns conversation context into a single prompt for a coding agent’s /goal orchestration loop — directional outcome, quality bar, and a self-verification path the loop can iterate against
• plan-for-mega-goal — Turns a multi-objective effort into a roadmap and a compact prompt for an autonomous goal loop