Claude Dev Skills

A portable Claude Code plugin — a set of eight general-purpose development skills that encode how to work: exploring a request before building, planning and debugging methodically, reviewing and auditing, designing UI, committing cleanly, and keeping memory tidy. Packaged as one marketplace plugin so the whole working discipline moves to a new machine in two commands.

Origin. These skills emerged while developing AutoBot-AI — a distributed autonomous-agent platform. They are the general, project-agnostic half of that discipline, extracted so the practice travels beyond the project that produced it. The AutoBot-specific skills (issue implementation, PR mechanics, stack debugging) live in AutoBot-AI-Claude-dev-skills.

Copyright © 2026 mrveiss · Apache-2.0.

---

The one rule

Every skill here assumes it, so it is stated once rather than repeated in each.

Say "I don't know" — never fabricate.

A plausible-sounding guessed cause, count or verdict is the one error treated as serious. Not because guessing is impolite, but because for an agent a wrong answer is not an opinion — it executes. A person who fabricates a cause is wrong until someone checks; an agent acts on it, at machine speed, against the environment it needs to keep working. A guess produces an action, the action produces new state, and the new state is then read as evidence.

And an admission is not a closure — it is an opening. "I don't know" means "I need help with this: let's research it, guide me, let's find the answer together." It is the rewarded move because it starts the conversation that solves the problem, not because it ends one.

So the follow-through is not silence. In an interactive session, ask right then — for the guidance, the decision, or the joint dig that closes the gap. Where there is nobody to ask, leave the criterion unticked and say why, fix it in scope or file it, and never drop it. An admission that opens no dialog and produces no issue has lost the finding politely.

A stated gap is a finding. An unstated one is the defect. The same missing knowledge produces either, depending only on whether someone wrote it down.

The operative form for anything mechanical: distinguish nothing found from did not look. A check that never ran and a check that passed report the same green, and only one of them means anything.

Install

This repository is a Claude Code plugin marketplace. Install from inside Claude Code:

/plugin marketplace add mrveiss/Claude-Dev-Skills
/plugin install claude-dev-skills@claude-dev-skills

or from a terminal with the Claude Code CLI:

claude plugin marketplace add mrveiss/Claude-Dev-Skills
claude plugin install claude-dev-skills@claude-dev-skills

What each step does. marketplace add clones this repo into Claude Code's marketplace cache and validates .claude-plugin/marketplace.json. install copies the plugin into the plugin cache and enables it for your user scope. Both are recorded in your ~/.claude/settings.json (extraKnownMarketplaces and enabledPlugins), so the state is declarative and reproducible.

Verify the install:

claude plugin list          # claude-dev-skills should show ✔ enabled

On a new machine, the same two commands restore the entire set — nothing else to copy.

Skills are read fresh on each invocation, so an update to the installed plugin takes effect without restarting a conversation. New skills added to the plugin appear after the next claude plugin marketplace update.

---

Using the skills

Skills load automatically and trigger two ways:

1. By intent. Each skill's description tells Claude when it applies. Ask for the underlying task in plain language — "help me plan this migration", "review this design", "audit the repo for the same bug" — and the matching skill activates. You do not name it.
2. Explicitly. Invoke one directly with a slash command: /process, /ui-design, /web-audit, and so on.

Each skill is self-contained: everything it needs is in its SKILL.md. There are no external files to fetch and nothing to configure before first use.

---

The eight skills

Process & discipline

• process — The approach discipline that goes before the code: brainstorming intent before building, writing a checkable plan, debugging by reproduction instead of guessing, test-driving the implementation, verifying a claim before calling it done, dispatching parallel agents only for truly independent work, and finishing a branch cleanly. Use at the start of any non-trivial task.
• commit — A standardized commit workflow: pre-flight branch and staging checks, auto-format, conventional-commit message shape, and retry logic when a pre-commit hook rewrites files.
• canonical-coding — The canonical-source discipline: before adding a helper, wrapper, or "v2" of anything, find the existing one and consolidate. Defines which source wins when a frontend and backend, or a doc and the code, disagree.

Review & audit

• review-lenses — Six domain lenses for judging whether work is good, not merely whether it runs — architecture, delivery, frontend, documentation, UX, and visual craft. Use for any "is this good?" review.
• gap-audit — After a batch fix, sweep the adjacent files for the same defect and file a discovery issue per gap found, so a fix in one place doesn't leave siblings broken.
• web-audit — A full multi-pass security, SEO, and AI-friendliness audit of any website — HTTP headers, TLS, CORS, DNS/subdomain recon, exposed panels, per-page SEO and performance — emitting a self-contained HTML report.

Design

• ui-design — One skill for every interface job: choosing an aesthetic direction that doesn't look templated, auditing an existing UI, redesigning without going generic, component polish and motion, and building a token-based design system. Self-contained, framework-agnostic.

Memory

• memory-cleanup — An end-of-session hygiene ritual for a file-based memory system: keep the index an index, ensure every entry is reachable from a topic file, and prune what has gone stale or duplicated.

---

Updating

claude plugin marketplace update claude-dev-skills   # pull the latest from this repo

Updates to an already-installed skill flow through on the next read. A newly added skill becomes available after the marketplace update above.

Uninstalling

claude plugin uninstall claude-dev-skills@claude-dev-skills
claude plugin marketplace remove claude-dev-skills     # optional: forget the marketplace too

---

Contributing

1. One skill = one responsibility. If a SKILL.md grows past ~700 lines, split it.
2. Every published skill must be self-contained. A skill's SKILL.md may not depend on files that aren't committed — an installer only receives what ships. Two skills (process, ui-design) keep optional local references/ enrichment on the dev machine; that material is git-ignored and not required for the skill to work.
3. Edit → PR → review. Workflow-rule changes get the same review as code. State in the PR what behaviour changed and why.
4. Write the description for triggering. It is how Claude decides when the skill applies — name the situations, not the mechanics.

Layout

.claude-plugin/marketplace.json          the marketplace manifest
plugins/claude-dev-skills/
  .claude-plugin/plugin.json             the plugin manifest
  skills/<name>/SKILL.md                 one self-contained skill each

---

Acknowledgments

Two skills distill the work of other skill authors into original, self-contained rewrites. Their SKILL.md text is written fresh for this repo — it does not redistribute the source material — but the practice it captures is owed to:

• process — the Superpowers skill suite by Jesse Vincent (obra): brainstorming, TDD, systematic debugging, verification, planning, worktrees, and the subagent workflows, plus Anthropic's skill-authoring guidance. For the full upstream treatment of any topic, see that repo.
• ui-design — five design skills: Emil Kowalski's philosophy on UI polish and motion, Anthropic's frontend-design, and the ui-ux-pro-max, taste-skill, and impeccable skills for systems, anti-slop redesign, and audit lenses.

Where an author's name was not recorded in the source, the skill is credited by its name. If you authored one of these and want different or removed attribution, open an issue.

Conventions

• One canonical skill per job. No -v2 or -fix variants; consolidate rather than fork.
• Self-contained by rule. A published SKILL.md carries everything it needs; it never routes to files that don't ship.
• A skill is a checklist, not a manual. Keep it scannable; push depth into clearly-named sections, not walls of prose.