A small, sharable collection of agent skills.
npx skills@latest add lmarqs/skillsFor claude.ai, download a skill's zip from the
skills-latest release and upload it in claude.ai's
skill settings. scaffold has no zip.
| Skill | What it does | Triggers when… |
|---|---|---|
hexagonal-layout |
Keeps a system's business logic independent of its technology by reasoning about which way dependencies point. Three kinds of code — the reason it exists, its connections to the outside, and how it's run (core · adapters · run). Language- and framework-agnostic. | You scaffold a new system, place a feature/integration/entrypoint, review how an existing one is shaped, or write a layout ADR — or a dependency points the wrong way and you're unsure where something belongs ("where should this go", "scaffold it ports-and-adapters", "review against hexagonal principles", "which layer owns this"). |
architecture-design |
Acts as the architect rather than the scribe, then writes the document: challenges the framing, writes stakeholders as usage roles and goals as outcomes, keeps external constraints and prior decisions out of the requirements, gives every requirement a goal, a scenario or a derived metric, and a proof, sources every claim about the current state, embeds the diagrams, and leaves a document that reads without opening a single link. A pragmatic blend of RFC and ADR, recording why a technical choice was made: size the decision, contextualize, requirements, design, tradeoff analysis (options and the baseline scored against requirement IDs, each risk with impact·probability·mitigation·contingency), the decision with its consequences and confirmation, then conclude & communicate. Matches the request's language. | You're choosing between technical options or documenting a decision ("write an RFC", "design doc", "architecture decision", "ADR", "tradeoff analysis", "technical documentation of an implementation") — including the retrospective case of documenting an implementation after the fact. |
sql-quality-check |
Audits SQL quality systematically — catches performance smells in raw SQL, ORM-generated queries, schema decisions, or PR diffs, explains the database-level impact, gives a corrected version, and makes the tradeoff explicit. Generic: no project/domain/ORM/language config; raises context (transactional table? intentional scan? expected volume?) during analysis. | You write, review, or optimize a query or data-access code ("review this query", "why is this slow", "is this index right") — or you spot N+1, SELECT *, a cartesian/row explosion, a missing date filter, OFFSET pagination, LIKE '%term%', NOT IN with nulls, an unindexed sort, or a long transaction. |
sql-migration-guide |
Senior-engineer knowledge for changing a live SQL schema and its data without outages, broken rollbacks or silent data loss. Classifies each change on availability, compatibility and reversibility, stages breaking changes with Expand–Migrate–Contract, designs batched idempotent backfills, sequences DDL against rolling deploys, and names verification and recovery before deploy. PostgreSQL and MySQL mechanics in references. | You plan, write or review a migration (ALTER TABLE, add/drop/rename a column, NOT NULL, a new index or constraint, a backfill, a zero-downtime change) or ask "will this lock the table", "is this migration safe", "can we roll this back". |
scaffold |
Scaffolds or audits a project's mise setup the house-standard way — a lean mise.toml, directory-namespaced .mise/tasks/* scripts (tf:apply, node:setup), a project "mise" house-rules skill, and command-skills for the few destructive tasks. Greenfield or brownfield (detect, report drift, fix on approval). Advisory: recommends the standard, explains the tradeoffs, lets you decide. |
You set up a new project/repo, add or standardize a task runner, want build/test/lint/dev/deploy commands made consistent, say "add mise" or "set it up like our other repos", or bootstrap tooling for terraform, python, node, go, localstack, docker-compose, aws, or arduino/platformio. |
ralph-tui-run |
Coordinates a beads epic or a prd.json task file through ralph-tui without writing code: reads the project's agent, ralph-tui, task source and git setup, proposes one run plan (scope, harness, model, effort, iteration budget, worktree, delivery strategy) for confirmation, starts ralph-tui headless so fresh subagents implement each task, monitors every N minutes, and delivers through the agreed PR strategy. | You explicitly ask to run an epic or task file with ralph ("ralph run", "run the epic with ralph", "implement the epic with ralph-tui", "run the prd.json with ralph", "start ralph on the beads"). |
clean-code-comments |
Decides whether a code comment earns its place — writes the ones that do, removes the ones that don't. Built on the Clean Code premise that a comment is a small failure of expression: prefer saying it in code (better name, extracted function, named constant) and keep only comments code can't carry — intent, warnings, external constraints, legal, TODO, public-API docs. Language-agnostic; findings in English, code quoted verbatim. | You write, review, or clean up comments ("should I comment this", "review these comments", "are these comments useful") — or you spot a comment that just restates the code, commented-out code, a changelog/byline comment, noise, a banner, a } // end label, or a stale comment that no longer matches the code. |
writing-for-readers |
Holds any text a person or agent will read to objective, reader-centered criteria, for writing and for reviewing. Declares the reader, what they know, their question and the scope first; puts the answer first; makes headings promise exactly what their bodies deliver; keeps each fact in one place so the text cannot contradict itself; resolves every pronoun and term; and keeps examples to their point. Grounded in reading research and style guides, each criterion with a test. | You write, rewrite or review a README, doc, spec, runbook, PR description, announcement or an agent's answer ("write the docs for X", "review this doc", "does this read well", "rewrite this for new hires", "this is confusing", "who is this for"). |
global/AGENTS.md holds the maintainer's personal, cross-project
agent instructions. It lives outside skills/, so the installer does not ship it.
Link it into an agent's global config to use it, e.g.:
ln -s "$PWD/global/AGENTS.md" ~/.claude/CLAUDE.mdEach skill includes skills/<name>/SKILL.md and agents/openai.yaml alongside it
for Codex's display name and short description. Optional
references/ holds progressive-disclosure detail. Skills are authored and eval'd with the
skill-creator tool vendored under
.agents/skills/ (pinned in skills-lock.json). See AGENTS.md for
the contributor guide.