skills
Agent skills for engineers who won't lower their standards to use AI.
30 skills, open source, used every day on real work. Install the set, or take only the ones you want.
- skills
- 30skills
- plugins
- 13plugins
- licensed
- MITlicensed
- agents
- 2agents
Install
Installed as plugins from one marketplace. Add it once, then install the plugins you want.
Any agent — pick the skills and agents you use. The installer writes editable files into your project.
npx skills@latest add smorinlabs/smorinlabs-harnessClaude Code — add the marketplace.
/plugin marketplace add smorinlabs/smorinlabs-harnessThen install a plugin.
/plugin install repo-hygiene@smorinlabs-harnessCodex — clone the repo, then register it in ~/.codex/config.toml.
git clone https://github.com/smorinlabs/smorinlabs-harness.gitThen register it in the config.
[marketplaces.smorinlabs-harness]
source_type = "local"
source = "/absolute/path/to/smorinlabs-harness"Source: smorinlabs/smorinlabs-harness
What a skill actually is
The problem
An agent will happily do the work the wrong way — skip the tests, guess at the API, commit straight to main — and sound confident doing it.
The fix
A skill is a written procedure the agent loads when it matches the work. Not a prompt you paste: a file the agent finds on its own, at the moment it applies.
What you type
A slash command, or nothing at all. Most of these fire on their own when the task fits, which is the point — you should not have to remember they exist.
Start here
Eight that earn their place on the first day.
The set
All 30, grouped by what you reached for them to do.
Nothing matches that search.
Shape the work
Turn a vague idea into decisions and tasks you can actually start.
8 skills
start with /question-walkthrough
- /question-walkthroughDecide a pile of open questions one at a time.Takes an unresolved pile — open questions in a doc, undecided to-dos, a plan blocked on four things nobody has answered — and turns it into a sequence of single decisions.
- /project-addCapture an idea in four questions, then stop.Registers a new project on the PROJECTS.md trunk: at most four questions, asked one at a time, then an idea stub and a trunk row written in a single atomic commit.
- /project-refineScope an idea up to ready, or break it into tasks.Three sub-modes over one project: refine its notes and references, scope a raw idea until it is genuinely startable, or decompose a scoped project into tasks.
- /project-nextAnswer "what should I work on?" from the actual trunk.Reads PROJECTS.md and projects/ and builds a composite menu: what is in progress, the next few unstarted projects by ID, and the files touched most recently.
- /project-auditEleven drift checks against your project tracking.Dispatches one auditor per check in parallel against PROJECTS.md and projects/, then walks the findings with you one at a time and asks before fixing anything.
- /using-project-harnessBootstraps the project skills on first use in a repo.Establishes the five project-harness skills in a repo, runs a four-step idempotent setup the first time, and routes any edit of project state to whichever sibling owns it.
- /guided-researchDeep research at the moments that actually justify it.An orchestration layer over deep research that decides when the expensive version is worth running — architectural decisions, new problem domains with established best practice, unfamiliar APIs you have already failed against twice — then organizes what comes back so it is reusable rather than read once.
- /design-by-elementsPrototype the form of an artifact before you build it.Decomposes an artifact — terminal output, a report layout, a table design, a page or email mockup, a status block — into elements, then iterates on each one in isolation before assembling.
Work the code
Review, refactor, and consolidate what is already written.
4 skills
start with /factor-scan
- /factor-scanSweep a code area for bugs, smells, and architectural drift.A broad sweep over an area of code that surfaces three different classes of problem at once — outright bugs, quality issues, and architectural smells — then walks you through the findings one at a time rather than dumping a report.
- /factor-architectArchitectural review against precedent and convention.Reviews a design against the codebase's own precedent, its stated conventions, and external references for the pattern in question.
- /factor-dedupConsolidate N implementations of the same idea.Walks per-implementation comparator reports across several similar implementations, classifies every divergence as intentional or accidental, and proposes a unified design — or documents why a divergence should survive.
- /using-factor-harnessRoute to the right one of the three factor skills.Orients across the factor-harness set and picks the right entry point for what you are actually doing.
- /pr-merge-flowDrive an open PR to merge, resolving every thread.Waits (with a bound) for the AI reviewer bots to weigh in, then triages every review thread: verifies claims by running the code where it can, fixes small in-scope bugs, defers out-of-scope work to the tracker, declines style-only asks, and escalates genuine redesigns to you.
- /ci-auditAudit, fix, and monitor CI and pre-commit hooks.Covers the whole CI surface: whether GitHub Actions are passing, why they are not, whether pre-commit hooks are actually running, and whether the workflow files themselves lint clean.
- /version-checkIs this version consistent, and is it published?Inspects and compares the versions a project claims across its manifest, its tags, and its registry, and reports where they disagree.
- /readme-syncDoes the README still describe the code?Verifies README.md against the current state of the project and updates it where it has drifted.
- /manual-test-guideThe commands to run to check this by hand.Produces concrete manual testing instructions for a project — what to run, what to look at, what counts as working.
- /explainExplanations anchored to a real example, every time.A fixed anatomy for explaining anything: what it is, just enough context to place it, and always a concrete example rather than a description of one.
- /html-codesignA decision page someone can operate and send back.Builds a single self-contained HTML file that turns a decision into something a person can operate: a step-through wizard, one question per screen, each with context and an argued recommendation.
- /clear-technical-communicationRewrite technical prose so the reader infers nothing.Reviews and rewrites technical communication against one rule: minimize what the reader has to reconstruct.
- /use-html-themePick a visual theme, apply it consistently.Offers a small catalog of visual themes whenever HTML is being generated or restyled, and applies the chosen one coherently across the page.
- /reader-stepsFormat the steps only a human can perform.A style spec for rendering actions the reader has to do themselves — interactive logins, UI clicks, approvals in other systems — as a bounded task block: numbered with a stable tag, grouped by surface (terminal, browser, phone, physical world), each step titled by its outcome with the literal command underneath and a verification line.
- /document-mergeMerge overlapping docs without losing a line.Consolidates several overlapping markdown documents into one, verifiably: every synthesis decision gets a stable ID, every source line range maps to an output section, the originals are archived byte-identical, and a validation script enforces the round trip.
- /html-explainAn explainer page where prose alone would not do.A read-only page for settled material, where diagrams, annotated figures, charts, and small interactive widgets do work prose cannot.
Run the session
Orient, hand off, and clean up the agent sessions themselves.
6 skills
start with /session-recap
- /session-recapCatch yourself up on a session gone cold.Reconstructs a session you have lost the thread of: what it was about, where it started and ended, what is done versus open, the live state of git, the PR, and the worktree — then ranks what to do next and ends with a verdict on whether to close it or carry on.
- /session-statusA "you are here" map of work still in flight.A mid-flight progress check on the active stream of work, in plain language: what the work is, what is done, what is in progress, what is left — scaled to the shape of the work rather than a fixed template.
- /session-handoffPackage a session so a cold one can continue it.Packages the current session into something a zero-context session — often on another machine — can act on cold: the outcome being pursued, full repo identity and paths, the resume state, and the tacit context that dies on compaction.
- /session-loose-endsFind what a session left dangling, and clean it up.Sweeps for four classes of loose end — uncommitted work and stale worktrees, running or zombie processes, promises made in conversation but never kept, and project-tracking drift — then reports each with its evidence and a recommendation in both directions before doing anything.
- /session-agent-listEvery agent session on this machine, as action cards.Finds Claude Code and Codex sessions across the machine and renders them as action-ready cards: how long ago, whether each is open mid-work or handed off or closed cleanly, what it was doing, and which repos and worktrees it touched.
- /repo-finderResolve a repo name to its real path in one command.Turns a bare repo name into every local copy plus the facts you need about each: origin, default branch, whether it is a checkout or a worktree, current branch, dirty state, build tooling.
What changed recently
The last five releases.
- v0.21.0Aug 12, 2026
8 skills adopted from the private harness
clear-technical-communication (reader-centered review and rewriting of technical communication), design-by-elements (element-by-element prototyping of an artifact's form), document-merge (verifiable consolidation of overlapping documents), and the five session-* skills — session-recap, session-loose-ends, session-handoff, session-status, session-agent-list — as the new session plugin. Marketplace goes from 9 plugins / 21 skills to 13 / 30.
- v0.20.0Aug 9, 2026
use-html-theme 0.11.0 — html-codesign wizard-by-default view + plain-language digest
The codesign page now opens as a step-through wizard: one question per screen, explicit Next/Back with a "Question N of M" progress line and clickable answered/skipped dots, and a last-step "Finish → review" that lands on the list view with every section folded. A Wizard|List control-bar toggle reaches the classic full-list view and back at any time, resuming the exact question. The spec's optional defaultView ("wizard" when absent) can open small pages in list view; everything after open is DOM-only view state, so identical picks export identically from either view. Every question is now required to be self-explanatory in clear, plain language — clear-technical-communication is applied when installed, with an embedded digest in the skill as the fallback proxy.
- v0.19.0Aug 8, 2026
repo-hygiene 0.8.0 — pr-merge-flow --one-pass
A single-pass override: run one triage/fix pass over the threads present, skip the step-5 re-review cycle entirely, and always end as a ready-report — never merging, because merging straight after a push without re-checking would merge over reviews still being written. It composes with any end mode (--auto/--confirm endings downgrade to the ready-report) and answers the "fix all the comments, just don't start another round" case that previously needed a repo-sticky cycle-bound: 1 pref plus a check-in answer. The step-3 ledger rule gains a one-pass carve-out — the work set is fixed at the pass's inventory, and a final fully-paginated refresh at the exit boundary reports late arrivals as open rather than triaging them (PR #37; the refresh itself came from the PR's own Copilot and CodeRabbit review round).
- v0.18.0Jul 27, 2026
repo-hygiene 0.6.0 — pr-merge-flow convergence guardrails
Scope-gated triage of every valid finding (fix / defer / decline / refute / escalate) with a value floor; per-wave trajectory measurement with a ratcheting bar (findings-received, never fixes-chosen); a check-in with three endings (continue until clean · merge and defer the residue · pause for redesign); deferral destinations detected from repo evidence (defer-target pref); new references/convergence.md; seven new/split Red Flags rows. Grounded in eleven measured review episodes across four repos (PR #31; auto-mode consent question deferred to #32).
- v0.17.1Jul 26, 2026
html-codesign 0.9.1 — the gate's worked example no longer demonstrates the violation it forbids
The example in the preflight ended with *"Generate it for these four?"* appended to the question list — inside the pre-read block. Eleven lines below, the same file rules that the pre-read **ends its turn, with no tool call of any kind after it**, and that the ask belongs in the dialog's option labels. An agent follows the concrete example over the prose rule, which is the entire reason the example exists, so the contradiction would have reproduced the same-turn bug v0.16.0 fixed. The shared context-triage.md had already been corrected during v0.17.0; this propagates it to the skill's own inline copy and states explicitly why the example stops where it does.
Also recommended
Not mine. Installed anyway, on most machines I work on.
Stops the guessing loop where an agent describes what it changed instead of showing you.
Turns the instruction file you keep meaning to tidy into something that actually steers behavior.
Maps a foggy piece of work into decision tickets, so planning stops being one long conversation.
Setup flows that ask rather than assume — the pattern most install skills should have copied.
Structural, syntax-aware code search and rewriting. The right tool the moment a regex refactor starts feeling brave.
Create, improve, and evaluate skills — including measuring whether a description actually triggers.
The methodology layer underneath everything else: brainstorming, planning, TDD, systematic debugging.
A second model family as an independent lens — adversarial review from something that is not Claude.
Install them and get to work
MIT licensed, one repo, no account. Take the whole set or one plugin.
Add the marketplace.
/plugin marketplace add smorinlabs/smorinlabs-harnessThen install a plugin.
/plugin install repo-hygiene@smorinlabs-harness