smorinlabs-skills

Architecture-aware review, refactor, and cross-implementation dedup — surfaces bugs, smells, and precedent gaps, then emits specs for Superpowers to execute.

4 skills · 3 agents · hooks

Lightweight PROJECTS.md project management — capture, scope, decompose, triage, and drift-audit projects as auto-triggering skills that hand execution off to Superpowers.

5 skills · 2 agents · hooks

Repo & release-readiness checks — CI fix and audit (measured, targeted, escalating), version drift, README sync, manual test guides, a PR merge flow that resolves every review thread before merging, and a Dependabot sweep that clears open dependency PRs across configured orgs.

6 skills

Session-state quintet for coding sessions: session-recap orients when returning to one (arc, done/open, git/PR/worktree state, ranked next steps, close-or-continue verdict — strictly read-only); session-loose-ends is its acting sibling — sweeps working-state artifacts, processes, unfulfilled promises, and tracking drift, reports each with evidence and a keep-or-clean recommendation both ways, then executes and verifies only what was confirmed item by item; session-handoff faces forward — it packages the current session into a self-contained, machine-portable handoff (outcome, repo identity, resume state, and the tacit context that dies on compaction) for a fresh zero-context session to pick up cold; session-status is the mid-flight glance — a manually-triggered, plain-language progress map of the active stream (done rolled up, what's left fine-grained where you stand), probe-free and read-only, with an ASCII kanban board of what's left when work runs in parallel or much remains; session-agent-list finds and lists Claude Code, Codex, Muse, and OpenCode sessions across the machine as action-ready cards — end-state triage (open / handed off / closed clean / loose ends), each tool's session name, what is running now, lineage, and copy-paste resume or transcript commands, filterable by tool.

5 skills

Themed-HTML toolkit, three pure skills sharing one context-triage preflight. use-html-theme: on any HTML-generation request, offer a catalog of visual themes (Birchline warm-editorial, Technical-minimal docs-site, High-contrast-dark dashboard), apply the chosen theme to all subsequent HTML in the session, and optionally persist it per-project — natural language and [theme: x] inline-flag control, side-by-side preview on request. html-codesign: interactive 'pick and export' decision pages, opened as a step-through wizard by default (Next/Back, progress dots, Finish-to-review; full-list view one toggle away) — every question opens with a collapsible context-and-recommendation preamble (free-form rich content on a clarity scaffold) in clear, plain, self-explanatory language (sharpened by clear-technical-communication when installed, an embedded digest otherwise), the reader answers, skips, or raises a question on each section, folds answered sections to dense summary rows for review, then exports the decision as Markdown/JSON (slim by default, full ADR-style by toggle) with stable IDs for diffable iteration. html-explain: read-only explainer pages for settled material — authoring maps which parts are intrinsically hardest and budgets the page around them, then enriches with diagrams, annotated figures, build-up sequences, and manipulable widgets ONLY where comprehension actually breaks; clarity canon (Mayer, Sweller, Minto, Heath, Tufte, Bertin, Cleveland) carried as advisory aids, with a standalone encoding fallback for surfaces where the bundled dataviz skill does not exist. Both page skills read recent context and ALWAYS confirm before generating — codesign for unanswered questions, explain for settled work — and hand off to each other when triage disagrees with the request. Both page skills close delivery the same way: the full absolute path on its own line, then conditional convenience offers — open it in the browser only when the session can reach one, add it to

3 skills

Prepare Windows in VMware Fusion for local standard-account checks or GitHub Actions, then start, inspect, stop, or explicitly reset the working VM.

2 skills

Decision requests from an agent to a human during a run. Checks existing authorization, investigates discoverable facts, and requests unavailable human-supplied information without turning it into an approval. Sizes preparation and explanation by seven axes into three tiers, rechecks uncertainty after verification, and keeps material caveats visible beyond the length targets. Uses same-input tables, worked examples, event sequences, and ASCII diagrams with exact verbatim content. Questions and options have stable meanings; revised choices supersede old IDs. Recommendations use option A when supported, with an explicit neutral form otherwise. Silence and skip never grant authority. Adapts a self-contained brief to the available question tool or ordinary text, with independent work continuing while required answers remain pending. Significant decisions retain their rationale and approval scope. Includes behavioral evaluation cases. Never produces HTML artifacts.

1 skill

Aggressively review, rewrite, and compose technical explanations, comparisons, status, handoffs, and decision requests so readers get the necessary context, explicit relationships, clear status, option effects, recommendations, and exact actions without reconstructing the author's mental model; restores stripped precision by keeping exact names paired with descriptions and frames every embedded code block, diagram, and example

1 skill

Iterative, element-by-element method for prototyping the visual or textual form of an artifact — terminal output, report and CLI layouts, table designs, page and email mockups — before building it: decompose into lockable elements, settle meaning before form, contrast opposed variants, lock with rationale, and promote every fix into a named rule

1 skill

Merge overlapping markdown docs into consolidated outputs without information loss, with full traceability of every synthesis decision

1 skill

Concrete-anchored explanation shorthand — /explain answers in a fixed anatomy: what it is, just-enough context, a real before/after example, and the payoff — with options/deeper/steps modes inferred from the target, and follow-up aware: a bare explain after an explanation diagnoses what's blocking action (step back to the bigger picture, clearer language, sharper example) and rewrites rather than repeats.

1 skill

Orchestration layer over the built-in /deep-research skill — decides when deep research is worth doing, shapes the prompt with project constraints, invokes the engine, and organizes results into a durable, reusable research tree.

1 skill

Install organization-member PR reviews and optional manual fix PRs with the OpenCode GitHub Action and Meta Muse Spark

1 skill

Adaptive one-question-at-a-time decision walkthrough. Takes a pile of open questions or undecided to-dos — mined from the conversation, read from a pointed-at document, given inline, or pulled from task systems — confirms the pile, sequences it by leverage (answers that could moot or reshape other questions go first), then walks it one AskUserQuestion at a time. Every non-obvious question is fronted by a pre-read — why it exists, impact, trade-offs, terms — with an ASCII visual when the question is complex (three or more relationships, a sequence or state to simulate, or diverging option paths) — delivered in a chat turn that ends BEFORE the dialog is raised (same-turn context is invisible context). Answer notes are classified and honored: option modifiers apply immediately; directives — extra work or a redirect that replaces the answer — get a one-line repeat-back and confirm, then run in an end-of-walk batch by default. After every answer the remaining pile is re-planned: mooted questions dropped loudly, order revised, implied follow-ups added with consent. Decisions are recorded back at their source and summarized in a final outcome table with parked items and directive outcomes called out. Pure-markdown skill, portable across Claude Code and Codex.

1 skill

Style spec for rendering actions only the reader can or will perform: a delineated task block toward the response's end — bounded frame (blockquote rail plus horizontal rules), a Completes line binding it to the tracked item by exact ID and title, steps numbered with a stable tag and grouped by surface (terminal, browser, desktop/system UI, phone, physical world) under merged dividers, each step titled by its outcome with the literal command or UI path below it and a ✓ verification line, closing with the immediate next move. Scales down to a one-line inline form for a single step and up with a map and stop points past eight; breadcrumb navigation promotes to one hop per line past three hops or on any caveat; reactive actions nest inside the step that triggers them; in-flight processes re-render across turns as a progress scoreboard; errors state cause and fix. Decisions are never steps — they are asked (question-walkthrough) before instructing. Canonical spec behind an always-on digest users keep in global instructions.

1 skill

One-command repo resolution for agents: resolve a repo name to every local copy with orientation facts (path, origin, default branch, checkout-vs-worktree kind, branch, dirty state, build tooling) via a config-bounded multi-root scan — falling back to the user's configured GitHub orgs over gh REST-first (GraphQL only as configured fallback) for exact clone names. Deterministic ranking (checkouts, then worktrees labeled with their main repo, then nested copies); recency displayed, not ranked, so the consuming LLM decides. Single-file zero-dependency uv Python script inside the skill; conforms to the CLI Design Standard (small-CLI profile, minimal tier). Replaces token-expensive ls/find cascades and repeated gh repo view identity checks.

1 skill