agent-skills

Dude, come on. You write at the density of your own thinking and leave the reader to unpack it. Apply whenever explaining anything to a human — not just end-of-run summaries, every message.

Use when creating or iterating staged planning artifacts for a complex feature: design discussion, structure outline, final executor plan, or supporting research/research questions. Also use when designing before coding, breaking work into vertical slices, or deepening a high-value roadmap item, NNN improve plan, architecture candidate, or HumanLayer-style task.

Use when turning a roadmap, repo priorities, architecture priorities, chosen opportunities, accepted feature outlines/plans, or “roadmap to plans” request into grouped improvement plan batches for future executors. Writes README.md indexes, 001-*.md/NNN-style plans, optional memo-*.md files, verification gates, drift checks, STOP conditions, rejected approaches, and executor handoff notes.

Use when a user asks “what should I work on in this repo?”, “what’s worth doing next?”, repo audit, tech-debt priorities, architecture priorities, old-plan reconciliation, or repo-grounded opportunity discovery. Produces .agents/ROADMAP.md with generated opportunities, user-seeded ideas, Now/Next/Later sequencing, rejected/not-now items, and next-artifact recommendations. Not for standard product/release roadmaps; if multiple repos are requested, create one roadmap per repo.

Use when reviewing code quality, asking whether a refactor or design is good, looking for code smells, or evaluating architecture tradeoffs. Identifies shallow interfaces, invalid state models, unclear boundaries, hidden effects, weak failure contracts, over-mocking, over-abstraction, compatibility glue, and clean-looking code that preserves bad models.

Build evidence-backed software/product roadmaps from repo state, tickets, ideas, audit findings, or user goals — prioritize and sequence work across Now/Next/Later, milestones, dependencies, risks, and handoffs. Use when asked for a roadmap, feature roadmap, technical roadmap, release plan, milestone plan, backlog prioritization, sequencing, what to build next, or to turn improve/ideate/brainstorm/design-discussion outputs into an execution sequence. Not for implementing code or writing PR-level plans.

Survey a codebase as a senior advisor and turn the highest-value findings into implementation plans for other agents to execute — strictly read-only on source code, never implements anything itself. Use when asked to audit a codebase, find improvement opportunities (bugs, security, performance, test coverage, tech debt, architecture, migrations, DX), suggest features or roadmap direction, or generate handoff plans for another agent to implement.

Write self-contained implementation plan files that a fresh, possibly weaker executor can run without the author's context — splitting work into PR-sized plans, with verification gates, STOP conditions, and a memo protocol for design forks. Use when asked to write/create a plan, turn a decided design or ticket into plans, or when another skill hands off plan-writing. Not for deciding what to build or executing plans.

Use when reviewing PRs, simplifying over-engineered code, judging architecture, or renaming unclear concepts — inline fake helpers, delete meatless ceremony, rename lying names, merge over-split files, reject premature abstraction, say no to speculative config/modes/layers, then end with the smallest safe next change. Triggers when code is too fancy, too abstract, too clever, too many files/helpers/layers, or too well-factored but painful to change. Embody Grug brain: complexity very bad, small words, no consultant speak, no hard pivot to opposite dogma.

Detect and eliminate common AI writing tropes from prose. Use when drafting, editing, or reviewing text to avoid the predictable patterns that mark AI-generated writing.

Use when authoring an agent skill that wraps a command-line tool — covers hands-on tool exploration, required vs. recommended sections, installation/usage structure, trigger-rich descriptions, task-grouped commands, progressive disclosure, and a pre-publish checklist. Triggers for CLI / command-line / terminal / shell-command tools and binary wrappers; for review, run the Checklist section.

Use when optimizing CLAUDE.md, AGENTS.md, custom commands, or skill files — diagnose the concrete failure first, then apply current documented Anthropic best practices (explicit instructions, context/motivation, examples, output and verbosity control, thinking/effort, CLAUDE.md size and skill-description rules) instead of inventing improvements. Triggers when a prompt isn't followed, a skill won't activate, CLAUDE.md is too long or ignored, or migrating prompts to current Claude models.

Jujutsu (jj) — the Git-compatible version control system. Activate ONLY when a .jj/ directory is present in the project or when jj/jujutsu is explicitly mentioned. Do NOT activate for plain git repos without .jj/. Use for any VCS operations in jj-managed projects: commit, push, pull, branch, bookmark, rebase, squash, merge, diff, log, status, working copy, change ID, revset, fileset, template, configuration, workspaces.

Use when authoring, creating, refining, or troubleshooting agent skills — scaffold SKILL.md and frontmatter, write and optimize the trigger description, structure the body with progressive disclosure, validate structure, and test or debug activation. Also when building a new skill from scratch, when a skill won't trigger, loads incorrectly, or the agent ignores it entirely. Use when a skill misbehaved in the current session and needs adjustment based on learnings.

Use when building or reviewing frontend UI — dashboards, admin panels, landing pages, marketing sites, web apps. Drives domain-specific design decisions (typography, color world, layout, CSS token naming, depth and spacing systems) instead of generic AI defaults; routes to app.md (product/data UIs) or marketing.md (public/creative pages) by context.

Mental-model reset for SvelteKit apps. Use when writing or reviewing routes, layouts, load functions, form actions, remote functions, hooks, auth, cookies, endpoints, redirects, errors, SSR, progressive enhancement, or app-level data flow. Triggers on SvelteKit, +page, +layout, +server, +page.server.ts, +layout.server.ts, hooks.server.ts, load, actions, fail(), redirect(), error(), cookies, locals, route groups, protected routes, sessions, form actions, enhance, remote functions, command(), query(), form(), getRequestEvent(), SSR, hydration, and serialization. Use svelte5 for component-level runes, snippets, accessibility, actions, transitions, and component API review.

Reviews Svelte components for idiomatic patterns, identifies anti-patterns carried over from React/Vue/Svelte 4, and suggests Svelte 5 refactorings. Use when writing or reviewing Svelte components, asking "is this idiomatic", "Svelte way", doing Svelte 5 migration, or encountering smells like: effect assigns state, prop copied to $state, global store by default, $bindable everywhere, clickable div, createEventDispatcher, export let, on:click, slot-shaped APIs, lifecycle-driven code, imperative DOM wiring, immutable-update ceremony, context value replacement, shadcn-svelte form structure, Field.* components, bits-ui form controls, or component APIs that hide ownership. This is the general-purpose entry point for Svelte component review; delegates to sveltekit for routes/load/actions/server concerns and to focused references for details.

Mental-model reset for Rust. Use when writing or reviewing Rust code to shift from "it compiles" to "thinks in Rust." Triggers on Rust code review, "is this idiomatic", borrow-checker errors, API design, domain modeling, ownership, lifetimes, errors, traits, async/Tokio, unsafe, serde, FFI, tests, performance, Cargo structure, .rs files, Cargo.toml, rustc diagnostics, clippy findings, Result/Option, thiserror vs anyhow, newtype, typestate, enum vs trait, dyn Trait, Send/Sync, Pin, Miri, PyO3, napi-rs, cxx, UniFFI, wasm-bindgen, serde attributes, or feature unification.

Mental-model reset for Salsa, the incremental computation framework for Rust. Use when building or reviewing Salsa databases, tracked functions, input/ tracked/interned structs, query pipelines, accumulators, cancellation, durability, LSP integration, memory management, cycles, or production Salsa architecture. Triggers on #[salsa::db], #[salsa::input], #[salsa::tracked], #[salsa::interned], #[salsa::accumulator], salsa::Storage, memoization, revisions, backdating, red-green algorithm, WillExecute, DidValidateMemoizedValue, Cancelled, returns(ref), no_eq, lru, cycle_fn, cycle_result, durability, or salsa::Event.

Use when writing, reviewing, or rewriting user-facing error messages, validation messages, form errors, empty/error states, auth errors, failure notifications, retry/support copy, CLI errors, or API errors humans will read. Helps make errors specific, actionable, non-blaming, accessible, and safe without leaking sensitive details.