Skill Claude CodeCodex
Rewrite prose that sounds synthetic, inflated, generic, or assistant-like. Use for AI-writing tells, vague attribution, suspicious citations, chatbot artifacts, or copy that should sound grounded and human. Applies to prose; for machine-written code tells use unslop.
Agent
How the engineering skills should consume this repo's domain documentation when exploring the codebase.
Agent
Issues and specs for this repo live in Linear, team SKI (Skills). Linear is the record - findings, plans and specs are written up as Linear items rather than markdown in this repo, and Linear wins when the two disagree.
Agent
The skills speak in terms of five canonical triage roles. This file maps those roles to the actual label strings used in this repo's issue tracker.
Skill Claude CodeCodex
Remove hidden fallbacks, swallowed errors, legacy aliases, permissive defaults, and silent compatibility paths. Use when configuration or control flow should have one canonical path and explicit failure modes, including deterministic environment handling with Envy.
Skill Claude CodeCodex
Add, check, explain, or repair JavaScript and TypeScript architecture boundaries with @howells/boundaries. Use for Turborepo package rules, app-to-app import prevention, source-layer profiles, package tags, or fixing violations without weakening the intended architecture.
Skill Claude CodeCodex
Test a rendered web app in a real browser and return evidence-backed QA findings. Use for dogfooding, localhost review, responsive or mobile defects, console or accessibility checks, persona walkthroughs, and frontend validation that must exercise the running interface rather than inspect code alone.
Skill Claude CodeCodex
Implement substantial production changes in foreman mode: the main agent plans, briefs, and reviews while subagents write code, routed by taste, heavy, or grunt work. Use for implementation or refactoring that benefits from delegated execution. Skip for tiny fixes or documentation-only work.
Skill Claude CodeCodex
Create, review, or revise a distinctive visual identity system: positioning, rendered direction options, OKLCH palette, typography, visual character, and Tailwind v4 tokens. Use when a product needs brand identity established before interface design. For screen-level UI direction use chiaroscuro.
Skill Claude CodeCodex
Find and refactor god components, oversized modules, tangled scripts, and multi-responsibility JavaScript or TypeScript files. Use for safe decomposition, responsibility splits, and extracting duplicated logic inside a codebase. For repeated UI components use componentize; for package extraction use aperture.
Skill Claude CodeCodex
Answer a question about intent, history, or provenance by sweeping every available evidence source in parallel, then return a cited answer and coverage map. Use when a claim, decision, or number needs proving beyond the code. For rebuilding current working context use muster.
Skill Claude CodeCodex
Add concise, useful JSDoc where IDE hover help matters. Use for public JavaScript or TypeScript APIs, exports, components, hooks, classes, complex types, generated API docs, or package publishing. Clarifies non-obvious behavior without commenting everything or changing code.
Skill Claude CodeCodex
Audit a Mastra codebase against execution failures first: workflow step size, fan-out keying, suspend and resume payloads, load-bearing writes, model settings, and visible tool keys. Use for pre-ship review or an existing Mastra implementation. For building Mastra features use $mastra.
Skill Claude CodeCodex
Report what this task is doing now from live state: original ask, completed work, remaining work, branch, uncommitted files, and drift. Use when returning to a long session or checking whether an agent is still on track. For a roll-call across tasks use muster; for rewriting only the last reply use what.
Skill Claude CodeCodex
Rebuild working context across concurrent or older agent tasks from transcripts, git state, peer sessions, and the tracker. Use for 'catch me up', 'where did I leave off', or 'what is in flight', returning every active thread and its next move. For evidence-backed investigation use inquest.
Skill Claude CodeCodex
Generate, critique, and validate names for products, projects, packages, CLIs, apps, brands, or features. Use for naming or renaming work where domain, package, GitHub, App Store, directory, or web conflict checks matter.
Skill Claude CodeCodex
Create, audit, or migrate agent plugins across Claude Code, Codex, and Cursor. Use for multi-host manifests, marketplace metadata, shared capability paths, release and version alignment, or deciding whether a skill collection should become a plugin.
Skill Claude CodeCodex
Build or extend an outside-in product description: a feature-by-feature, event-by-event account of user-visible behavior drafted from code and tests, verified against the running product, and consolidated into bug triage. Use for behavior-spec repositories or complete product-experience documentation across web…
Skill Claude CodeCodex
Rescue work that exists in only one place - detached HEADs, worktrees, unpushed branches, or stashes - then remove only what is proven merged and pushed. Use for repository cleanup where ambiguous work must be preserved and each surviving thread named. For status reconstruction without cleanup use muster.
Skill Claude CodeCodex
Grade an entire codebase with a stage-calibrated verdict, clustered findings, and comparable scores. Use for repository health audits where mechanical checks, source-confirmed findings, lifecycle stage, and multiple review lenses matter. For a diff or PR use code review; for Mastra use mastraudit.
Skill Claude CodeCodex
Remove machine-written code tells from a diff without changing behavior: narrating comments, one-caller wrappers, impossible guards, unused options, decorative logs, and leftover scaffolding. Use before commit or review. For behavior-changing fallback cleanup use fail-fast; for prose use deslop.
Skill Claude CodeCodex
Rewrite only the previous agent reply into three short, plain-language lines: what happened, where things stand, and what comes next. Use when the last message was dense, jargon-heavy, or unclear. It re-explains existing content; for whole-task status use memento.