Core skills

Compress markdown and text guidance files locally without external LLM or Claude CLI calls. Use when Codex needs to shorten SKILL.md, CLAUDE.md, prompts, runbooks, or other natural-language instruction files while preserving headings, fenced code, inline code, links, URLs, paths, and markdown structure.

Scaffold a version-focused 11ai operator plugin for a tool, CLI, service, or SDK so it matches the existing operator plugins exactly, covering version selection, identity naming, first-party native-skill discovery and compatibility, skill archetypes, safety, manifests, and catalog wiring. Use when asked to create or extend an operator plugin, add skills for a technology, or check that an operator plugin follows the house pattern and targets the requested or latest stable version.

Analyze an existing 11ai plugin and create or update its FAQ skill: a question router that maps user questions about skill behavior, usage, outputs, and troubleshooting to the plugin's own contracts, references, scripts, and tests, plus a comparison table and glossary the siblings do not carry. Use when asked to add a FAQ skill to a plugin, refresh a plugin FAQ after skills changed, or check that a FAQ still covers its plugin.

The house styleguide for generated, self-contained HTML reports: canonical design tokens with a dark default and a light toggle, embedded Inter and Geist Mono fonts with cross-platform fallback stacks, collapsed-disclosure section structure, and a table interaction contract covering descending-first sorting, multi-row click highlighting that survives sorting, and column resizing down to zero width, none of it persisting across reloads. Ships copyable tokens.css and fonts.css, a complete working report-template.html, the interaction spec with its known failure modes, and a regression checklist against previous generations. Use when building, styling, restyling, or reviewing a generated HTML or markdown report, or when asked to apply the house report style.

Reverse engineer a locally cloned repo — extract core logic, data flows, and key implementation pieces (sanitized of all branded/personal/company information), then produce a single markdown blueprint with instructions to rebuild it better using modern technologies and best practices. Use when the user wants to study, reverse engineer, modernize, or rebuild an existing repo.

Apply the house web design language to a Next.js application built with shadcn components and Tailwind CSS v4: oklch design tokens in light and dark modes, square corners with hairline borders, a single accent family with a five-step chart ramp, Inter and Geist Mono via next/font variables, next-themes with a system default and a d hotkey, and dual-theme code highlighting. Tokens and fonts are user-parameterized: whatever the user specifies is kept, missing tokens and fonts are generated as balanced equivalents, a single accent expands into a full palette, a one-mode palette derives its counterpart with contrast gates enforced, and the bundled defaults apply when nothing is specified. Use when styling, theming, or restyling a Next.js and shadcn application, generating design tokens from a brand color, or asked to apply the house www style.

Speak in the short register, always extremely pragmatic and objective: extremely concise lines, grammar sacrificed for concision, lists over paragraphs, no em dashes, an example or snippet for every claim, sources cited. Four working modes on top: a fix beside every problem found when reviewing, no file changes while planning until the action plan is approved, a pointer to where changes landed when implementing, and a copyable conventional commit message with scope after any repo work. Use when the user asks for terse, punchy, or no-fluff replies, sets a brevity rule for the session, or says an answer reads long-winded, padded, or hedged.

Review, critique, and roast code, documents, designs, or any work product for improvements — read-only, no file edits. Use whenever the user asks to review, critique, roast, tear apart, poke holes in, or get honest/brutal feedback on something, or asks "what's wrong with this" or "how would you improve this", even if they don't say the word "roast". Delivers blunt, prioritized suggestions only; never modifies files.