skills

Review a GitHub pull request using the passive, neutral, assertive, or aggressive profile, optionally paired with a named reviewer persona that sets the review voice, by statically reading the PR diff, metadata, comments, and discovered issue/context links to determine whether it solves the stated issue. Use for automated or manual PR review flows that should leave an emoji-marked top-level review plus targeted inline comments or suggestion blocks, without running code or blocking except for malicious-looking changes.

This skill should be used when considering delegating a coding task to the OpenAI Codex CLI — codebase exploration/Q&A, well-specified implementation, mechanical edits at scale, second-opinion or adversarial review, or bugs stuck after 2+ attempts. Codex spends ChatGPT quota instead of Claude tokens and its models are token-economical. Invoked for "use codex", "ask codex", "delegate to codex", "codex exec", "second opinion", or when offloading token-heavy scoped work.

Create or extend generated profile/portrait emoji atlases for Slack-style custom emoji. Use when making a first-run or extension set of pixel/profile/person/mascot emoji, generating a chroma-key atlas, splitting it into individual transparent PNGs, fixing green-screen/background artifacts, validating emoji-size readability, creating manifests/previews/zips, or optionally uploading custom emoji with agent-slack.

This skill should be used when a non-Claude agent (e.g. Codex) is considering delegating a coding task to Claude Code's headless mode (claude -p) — hard long-horizon bug fixing, implementation where first-pass completeness matters more than token cost, test authoring, or a cross-model-family second opinion on a plan or diff. Claude Code spends an Anthropic subscription/API budget instead of the caller's tokens. Invoked for "use claude", "ask claude", "delegate to claude", "claude -p", or "second opinion from claude".

Sync a forked repository with its upstream. Fetches both remotes, shows divergence, resets shared branches to upstream, re-merges local-only branches, cleans up branches already merged upstream, and pushes. Use when upstream has accepted PRs or moved ahead and you need to bring your fork in line.

Analyzes and improves code structure — decomposes long functions and files, reduces complexity, extracts shared patterns, reframes complexity so whole branches or layers disappear, flags abstractions that don't fit (thin wrappers, over-generalization, reinvented helpers), assesses test coverage on critical paths, and cleans up dead, unreachable, or orphaned code that accumulates as a side effect of refactoring. Use when the user wants to refactor for clarity, split large files, reduce nesting, DRY up code, simplify an over-engineered or wrong-fit abstraction, replace a reinvented helper with the repo's canonical one, improve testability, or sweep for dead code after restructuring. Not for feature changes, bug fixes, or performance optimization — this is structural refactoring only.

Audit a codebase's module boundaries — enumerate modules, map their seams (import edges between modules), produce a layered topology diagram, and classify each module as narrow, hub-by-design, or accidental hub (with separate flags for cycles, layer violations, and uncertain import graphs). Outputs a diagram plus a flagged-for-review list; does not change code. Use when assessing whether abstractions live at the right boundaries, before/after a refactor to verify the boundaries improved, or when an unfamiliar codebase needs an architectural map. Not for intra-module refactoring (see improve-code-structure), bug hunting, or feature work.

Speak using pop-culture quotes, movie lines, song lyrics, and memes, verbatim or playfully corrupted to fit. The defining rule is that quotes SUBSTITUTE for plain prose, not decorate it. Quote bank lives in references/quotes/ organized by theme of work × mood. Intensity levels are subtle, full (default), unhinged. Use when the user says "quotespeak", "talk in quotes", "meme mode", "quote mode", or invokes /quotespeak. Stays active every response until "stop quotespeak" or "normal mode". Make sure to use this skill whenever the user invokes it OR continues a quotespeak session, even when a particular turn looks like routine technical work, structured proposal-writing, or long synthesis. Those modes are exactly when persona density most often drops, and the skill is most needed.

Repeatedly invokes a target skill until its output settles — meaning the target makes no further substantive changes or recommendations. Detects cycles (oscillation between two or more states) and stalls (inner skill keeps recommending but changes don't land) and handles each appropriately. Optionally chains to a follow-up skill after settling. Use when one pass of a skill is rarely enough and each pass tends to uncover more work until eventually there's none — e.g. "repeat-until-settled improve-code-structure then release".

Create, update, or apply a macOS dotfiles repo. Use when the user wants to back up their system configuration, set up a new Mac from dotfiles, capture current configs into an existing dotfiles repo, or manage dotfiles with GNU Stow.

Extract and sharpen the user's mental model of an idea, concept, design, or plan through targeted questioning. Use when the user wants to articulate something they're thinking about, says "ask me about", or wants to be interviewed about a topic. Not for simple clarifying questions — use this when the topic deserves a dedicated interview.

Convert raster images (photos, illustrations, AI-generated art) into high-quality SVG recreations. Breaks the image into isolated features, builds each as a standalone SVG layer, then composites them. Use when the user wants to recreate an image as SVG, create vector versions of artwork, or extract specific elements from images as scalable graphics.

Review code changes from multiple specialist perspectives in parallel. Use when you want a thorough review of a PR, branch, or set of changes covering security, performance, correctness, edge cases, and ripple effects. Spawns parallel reviewer agents that each focus on a different lens, then synthesizes into a unified review.

Brainstorm competing solutions to a problem using parallel agents. Use when you need to explore multiple different approaches to the same problem, compare trade-offs, and choose the best path forward. Spawns parallel proposer agents who each design an independent solution, then peer-review each other's work before a structured comparison.

Debug problems by investigating multiple hypotheses in parallel. Use when you have a bug, unexpected behaviour, or mystery where the root cause is unclear. Spawns parallel investigator agents each pursuing a different theory, then compares evidence to identify the most likely cause and fix.

Manage stacked branches — rebase cascades, detect landed PRs, show stack status. Use when branches are stacked (B on A on main), trunk has advanced, a mid-stack branch changed, or a PR has landed and descendants need rebasing. Lightweight alternative to Graphite that infers the stack from git history.

Investigate and solve problems using a team of specialist agents. Use when facing complex, multi-faceted problems that benefit from parallel research and structured implementation.

Activate orchestrator mode for complex multi-task work using subagents. Use when you need to coordinate multiple independent Task subagents to accomplish work while keeping the main context window clean.