yasyf

Get a second opinion from OpenAI Codex CLI on difficult debugging, code analysis, or architecture problems, run a code/diff review (finder or adversarial-refuter passes over a diff or working tree), run a security review/audit or verification of security-sensitive code (auth, input validation, crypto, secrets), diagnose a bug, fan a repetitive bounded sweep (N-unit migrations, test conversions, mechanical refactors) across parallel codex lanes (individual bounded changes and large net-new code stay on Claude — opus, fable only when the surface is very sensitive or error-prone), generate images (logos, mascots, banners, illustrations) with Codex's $imagegen skill, or offload rote throwaway work (one-off scripts, data munging) where code quality doesn't matter and nothing can go wrong. Use when reviewing code or a diff for defects, when auditing or verifying security-sensitive code, when diagnosing a bug, when stuck after multiple attempts, for a sweep of fully specified edits, when asked to generate an image,

Bootstrap or retrofit a new repository with proven agent docs, README and brand assets, Claude Code settings, ccx, capt-hook guards, tests, linting, CI, and release conventions. Supports Python packages and CLIs using uv, Click, pytest, ruff, and ty; Go CLIs using cobra, slog, Task, and golangci-lint; Swift packages and CLIs using SPM, Swift Testing, SwiftFormat, and SwiftLint; and synced-project SwiftUI iOS apps. Optional publishing covers Great Docs on GitHub Pages, trusted PyPI releases, and Homebrew releases for Go and Swift CLIs. Use when creating a repository from scratch, scaffolding a Python, Go, or Swift package, CLI, or iOS app, or applying these conventions to a young repository.

Open a pull request that reads like the repo's own maintainers wrote it — orient with `ccx vcs lane` and `ccx vcs guidelines`, learn the house style from a bot-filtered sample of recent human-authored merged PRs via a cached per-repo style card, calibrate register against a local wlm voice profile, run a politeness and standards-conformance pass with issue-first handling on repos the user does not own, write the title, body, and commit messages to the repo's own conventions, ship through `ccx vcs ship`, and hand the CI and bot-comment watch to a background pr-watcher agent. Use when opening, submitting, or contributing a pull request, drafting a PR title or body, contributing a change to an open-source repo, or shipping a branch as a PR.

Attach to an already-open pull request and iterate it until CI is green and the bots are quiet — poll checks, review verdicts, and bot comments through a Monitor on the bundled poll script, apply the fixes the failure itself determines (formatter output, single-resolution compile and type errors, stale lockfiles and snapshots, review suggestion blocks, bot nits naming the exact replacement), ship each through ccx vcs ship, and bring every real decision back with 2–4 concrete options. Four mechanical safety gates precede every edit and a per-PR state file caps attempts per check across restarts. The human-driven counterpart to open-pr's background watcher. Use when the user says "watch this PR", "fix CI on PR 123", "get the PR green", "address the review comments", or "why is CI failing on my PR", when driving a PR someone else opened, or when resuming a PR whose background watcher already reported.

Research-backed, provider-agnostic guidance for writing LLM prompts and agent instructions — positive framing, XML structure, contrastive examples, outcome-first agentic contracts, and provider tuning knobs. Use when writing or revising a system prompt, agent or tool instructions, a tool description, or few-shot examples; or when the user asks to improve a prompt.

Check a file (or piece of text) for LLM-generated prose tells using the slop-cop CLI and report the violations. Use when the user asks to "check/lint/scan this file for slop", "find LLM tells in <file>", "run slop-cop on <path>", or wants a violations report for a doc, README, blog post, PR description, or commit message. Report-only by default; offers to apply fixes if asked.

Generate an animated SVG terminal demo of a CLI using evp — write a .tape, render it, inspect the keyframes, and refine in a loop until it looks good. Use when the user asks to "make/record a terminal demo of <cli>", wants an "animated SVG/GIF demo of a command", a "VHS-style terminal recording", or a demo GIF/SVG for a README. Renders natively on Linux x86_64 and via Docker (linux/amd64) on macOS/ARM.

Run an assumptions-first architecture design and ship it as an interactive design doc. Ground truths land in a register with stable IDs, every design fork goes through a question round with a recorded escape hatch, an adversarial review attacks the middle draft, and every estimate is gated on a named spike. The deliverable is a registers.json-driven single-file HTML doc plus a generated PDF, written as a humble proposal that explains and asks for feedback. Use when asked to "write a design doc", "architecture proposal", "help me design <system>", "redesign <system>", "assumptions-first design", or to turn a systems discussion into a reviewable design document.

Drafts one capt-hook (captain-hook) hook from a durable correction — the user's verbatim feedback plus its context — as a new .claude/hooks/<slug>.py, or (FIX mode) amends an existing misfiring hook with a mandatory regression test reproducing the misfire, or (EXTEND mode) broadens an existing hook to cover a newly mined rule without weakening its existing tests. Picks the right primitive (nudge for one-shot advice, gate for one-shot stop checks, hook(block=True) for always-on enforcement), writes the narrowest condition that captures the correction, a message that cites the correction, and inline tests (one Input firing on the offending shape, one Allow() on a benign neighbor), then proves the file with uvx --isolated capt-hook test before it goes live. Use when the user says "author a hook", "draft a hook from feedback", "encode this correction as a hook", "fix this misfiring hook", "broaden this hook", or when the bootstrapping-hooks or scanning-sessions skill delegates a hook to write or amend.

Surveys a repository and sets up captain-hook (capt-hook) guardrails for Claude Code — blocking gates, advisory nudges, command blocks, and test-integrity checks mined from the repo's own docs, CI workflows, lint configs, and git history. Scaffolds the framework and enables the session reviewer up front (Step 1), then proposes categorized candidates for user approval before writing anything, then writes .claude/hooks/*.py with inline tests and verifies with capt-hook test — the captain-hook plugin registers every hook event, so nothing else needs wiring. Use when the user asks to "set up captain hook", "set up capt-hook", "set up hooks", "bootstrap capt-hook", "add guardrails", "enforce our conventions with hooks", "protect this repo", or "make Claude follow CONTRIBUTING.md".

The headless session-reviewer brain — turns a watched repo's PR-eligible candidates into pull requests, both kinds. Invoked as /captain-hook:scanning-sessions --transcript <path> inside the target repo by capt-hook's detached SessionEnd reviewer pipeline. Enumerates judge-accepted, threshold-eligible candidates via uvx --isolated capt-hook review, re-verifies every cited quote verbatim against its transcript, screens each create candidate against the hooks already active in the repo, then drafts a new hook, broadens an existing one (EXTEND), or amends the attributed misfiring hook (FIX) by delegating to the authoring-hooks skill — routing follows the change shape, so a new hook lands repo-local and an edit lands in the hook's home repo. Proves it with uvx --isolated capt-hook test, opens exactly one PR with the verbatim evidence, and records the PR on the candidate. Use when a prompt starts with /captain-hook:scanning-sessions, or to review eligible capt-hook candidates and open hook PRs.

Translates a prose style guide (STYLEGUIDE.md, CONTRIBUTING style sections, CLAUDE.md conventions) into enforced captain-hook rules — AST Matcher StyleRules where the rule is mechanical, custom check() logic for complex-but-deterministic rules, LLM gates/nudges for semantic rules — each with inline tests synthesized from the guide's own good/bad examples, plus a final report of what is and is not enforced. Use when the user says "enforce our style guide", "turn STYLEGUIDE.md into hooks", "make Claude follow our coding conventions", or when bootstrapping-hooks finds a style guide.

Plans award trips with judgment — flights across 28 mileage programs via seats.aero, hotel award stays via rooms.aero. Triggers on the dense trip ask — "I want to go away for roughly a week... warm, beachy, cheap points tickets for business class... avoid the common places we always go like seoul or tokyo" — and whenever the user wants to plan a flight or trip on points or miles, find award availability between airports or across a region ("west coast to Asia"), plan a round trip or open jaw in one pass, get creative with routings (lie flat on points to a hub like NRT with a cash hop onward, two awards stitched across programs), find hotel award nights at the destination, weigh how interesting the layover city is, price a cash positioning flight, compare mileage programs for a route, pull booking links or taxes for an award, or resume a trip already in planning — or mentions seats.aero or rooms.aero. Needs a seats.aero Pro API key, from SEATS_AERO_API_KEY or a 1Password reference in ~/.getaway/preferences.jso

Sets up getaway travel preferences. Triggers when the user wants to set up getaway ("set up getaway", "set up my travel preferences", "run getaway onboarding") or to record airports, airline and hotel points balances, elite statuses, status goals, travel instruments (airline eCredits, vouchers, companion certificates, hotel free-night certificates), credit cards held (Amex Platinum, Chase Sapphire Reserve), travel documents (passports, residency, standing visas), layover preferences (minimize or explore, connection floor, long-stop cities), or avoid lists for award planning. Auto-fills from AwardWallet, Gmail, and logged-in airline and bank sites; nothing is written until the form's Submit. Refreshing balances already on file is /getaway:refresh.

Refreshes getaway points, miles, elite statuses, and travel instruments. Triggers when the user wants to refresh or update balances ("refresh my balances", "update my Aeroplan balance", "how many points do I have now"), re-check elite status, pull current credit-card points — Amex Membership Rewards, Chase Ultimate Rewards, Citi ThankYou, Capital One miles — refresh hotel points (Hyatt, Hilton, Marriott, IHG, Choice, Wyndham), or re-check travel instruments ("refresh my credits", "what credits are expiring", companion certificates, hotel free nights), or pull balances from AwardWallet ("refresh from AwardWallet", "pull my AwardWallet balances"). Reads AwardWallet first when a key is on file, then logged-in airline, hotel, and bank sites, with a Gmail statement fallback for banks and a Gmail mining pass for instruments. First-time setup is /getaway:onboard.

Detects and fixes LLM-generated prose tells using the slop-cop CLI. Triggers whenever the user asks the agent to write, draft, revise, polish, or edit natural-language prose (blog posts, docs, PR descriptions, commit messages, release notes, marketing copy, emails). The agent pipes its own draft through `slop-cop check -`, revises based on the JSON violation report, and delivers only the cleaned result.

Detects and fixes LLM-generated prose tells and plain-language clarity problems using the slop-cop CLI. Triggers whenever the user asks the agent to write, draft, revise, polish, or edit natural-language prose (blog posts, docs, PR descriptions, commit messages, release notes, marketing copy, emails). The agent pipes its own draft through `slop-cop check -`, revises based on the JSON violation report, and delivers only the cleaned result.

Author a block pack — custom block types for cc-present boards, shipped as React components declared by JSON Schemas and one prebuilt bundle. Use when the user wants a new block type for cc-present, a custom component or custom block on a board, asks to author, build, or write a block pack, or wants to extend cc-present's blocks beyond the built-ins.

Present options, drafts, or work-in-progress to the human as a live interactive artifact — an approval board at a localhost URL whose every click streams back into this session in realtime. Use when the user asks to present something interactively, wants an approval board, says "let me pick", "present this for approval", or "show me the options", or at any moment you would otherwise dump a static HTML page for sign-off.

Cache expensive pipeline artifacts (OCR passes, embeddings, parsed corpora, derived datasets) in the shared athome cache instead of ad-hoc pickle files, temp dirs, or hand-rolled digest schemes. Use whenever code re-derives bytes it already computed, needs atomic multi-file artifact writes, or needs record/replay caching of LLM API calls in tests.

Run scripts overnight or on a schedule on macOS: detach long runs from the terminal with completion sentinels, schedule recurring jobs as launchd agents, and make sweeps resumable so a crash retries only the failed units. Use instead of nohup, screen/tmux, cron, or hand-written plists whenever a command should outlive the session or recur on a schedule.

Use cc-notes to record durable tasks and notes that outlive a session, stored as git objects on refs/cc-notes/*. Triggers when an agent records a task or note for later; stores an artifact, evidence file, or dump — a VM or CI run log, a panic or crash dump, a repro archive — durably as a log attachment instead of committing it to the repo tree; adopts cc-notes in a repo that has not run init yet; runs status to orient on the backlog and who holds what; claims or starts a task; coordinates work across branches and multiple agents; manages leases and reclaims stale claims; verifies or supersedes a durable fact; records or executes a repeatable procedure as a runbook with per-run step tracking; records an approved plan verbatim and tracks its execution to an outcome; syncs tasks and notes with a remote; reconciles tasks after merging a branch; or links commits to the task they implemented.

Read code, find symbols, search a codebase, review diffs, edit a span in place, outline and search web pages, and compose multi-call pipelines with token-bounded outputs instead of raw file reads or whole-page fetches. Use whenever you need codebase context: reading a file, locating a symbol or definition, searching code by intent or text, listing files, reviewing changes, editing a line range you already have an anchor for, re-encoding a command's JSON output, or chaining tool calls and keeping only the distilled result — and whenever you need web content: reading or fetching a web page or docs URL, or asking a question of a page. Triggers on "read this file", "where is X", "find the Y function", "how does Z work", "search the code for", "show me the diff", "what calls this", "change line N", "replace this span", "filter this output", "for each file", "combine results from", "fetch this page", "read these docs", "what does the page say about X", or running any command that emits JSON. Reach for ccx before Re

Start or resume a cc-review review of the code Claude just wrote. Opens a PR-like web UI at a localhost URL, streams the human's inline comments back into this session in realtime so Claude can ask clarifying questions under each comment, and blocks edits until the human presses Submit (a review idle for 24 hours expires and lifts the block). Use when the user asks to review changes, says "/cc-review:start", "review this", "let me review before you change anything", or wants to give feedback on a diff before Claude proceeds.

Investigates the user's past Claude Code sessions via the cc-transcript CLI. Triggers on questions like "what did I ask Claude yesterday", "analyze my past sessions", "find the session where we fixed X", "how often do I interrupt", "which tools do I use most", and on debugging hooks, session resume, compaction, sidechains, or token blowups from transcript evidence. Rule, never read a raw .jsonl transcript directly — use list, stats, grep, then show.

Start or resume a {{DISPLAY_NAME}} session over the code Claude is working on. Opens a localhost web UI, streams the human's feedback back into this session in realtime so Claude can ask clarifying questions, and blocks edits until the human submits. Use when the user asks to start {{DISPLAY_NAME}}, says "/{{SKILL_NAME}}", or wants to give feedback before Claude proceeds.

Drive cc-orchestrate (the cco CLI) to spawn, watch, message, and kill fleets of Claude Code agents across backends (herd, superset, cmux, zellij, tmux). Use when orchestrating agents with cco or cc-orchestrate — creating repos, workstreams, and sprints; spawning agents into git worktrees; streaming agent status; or driving a fleet from a parent session over MCP.