.Config

Create Product Requirements Documents (PRDs) that define the end state of a feature through iterative design interview. Use when asked to "write a PRD", "spec this feature", write a requirements doc or RFC, or when planning new features, migrations, or refactors that need a reviewable spec. Generates structured PRDs with acceptance criteria, testing strategy, and architectural decisions.

Control interactive CLIs and TUIs via tmux - safely target panes, send keystrokes, scrape output, and operate modal prompts with observe-before-commit guardrails. Use when a command needs a real PTY (REPLs like python/psql, debuggers like gdb/lldb), when an installer or wizard presents modal prompts or menus, when a long-running interactive command must be watched without blocking, when something asks for confirmation or a permission prompt mid-run, or when a tool has no non-interactive flag and the only path is driving its TUI. For observing or prompting coding-agent panes, use the coding-agents skill.

Channel the mental models, decision frameworks, and communication style of real experts (Steve Jobs, DHH, Rich Hickey, etc.) to approach problems the way they would. Use when the user says "summon", "channel", "what would [name] think", "ask [name]", or wants an expert perspective on a problem.

Scaffolds a loop directory for automated agent task execution. Use when asked to "create a task loop", "set up a loop", "scaffold a loop directory", "prepare tasks for rl", or "set up automated execution" for a backlog. Takes an existing backlog and generates PROMPT.md (loop contract), run-log.md (execution history), and .gitignore for ephemeral loop-state.md.

Decompose input into a structured task backlog for automated agent loops. Use when asked to "create a task plan", "break this into tasks", "decompose this PRD", "turn this into a backlog", or "plan tasks from" any input source (PRD, Figma feedback, GitHub issues, user requirements). Also use when the user provides a PRD, design doc, or requirements and wants executable tasks, even if they don't mention "task plan" explicitly.

Gets an adversarial, report-only code review of the current change from a different AI model (Claude, Codex or opencode), run headless and read-only, and relays the findings. Use when the user asks for a second opinion, a cross-model or adversarial review, to "have another model check this", red-team a diff, or review working-tree changes, unpushed commits, or a branch before committing or opening a PR - especially to pair a review rubric skill (e.g. thermo-nuclear-code-quality-review, code-review) with a different model via skl. Not for shipping a session to another host (that is agent-teleport / atp), branching your own session into a pane (tmux Alt+b), or auto-applying fixes - this reports only.

Finds what is consuming disk space on this Mac and reclaims it safely. Use when the disk is low or full, or the user asks to analyse disc usage, find what's eating space, free up space, or clean up their Mac - including mentions of ncdu, du, dust, or "startup disk full". Not for pruning a single project's build output, or for disk on a remote server.

Captures a brand kit from a company or product name, domain, URL, website, or brand-guidelines link: official source resolution, logos, favicons, social images, screenshots, colours, typography, visual style notes, design tokens, and a basic on-brand page brief. Use when the user asks to copy, capture, match, recreate, clone the vibe of, extract tokens/assets from, or make a page feel on-brand from a web link or just a brand name. Not for phishing, fake affiliation, counterfeit pages, or exact replicas of protected flows.

Observe and drive coding agents (Claude Code, Codex) running in tmux panes: list live agents, wait for one to finish or need input, send it a prompt, or name a pane. Use whenever a task involves another agent's pane - "is the agent done", "when it finishes", "send this to the other agent", waiting on an agent before a follow-up step, or scripting multi-agent hand-offs. Owns the `agent` CLI and the `@agent_state` vocabulary. For driving any other TUI (REPLs, wizards, editors) use the tmux skill.

Downloads images from GitHub issues and PRs using authenticated signed URLs. Use when a user shares a GitHub issue/PR URL and the agent needs to view screenshots or image attachments, especially from private repos, or asks to "grab the screenshots from that issue" or similar.

Create and manage Hetzner Cloud servers. Use when creating VPS/cloud servers, managing Hetzner infrastructure, or setting up dev/remote servers. Requires hcloud CLI.

Picks the right local tool and command to mirror, archive, or snapshot a website or page for offline use, and avoids the silent failure where a saved page renders blank with no network. Use when asked to mirror, archive, download, save, or snapshot a site or page for offline viewing, make a browseable offline copy, or grab a page as a single self-contained file - covering wget, monolith, firecrawl and yt-dlp, robots and soft-block handling, and the JS/API offline-content trap. Not for live web extraction into markdown when no local copy is wanted - use firecrawl directly for that.

Scaffolds a SUPERVISOR.md runbook for watching a long-running agent loop in tmux. Use when asked to "supervise a loop", "watch a loop", "babysit a loop", "set up a supervisor", or when a task-loop / rl run needs someone operating the harness around it.

Sweep the authored skill catalogue for truth decay - re-verify dated caveats, version literals, counts, and executable claims against live sources, and re-baseline skills against what the model now does unaided. Use when asked to check skills for staleness, run a freshness sweep, re-verify skill claims, or after a major model upgrade; also for a single skill when revising it feels risky. Not for vendored skills (update-vendored-skills owns upstream refresh) and not a lint (writing-skills' check.sh owns phrasing).

Bundles an HTML page or URL into one self-contained, shareable HTML file by embedding images, icons, CSS, JavaScript, fonts, and other page assets as data URIs. Use when the user asks to make a webpage portable, offline, standalone, single-file, self-contained, shareable, or not depend on an assets folder. Not for rebuilding or redesigning the page.

Creates video thumbnails (YouTube, LinkedIn) from a talk or screen recording: scouts frames with ffmpeg contact sheets, cuts out the speaker with rembg, composites panels/logos/big text as HTML/CSS and screenshots with playwright-cli. Use when the user asks for a thumbnail, cover image, or social preview for a video, talk, or upload, or wants to match a reference thumbnail's style. Not for image edits with no video source.

Product positioning, customer-research-driven messaging and website copy, and brand identity briefs, informed by the work of April Dunford, Georgiana Laudi, Claire Suellentrop, Michael Johnson and Alina Wheeler. Use when positioning or repositioning a product, choosing a market category or target segment, planning customer or JTBD research, writing or rewriting homepage, landing-page or sales copy, a value proposition or tagline, naming a product or company, briefing a rebrand, logo, visual identity or brand architecture decision, or critiquing existing positioning, messaging, a name or a brand against expert criteria. Not for ad campaigns, SEO, or channel/growth tactics.

Renders investigation findings as honest tickets, issues, bug reports, and incident writeups filed in the user's name: every claim carries its evidence, hedges mark real uncertainty instead of smoothing over gaps, and proposed fixes never outrun the diagnosis. Use when drafting or reviewing a ticket or writeup, when deciding how confident a claim may sound, when the user asks whether a ticket overclaims or is bullshitting, or when writing first-person prose on the user's behalf. Not for conducting the investigation itself - that is prove-it; this skill owns how the result is written down.

Authors fake terminal sessions and coding-agent TUIs that render to SVG, PNG, or MP4 via termctrl, without running the real program. Use when you need a fabricated, idealised terminal image or recording - a mocked-up Claude Code / Codex session, or a fake shell / REPL / CLI transcript for a slide, doc, README, or blog - instead of capturing a real run. Covers authoring raw ANSI stills and hand-built .termctrl timelines. To capture or render a REAL running program, use the terminal-control skill instead.

Concrete rules for building or restyling UI in code: visual hierarchy, spacing and layout, type scales, colour palettes, depth and shadows, images, and finishing flourishes. Consult WHILE building, styling, restyling, or design-reviewing an interface - whenever something needs to "look better", feel more polished, or when setting up spacing / type / colour systems. Framework-agnostic, with a separate Tailwind v4 mapping.

Safely review and update sha-pinned tmux plugins. Use whenever the user mentions tmux-upstream, tmux plugin pins or updates, a tmux plugin being commits behind upstream, pin_tmux_plugin in home-shared.nix, asks whether a tmux plugin update is dodgy/compromised/safe, or asks to bump/update a tmux plugin. Default to review-only and ask before bumping unless the user explicitly requested automatic safe updates.

Safely refresh the vendored third-party agent skills in this dotfiles repo. Use whenever the user wants to update, refresh, upgrade, or re-pull vendored skills (`skills update`), or asks to check whether a skill refresh is safe / dodgy / compromised before committing. `skills update` is an unauthenticated git clone with no quarantine, no signature, and no scan - and skill files are instructions injected into every agent session - so this skill gates each refresh by reading the diff and only commits clean ones, holding dodgy diffs for sign-off.

Analyze OpenCode conversation history to identify themes and patterns in user messages. Use when asked to analyze conversations, find themes, review how a user steers agents, or extract insights from session history.