Shape agent capabilities into your workflow.

debrief is a conversational learning ritual, not a document generator. It narrows the session to a few ideas, asks you to reconstruct each one in your own words one question at a time, reveals and corrects gaps only after the attempt, and distills a transferable principle with an optional durable note. It deliberately stays out of mechanical tasks, mid-task flow, and ordinary status recaps.

codex-image-gen drives an interactive Codex pane through tmux, requires an explicit absolute save path, waits for the file, and verifies the image and target aspect ratio with ffprobe or PIL. If the expected file is missing, its helper searches the Codex generated-images cache and can recover base64 data from the newest session rollout JSONL. It is a reliability wrapper around an interactive workflow, not a standalone image model.

trip-planner builds an itinerary around geography, dates, and logistics before filling it with attractions. It red-teams the plan for long transfers, closures, holiday traffic, weather, vehicle capacity, and other failure modes, then verifies places against the actual travel dates. The output is a clean group-friendly PDF with route context, book-now warnings, and Google Maps links that open when a traveler taps each stop.

launch-doctor-kr scans a Korean Android app before Google Play submission, with checks for missing Maps keys, outdated targetSdk, release debuggable flags, Capacitor Hangul input issues, location-service reporting, AI-content labeling, and privacy-policy gaps. Findings include a file location, a suggested fix, and a supporting link. It is aimed at small teams using Capacitor, React Native, or Flutter.

visualizing-data helps an agent choose a chart from a message-first framework and a catalog of more than 24 chart types. It covers decision trees, annotations, operational dashboard patterns, perception principles, colorblind-safe color systems, performance, and HTML or document output. The emphasis is on making the intended message legible and accessible, not merely decorating a dataset.

genai-prompt-optimizer finds every LLM call in a GenAI app, pins one model per run, and derives adversarial cases from the app’s prompts and JSON schemas. It first records the expensive model as an oracle, then iterates on the cheaper student’s prompt and re-verifies each cohort with deterministic checks, semantic comparison, and limited judge review. The result reports measured token cost and projected savings; it asks for consent before the token-heavy run.

schedule-reminder is a persistent, queryable store for todos, events, deadlines, and progress rather than a calendar UI. Its state machine covers pending, doing, done, blocked, and cancelled; a local Discord relay dispatches due reminders, while other skills can use the stable JSON CLI. It also handles idempotent ticks, missed-fire catch-up, retries, and concurrent read/write checks without becoming a cloud service.

box-rclone-binder uses Box server authentication so each host mints its own short-lived access token instead of sharing a rotating OAuth refresh token. The box-binder CLI deploys one Box drive to multiple servers through rclone, runs read-only health checks, performs bounded refresh or self-heal, and can schedule checks with cron or systemd plus Discord alerts. Fleet configuration stores pointers to secrets rather than secret values.

screen-vision treats the operating system’s accessibility tree as the primary source for desktop UI names, states, types, and rectangles, using OCR or icon vision only to fill gaps. Its CLI probes the host, captures a screenshot plus structured elements, and can click by element ID with dry-run as the default. It targets native, Win32, WinUI, Electron, game, or remote-desktop windows—not browser pages, which should use Playwright.

promotion-assistant coordinates email, posts, forum replies, and direct messages for a shipped product, tracks a six-layer conversion funnel, and uses Thompson Sampling to adjust tactics from measured feedback. It keeps copy, audiences, policies, and credentials in a separate private config repository. Dry-run is the safe default: no outbound action is allowed until the product and that individual channel are explicitly enabled, so it is neither a spam cannon nor a scheduling engine.

email-monitor incrementally watches Gmail over read-only IMAP using UID watermarks and message deduplication, then classifies new mail through L0 rules, L1 cheap scoring, and an optional L2 LLM hook. It sends only a redacted title to Discord and can archive or summarize messages, but replies remain drafts for the user to send. Mail bodies stay with the local model, while account topology and secret pointers live in a private config repository.

daily-hotspots gathers opportunity candidates from multiple sources, merges them across days, applies a reproducible scoring rubric, and delivers tiers through Discord with a private archive. A fail-closed gate requires at least two independent origins and keeps deep dives within a daily limit; deeper research is delegated to market-intel or small-cap-deepdive. It owns cadence and ranking, not the underlying search or verification engines.

auto-support is a Claude Code plugin for answering a product’s Discord users from an allowlisted set of public documents. It uses permissions, a fail-closed tool hook, standard-library detection, and an egress DLP gate so the model cannot open secrets or source files. Out-of-scope or uncertain questions are refused and escalated to founders; the MVP produces drafts for a user to review rather than auto-posting.

demand-mining turns a shipped product’s Discord conversations into a daily demand radar. It removes PII before the model sees messages, extracts explicit and implicit jobs-to-be-done, deduplicates across days, and ranks the pool with RICE, Opportunity, WSJF, and Kano views. The result is an end-of-day brainstorm and prioritized iteration queue; search, verification, and listening are delegated to companion tools.

five-step-algorithm forces a process review in the order question, delete, simplify, accelerate, automate. It prevents addition bias by making removal and simplification precede new wrappers, retries, agents, or optimizations. A Backward Check stops repeated rounds of fixes and asks whether the problematic step should exist at all. Use it before extending workflows, building agents, or carrying dead requirements into a migration.

seedance-storyboard converts a screenplay, scene description, ad concept, or even a one-line idea into a storyboard for Seedance or Jimeng text-to-video generation. It first fixes the global visual style, palette, mood, and continuity constraints, then lays out a timed shot sequence with framing, camera movement, subject action, and detail. The result is a shot-writing aid, not the video renderer itself.

ideation-engine reads the shape of an ideation request and routes it to methods such as affinity mapping, product framing, or provocation. It is strongest on transcripts, interviews, social posts, and research notes: the output must retain real proper nouns, concrete mechanisms, and implementable details instead of generic categories. A quality gate screens out vague ideas and helps turn raw source material into feature or experiment candidates.

design-effects-skill teaches an agent to select visual effects for UI and branding contexts, with decision rules that also say when not to use them. Its references cover animated backgrounds, text and scroll motion, liquid or glass treatments, shader gradients, and 3D UI, with implementation examples tied to common libraries. A browser-ready examples page lets you preview the effects without a build step.