Explore

How to read winnow stubs in tool results. A block marked "[winnow] ... hidden" was judged unlikely to matter for the current task and replaced with a short summary plus a recall key; call winnow_recall to restore it. Use when a tool result contains "[winnow]" lines or when winnow_recall / winnow_stats tools are available.

Edit AI-sounding text so it reads like a careful person wrote it, without changing what it says or commits to. Built for work writing where the wording matters: sales and marketing copy, support replies, HR and legal letters, emails, docs, release notes, posts, landing pages. Two modes. Rewrite: edit the text you are given. Interview: when a draft is AI-composed and thin, ask the writer 3-5 sharp questions and build the piece from their answers, in their words. Also writes from the writer's own notes or a voice-memo transcript. Calibrated on measured differences between human and AI writing.

Edit AI-sounding text so it reads like a careful person wrote it, without changing what it says. Two modes. Rewrite: edit the text you are given (posts, emails, docs, release notes, replies, landing pages). Interview: when a draft is AI-composed and thin, ask the writer 3-5 sharp questions and build the piece from their answers, in their words. Calibrated on measured differences between human and AI writing.

Edit AI-sounding text so it reads like a careful person wrote it, without changing what it says. Two modes. Rewrite: edit the text you are given (posts, emails, docs, release notes, replies, landing pages). Interview: when a draft is AI-composed and thin, ask the writer 3-5 sharp questions and build the piece from their answers, in their words. Calibrated on measured differences between human and AI writing.

Use when the awwwards MCP server or its skills are invoked for the first time in a session or on a new machine — when no awwwards preferences exist yet, or the user asks to set up, configure, or change awwwards-mcp settings (result density, viewport, live captures, local index, winner watches). First-time onboarding asks the user a short set of preference questions, persists them to ~/.awwwards-mcp/preferences.json, and runs any one-time installs they opt into.

Use when building a site whose reference has animation — before writing any motion code, and again when verifying a build. Teaches what to record from a live site (and what to skip), how to study the recordings frame-by-frame (video input or ffmpeg tile-per-element), how to turn the study into a motion inventory, and how to verify the build's motion against the reference. Complements awwwards-inspiration (step 6) with the full capture→review→build procedure.

Use when building a site that needs design references (e.g. "make it feel premium", "dark 3D portfolio vibes"), when the user asks for standalone inspiration research ("show me award-winning e-commerce sites", "what's trending in brutalism"), or when setting up or maintaining the awwwards-mcp local index or live-capture tooling (awwwards-index, capture_live_site).

Use when awwwards-mcp scraping or tools stop working — BlockedError, "layout may have changed" parser errors, empty results, failed captures, stale searches — or when the user reports the MCP behaving oddly. Guides running the doctor/drift/index scripts, applying their fixes, re-anchoring parsers after real awwwards.com drift, and recovering the in-flight task that surfaced the failure.

Coleta, pré-visualiza e entrega B-rolls com revisão humana e origem registrada. Use quando alguém pedir b-roll, vídeos de apoio, imagens de apoio, cutaways, inserts, footage, "um corte do X falando Y", um print da tela de um site ou de uma notícia, ou material para ilustrar um vídeo, Reel ou aula — buscando em YouTube, Instagram, TikTok, Wikimedia Commons, NASA ou bancos (Pexels, Pixabay), gerando prévias para revisão humana e entregando os trechos com origem e condições de uso. Also in English: collect B-roll, cutaways, inserts, supporting footage, stock video, screen grabs. Não serve para editar, montar ou renderizar o vídeo final. Not for editing or rendering the finished video.

Check that a web page you wrote or changed actually works, by opening it in a browser and driving every control on it. Use after creating or editing an HTML page, or the JavaScript or CSS that a page loads.

The definition of done. Run before calling any change done, fixed, verified, ready or shippable, for features and bug fixes alike. Checks the pre-mortem was answered, proves the tests fail on the old code, gets a fresh breaker review, runs CI's own checks in a clean checkout, checks the user-facing words, and writes the evidence report. Also use when asked "is it done", "is it ready", "did you verify it" or "can we merge".

Rewrites the prompt typed after the command before any work starts, then works from the rewrite. Splits a mixed-up message into a numbered list of asks, turns words like "be 100% sure", "cover all cases" or "don't break anything" into the checks they stand for, keeps every name, path, number and quote exactly, and stops to ask only where it would have to guess. Runs only when the user types it.

Reviews a change before it is merged, from little or no input. With nothing typed it reviews the branch you are on, uncommitted work included, against its base, its PR and its ticket, so a developer can check their own work before asking for review. With PR links, repo#N or a branch it reviews someone else's change the same way, several repos at once. It judges the failed CI checks and every Bugbot and reviewer comment, gets a fresh breaker review, traces every other piece of code that uses what changed, proves each finding or marks it unproven, says what the merge needs (a build, env vars, migrations, deploy order) and how to check the deploy, never posts to GitHub, and offers fixes only after the report. Runs only when the user types it.

Sets up, checks or changes the optional Jev judge, the TypeSafe decision model that ship-check asks about review findings (is this real harm, is a small one worth fixing now, what proof does a small fix need). Shows whether it is on for each repo, which key it uses, switches it on or off, and sends one test request. Runs only when the user types it.

Lifts the task's scope for one task. The session, its reviews and fix-the-class then look across the whole codebase and every repo that shares its code, data or vendors, and every problem found anywhere is handled by the usual rules. Runs only when the user types it.

Learns the words a user habitually writes to their coding agent that make its answers worse ("be 100% sure", "don't assume", "full review", "are you sure?", "all fine, right?"), from what they actually typed in their recent Claude Code sessions. Shows what each one does, what went wrong after it, and what to say instead, then writes the habit words block that maps each word to the checks it should trigger, so the checks run without the words. Use when asked to learn, read, refresh or review the user's habit words or prompts, "which of my words make you worse", from setup-first-pass's profile step, or when the start-of-session check says the habit words are due.

Bug-fix routine that fixes the whole class of bug, not just the reported instance. Use for any bug report, failing production behaviour, monitoring alert, audit finding or review finding that does real harm (a smaller review finding goes on the item's list, as ship-check says), and whenever the same kind of bug has been seen before. Reproduces it with a failing test, names the failure class, searches the task's scope for the same pattern (the whole codebase when the user's own prompt asks for the fix, not when it is picked from a task's list), fixes or records every hit, and adds the rule, invariant, helper or check that stops it coming back.

辞达 · 现代中文高质量对话型表达引擎。用于中文写作、重写、润色、诊断与文体校准： 文章重写、段落优化、口述/访谈转可发表文章、书面稿自然化、公众号/小红书/知乎/ 视频口播/播客等平台适配，以及按用户真实文本校准个人风格。本 Skill 不做"去AI味"， 不做AI检测对抗，不模仿任何在世作者；它以"辞达"为纲——先诊断，再按 P0意义逻辑→P1结构清晰→P2流转语域→P3节奏修辞→P4打磨的顺序提升文本， 目标是让有价值的思想被说得清楚、自然、漂亮、亲近、有分量。 Use for Chinese conversational-quality writing, rewriting, diagnosis, and style calibration.

발표 브리프를 스크롤로 진행하는 발표 덱(Scrolline v2)으로 만든다. 무대 1개·연속 timeline 1개·handoff move·컨셉 6종(rail·data·launch·zoom·canvas·kinetic)별 움직임 문법. deck.yaml 작성 → node plan 확인 → 브라우저 smoke → HyperFrames MP4 내보내기까지. 트리거 — "스크롤 덱 만들어", "스크롤 발표", "스크롤텔링 덱", "스크롤로 넘어가는 발표자료", "연표 덱", "데이터 스토리 덱", "제품 발표 스크롤", "줌 드릴다운 발표", "덱을 MP4로", "scrolline", "scrolline deck", "scroll deck", "scrollytelling deck", "scroll-driven presentation", "export deck to mp4". 후속 — "이 beat만 move 바꿔", "정지점 추가", "홀드 길게", "컨셉 바꿔서", "smoke 다시", "MP4 다시 뽑아" 도 이 스킬. ※ 이산 슬라이드·발표자 노트 중심이면 slideshow, 덱과 무관한 영상 제작이면 hyperframes 소관.

Cut a prismantis release - version bump, changelog, verify, tag and push. Use when asked to "release", "bump the version", "ship 0.x" or "publish".

Load the working copy of prismantis into a real Claude Code session and screenshot how replies render. Use after any visual change, when asked to "screenshot", "see it", "try it live", or before claiming a rendering fix works.

Print the prismantis demo reply to eyeball every rendering feature after a change. Use after any visual change, when asked for "the demo", "show all features", "visual test", or before a release screenshot.

Edit any video by conversation. Transcribe, cut, color grade, generate overlay animations, burn subtitles — for talking heads, montages, tutorials, travel, interviews. No presets, no menus. Ask questions, confirm the plan, execute, iterate, persist. Production-correctness rules are hard; everything else is artistic freedom.

Use when the user wants a Khan Academy style explainer video, a blackboard or whiteboard lesson, a hand-drawn chalk-talk animation, a short narrated video that teaches a concept ("explain X in a video", "explainer video", "teach X in 30 seconds"), or a tutorial that draws hand-written arrows, circles and captions over product screenshots ("draw over the UI", "annotated walkthrough"), or clips cut from a YouTube video or podcast where the speaker's real audio plays under a blackboard drawn to it ("khanify this video", "clip this with the Khan visual"), especially when it must be free, local, and use no paid APIs or keys. macOS only.

Manual desktop QA with CUA. Use only when explicitly requested.

An ongoing ADHD-friendly study companion for any subject, inside a normal Claude chat. Use when the user wants to study or learn something, be quizzed or reviewed, wrap up a session and reflect, or check in on their learning patterns. Say things like "let's study X", "quiz me", "I'm done for today", or "how am I doing lately". Runs spaced-repetition warm-ups, small-chunk teaching, teach-it-back reflection, and keeps progress in a copy-paste Progress Card.

Use before a step that judges material you already have — a claim against evidence, screening fetched text before reading it, classifying or keeping or dropping many items, finding or reranking candidates, one bounded choice, comparing two passages, grading on an ordered rubric, judging a file without reading it into your context, extracting a regex-matched value, reviewing a patch, or checking that work finished. Batch every question about one state into one call. Jev is invoked when an unresolved judgment earns a model decision. Deterministic evidence takes precedence; Jev is not a mandatory ceremony. Write the text yourself when the step produces new words, code, or options you cannot list. Not for building on the Jev API; that is the jev skill.

Use when code needs a judgment about natural language that rules or regex cannot make (classify, route, triage, moderate, score, rank, match, dedupe, filter, extract, verify, or gate an action), when brainstorming where AI could fit in an app, when an LLM call exists only to pick a label, score, or yes/no, or when code uses TypeSafe AI, Jev, System One, typesafe-sdk, or @typesafe-ai/sdk.

Invoke whenever writing, changing, reviewing, or sweeping tests in jev-judge-mcp. Authoring gate for new tests plus audit workflow for low-value, implementation-coupled, or duplicative tests and the test-only production seams they demand, under this repo's rule that a test is never deleted or weakened to make a failing build pass.

用本地命令行 mcd-git（git mcd --json）帮用户点麦当劳：选门店、看菜单、加购物车（去酸黄瓜、少冰、套餐换搭）、自动凑套餐挑最省的券、核价、下单拿支付链接、查取餐码、再来一单、取消订单、领券（pull，先确认），以及生成 1024 吃麦贡献图（card）。用户提到点麦当劳、麦当劳到店 / 得来速 / 麦乐送、取餐码、再来一单、取消麦当劳订单、领麦当劳的券、吃麦贡献图时使用。下单、取消和领券都必须先把确认摘要给用户看，用户明确同意后才能带 --confirm 执行。

王佩专用的去 AI 味系统。触发词：去AI味、去味。合并三家——王佩写作系统 v4.0、humanizer-zh、lieflat-less-ai-tone（含实测数据）。先选模式（自家稿 / 白名单），再过硬边界，再下刀。不判断作者身份，不保证通过 AI 检测器。

Work out what is actually worth defending in this system — the trust boundaries data crosses, the assets behind them, what an attacker can already do, and the small number of threats that justify a control. Use when designing or changing a system's shape, before building auth, payments, uploads, file sharing or multi-tenancy, when adding an integration that crosses a boundary, or when asked whether a design is safe. Not for finding bugs in code that already exists (that is secure-code-review), not a compliance questionnaire, and not incident response.

Review a change or a codebase for security defects with a deliberate reading order — where bugs cluster, trace source to sink, confirm reachability, and write a finding someone can actually fix. Use when asked to security-review a pull request, feature or repository, or before shipping anything touching auth, money, tenancy, uploads or secrets. Not a scanner run (its output is leads), not design-stage threat modeling, and not a penetration test — nothing here authorizes touching a running system.

Hunt the web failure classes that actually recur — broken access control, injection, XSS by output context, SSRF, unsafe deserialization and mass assignment, session and token handling, secrets leakage, file handling — as patterns to find in this codebase. Use when writing or reviewing request handlers, queries, URL fetches, templates, uploads or auth code, or when asked to check a web app for the common vulnerability classes. Not for design-stage modeling (threat-modeling), not infrastructure or network hardening, and not a list to recite without reading code.

Judge dependency risk instead of reciting it — resolve the advisory to a real path through the lockfile, decide whether the vulnerable code is reachable in this codebase, and pick upgrade, override, removal or a written acceptance. Use when an audit or advisory alert fires, before adding a new dependency, when a lockfile change needs reviewing, or when someone asks whether a named CVE actually affects this project. Not for finding vulnerabilities in code you wrote, not for secret scanning, and a clean audit output is not a claim that the application is secure.

Attack and harden an existing auth surface — session fixation and rotation, token verification, horizontal and vertical privilege escalation, password reset and account recovery, MFA bypass. Use when reviewing login, session, token, reset, invite, impersonation or role-elevation code, when someone reports seeing another user's data or an account takeover, or when auth changes are about to ship. Not for designing the login mechanism or permission model in the first place (authentication, authorization), not for infrastructure IAM, and never run against a system you have not been told you may test.

Secure an agent system as a permission surface — what authority each tool call runs under, where the confused deputy sits, which controls are enforced outside the model and which are only prompt text, and how far one bad call reaches. Use when granting an agent tools or credentials, wiring in MCP servers or subagents, reviewing an agent that acted beyond what the requester could have done, or before letting an agent touch a shared or production system. Not for wording the ingestion trust boundary in detail (prompt-injection-defense), not for scoping what an agent is for (agent-design), and it never grants an agent permission it did not already have.

Decide what to test and at which level before any tests get written — enumerate failure modes by consequence, give each one the cheapest level that can actually observe it, and name what is deliberately left uncovered. Use when starting testing on a feature or a codebase, when asked for a test plan or why coverage sits where it does, or when a suite is large and still missing bugs. Not for writing the tests themselves, and not for proving a finished change works.

Write unit and integration tests that fail for the right reason — behaviour-level assertions, boundary values, the smallest honest double, deterministic setup and teardown. Use when adding tests for new code, pinning a bug before fixing it, or repairing tests that stay green while the code is broken. Not for deciding what to test or at which level, not for browser journeys, and not a way to call a change verified.

Find the cause of a failure by evidence — reproduce it, read what the system actually did, form hypotheses that predict something you have not looked at yet, bisect to the smallest failing case, fix where all callers route through, then re-run the original reproduction. Use when something is broken and the cause is not yet known, when a previous fix did not hold, or when an edit is about to be made because the code "looks wrong". Not for an outage still burning (mitigate first), not for a cause already established that only needs implementing, and not a licence to skip reading the stack trace.

Turn a fixed bug into a test that fails on the old code and passes on the new one. Assert behavior at the cheapest useful level and compare old behavior without reverting unrelated work or leaving scratch files. Use after a fix or when recurring defects lack coverage; not for a new feature's test strategy or an environmental/cosmetic failure.

Find where the time actually goes before changing anything — define the metric and workload, build a repeatable baseline with its spread, profile on the right axis, confirm the suspected hotspot by removing its work, then re-measure under identical conditions. Use when something is slow, when an optimization is about to be written from a hunch, or when a speedup is being claimed without comparable numbers. Not for reading code to guess what is expensive, not for micro-tuning with no measured user impact, and not a licence to profile production without asking.

Build and keep end-to-end suites that are worth their runtime — which journeys belong at this level, selectors that survive refactors, waiting on state instead of sleeping, per-spec data isolation, and flake triage that finds the cause. Use when adding or repairing an e2e suite, when specs are slow or intermittently red, or when deciding whether a case belongs here at all. Not for one-off checking that a change renders, not for unit or integration tests, and not for load testing.

Prove a UI change actually renders and works — desktop and mobile, controls operated, console clean. Use before reporting any change to rendered UI as done, when asked whether a screen actually functions, or when checking someone else's frontend work. Not for judging whether a design is good, and not a substitute for tests you intend to keep.

Choose what not to do and make the tradeoff legible — the constraint being spent, each candidate's cost and value with the evidence behind it, what is reversible, and what would change the answer. Use when a list is longer than the capacity, when asked what to cut or what ships first, or when a decision was made and nobody can reconstruct why. Not for deciding whether a problem is real (product-discovery), not for writing the scope of what wins (prd-and-stories), and not for committing the team on their behalf.

Write a scope someone else can build from without asking you what you meant — problem, users, the flow end to end including its unhappy states, acceptance criteria phrased as observable behaviour, and explicit non-goals. Use when turning a settled problem into a brief, spec, PRD, epic or set of stories, or when repairing a ticket that is too vague to estimate or to check. Not for deciding whether the problem is real (product-discovery), not for ranking work (prioritization), and not for choosing the implementation.

Design an A/B test that can answer its question — hypothesis, randomization unit, power, duration, pre-registered analysis — and read the result honestly, including a null. Use when someone proposes testing a change, asks how long a test must run or whether a result is real, when a test is about to be stopped early, or when a flat result is being read as "no difference". Not for defining the metric itself (product-analytics), not for a change too small or too rare to power, and not for authorizing a rollout.

Write or repair documentation that stays true to the system it describes — README, setup guide, how-to, API reference, architecture explanation, release notes, runbook — by fixing the reader and the task first, then grounding every factual claim in code you actually read. Use when documentation is being written, restructured or found to be stale, or when a reader cannot get from the docs to a working result. Not for interface strings inside a product, not for persuasive or marketing copy, and not itself proof that the documented commands work.

Judge whether a source can carry the weight you are about to put on it — trace it upstream to something primary, check its date and version, name who benefits if you believe it, test whether corroboration is actually independent, and cite it so a reader can reopen exactly what you read. Use before relying on a claim that changes a decision, when two sources disagree, when a number arrives without its method, or when writing citations. Not for running the overall investigation, and not for deciding between products.

Make a page findable for a real query — match the intent behind it, structure the page around that intent, and clear the technical floor that decides whether it can be crawled and indexed at all. Use when a page is written for search, when traffic drops or a page is missing from results, when someone asks "how do we rank for X", or before publishing content meant to be found. Not for writing the copy itself, not for page speed work, and not for paid search or social distribution.

Say what a product is for, which alternative it beats, and why — in one segment's own words, with every claim tied to something the product actually does today. Use when a homepage or deck says what the product is, when "who is this for" or "how is this different" has no crisp answer, when entering a new segment, or when sales keeps losing to one competitor. Not for writing the final copy, not for pricing, and not for deciding what to build next.

Get a defensible answer out of a dataset — grain and denominators pinned, missing data accounted for, aggregates checked against their segments, and every number reproducible from raw input by a script. Use when asked what a dataset shows, when a number needs explaining or two numbers disagree, or before any finding from data is reported to someone who will act on it. Not for defining product metrics or instrumentation, not for building the pipeline that produced the data, and not for designing an experiment.

Compare products, libraries or approaches on criteria derived from your own constraints rather than from anyone's feature list, and fill each cell with evidence you obtained — hands-on trial, primary docs, source, issue tracker — marking clearly what is only a vendor claim. Use when choosing between real alternatives, when asked how something stacks up against the competition, or when a decision is being made from a marketing comparison table. Not for general investigation method, not for judging a single source, and not for writing positioning or marketing copy.

Prove a UI looks right across every state and viewport it has to survive — enumerate the state matrix, make captures deterministic, compare against a baseline captured the same way, and triage every diff. Use before calling a visual change, a style refactor or a redesign done, when running or trusting a screenshot-comparison suite, and when asked whether a screen still looks correct. Not for judging whether a design is good, not for proving controls actually work, and never on its own sufficient to call a page verified.

Non-animation creative direction for HyperFrames videos. Use for design spec (frame.md / design.md) handling, palettes, typography, narration, beat planning, audio-reactive visuals, composition patterns, and brand / style decisions. For atomic motion patterns and scene blueprints, use `hyperframes-animation`.

Mandatory entry point: read this first for any request to make, create, edit, animate, or render a video, animation, or motion graphic, including a promo, explainer, captioned clip, title card, overlay, slideshow or interactive deck, Remotion port, or any HyperFrames HTML composition. Also use it to inspect, diagnose, validate, preview, publish, or batch-render an existing HyperFrames project. Inputs may be a website URL, GitHub PR, Figma design or URL, text or brief, existing footage, or music. It resumes project state, captures intent when applicable, selects and installs the owning workflow, and routes domain capabilities. HyperFrames is the default output framework unless the user explicitly chooses another framework for the deliverable or asks only to record a browser session.

Author or edit a custom HyperFrames composition when no specialized workflow fits, or when BRIEF.md sets flow: companion. Use for longer or multi-scene pieces, brand and sizzle reels, montages, static loops, static title cards, footage remixes, and freeform builds. Use motion-graphics instead for a short unnarrated motion-first unit, including an animated title. Route fresh creation through hyperframes before using this skill.

Import Figma content into a HyperFrames composition — rendered assets, brand tokens, components, storyboard sections → reconstructed motion (frames read as states, not slides) (REST/CLI), connector-assisted motion when available, and shaders from a connector or native export. Use when the user pastes a figma.com link or asks to bring a Figma design, frame, logo, brand, or animation into a video/composition.

Turn arbitrary text — an article, notes, a topic, a brief — into a faceless explainer video: there is no site or footage to capture, so the visuals are invented per scene (typography, abstract graphics, diagrams, data-viz). Use for topic explainers, concept breakdowns, how-tos, listicles. Not a video built from a website (/product-launch-video — promo or tour). Unclear → /hyperframes.

Premium 3D product films rendered in Blender (headless, GPU) — dark studio, rim light, light sweeps, glass/liquid/metal materials, condensation and bubbles, macro and orbit camera moves, shallow depth of field, then crisp 2D titles, transitions, music and SFX in the edit. Builds a round product (bottle, can, jar) directly from a user's product photo (real silhouette + projected label), or a procedural one. Use for product launch / hero / commercial films whose reference is 3D or photoreal CG (Apple-style "black stage + a line of light" films), for any request to make a product look expensive, and whenever the video-clone router picks the premium-product-3d style.