sorawit-w
UserAgent-agnostic skills for coding agents, verified on Claude Code, Codex, and Copilot. Includes screenwright, which builds a UI from a design spec, then renders it, audits accessibility, and fixes what does not match.
Categories
Indexed Skills (25)
define
Contextual definition and translation of a single word or phrase, resolved from the sentence or paragraph around it — the true in-context sense, not the dictionary default. The user names a target word/phrase and supplies surrounding text; this skill disambiguates the sense and explains it for a language learner (meaning, why this sense, register, etymology, related senses, difficulty), and can translate it contextually into another language. Triggers on "what does X mean here", "what does this word mean in this sentence", "define X in context", "meaning of X as used here", "translate just this word in context", or any word-sense / vocabulary / reading-comprehension request that supplies a target plus its context. Does NOT trigger on editing, creating, or syncing translation/locale files, adding a language to a product, or i18n JSON/YAML/TS keys — that is the `i18n` skill. Does NOT trigger on full-document translation, or a context-free dictionary lookup with no surrounding text.
ghostwriter
Drafts messages sent AS the user, in their natural personal voice, with zero AI tells. Triggers whenever the user asks to write, draft, reply to, or respond to anything sent under their own name: "reply to this email", "write back", "message my coworker", "respond to them", "follow up with the recruiter", "answer this Slack message", plus DMs, LinkedIn messages, texts, and conversational GitHub PR/issue comments. Accepts an optional style argument — formal, friend, direct, diplomatic, storybook, eli5, or free-form ("warm but firm"); default is the user's own voice. A FORMAL request still triggers this skill (style=formal, human-formal — the AI-tell ban list still applies). Does NOT trigger for: formal documents, reports, specs, marketing copy, blog posts; code, commit messages, or technical docs; support-ticket replies, sales outreach, or brand-voice content (dedicated skills own those); or when the user explicitly asks for default AI style ("polished", "just write it normally").
startup-audit
Fast, code-grounded triage of an ALREADY-BUILT product. Ingests a codebase (primary) and/or live URL (secondary), infers the model from what's built, diffs it against the claimed story, and ships an HTML dossier with a Continue/Pivot/Kill verdict and a Red/Amber/Green band. Claims are confidence-tiered (observed/inferred/unknown) with provenance; the verdict is explicitly OPINION, not financial or legal advice. Use when the user says "audit my startup", "grill my built product from the repo/URL", "should I continue, pivot, or kill this", "does the code match the pitch", or points the skill at a repo/URL asking whether to keep going. The INPUT decides, not the verb: a codebase/URL plus a verdict request is THIS skill even if they say "grill"/"kill". Do NOT use for: pre-build idea validation (`validation-canvas`); a deep adversarial verdict on belief artifacts or to CONFIRM a Kill/Pivot (`startup-grill`, which reads a canvas/deck, not code); the full 5-artifact kit (`startup-launch-kit`).
ai-eval-review
Audit the evaluation layer of an AI product against design-completeness questions: offline criteria, ground-truth quality, online signal, cohort/disparate-impact, adversarial + robustness coverage, and drift detection. Eval-side companion to `ai-ux-review`. Produces an editable Markdown artifact plus a self-contained HTML report under `docs/ai-ux/`. Use when the user asks to "review my AI eval setup", "audit my eval design", "is my AI eval rigorous enough", "responsible-AI eval review", "fairness eval check", or "drift detection design", or ships an LLM/ML feature and wants an eval-rigor check. Invoke even if only one block is named — the others stress-test it. Does NOT trigger for: human-AI UX review (`ai-ux-review`); lean-canvas work (`validation-canvas`); adversarial pre-mortem with a verdict (`startup-grill`); SKILL.md audits (`skill-evaluator`); implementing eval pipelines, writing eval code, or labeling datasets (this skill names gaps, it does not build them).
ai-ux-review
Audit an AI product against structured human-AI design questions — classical UX (user need, mental models, trust calibration, feedback + control, graceful failure) plus the gen-AI integrity surface older frameworks miss: hallucination handling, output verifiability, provenance + citation, prompt-injection exposure, and agent autonomy. Produces an editable Markdown artifact plus a self-contained HTML report under `docs/ai-ux/`. Use when the user asks to "review my AI feature", "audit my AI product", "is my AI experience trustworthy", "responsible AI UX review", or ships an LLM feature and wants a pre-launch trust review. If `validation-canvas` output exists, assume the business model is settled and focus on UX execution. Does NOT trigger for: lean-canvas work (`validation-canvas`); adversarial pre-mortem with a verdict (`startup-grill`); SKILL.md audits (`skill-evaluator`); brand/logo work (`brand-workshop`); building eval pipelines or writing eval code (names gaps only).
brand-workshop
Simulate a collaborative branding team to create a launch-ready brand identity package from a business overview. Use whenever the user asks to create a logo, brand identity, tagline, or brand concept for a business, product, or app — "help me brand this", "I need a logo", "come up with a tagline", "brand identity for my app" — or gives a business description and asks for visual identity work. Runs a multi-role workshop producing: a brand strategy brief, tagline, code-generated SVG logo, favicon pack, optional mascot + character sheet, social banner set (OG/X/LinkedIn/Instagram), descriptions pack, and a starter DESIGN.md. Use it even for just a logo or just a tagline. Do NOT trigger for pitch decks or investor materials (hand off to `pitch-deck`). Do NOT trigger on revising a LIVE brand — "refresh our brand", "update our style guide", "evolve our identity", "audit our logo" — those go to `team-composer` (@brand_strategist / @senior_copywriter) and the `brand-voice:*` family. Generates net-new identity only.
gamification-fit
Analyze a product, idea, or resource set (code, docs, URLs) against a stated goal and produce a restraint-first report on where gamification genuinely fits, and more prominently where it deliberately does NOT. Most features should not be gamified; the value is discrimination, not idea volume. SDT-anchored; avoids default Points/Badges/Leaderboards; refuses manipulation (non-droppable ethics veto). Produces an editable Markdown artifact plus a self-contained HTML report. Requires TWO inputs: the resources AND a goal/metric (it infers and confirms the goal if missing). Use when the user asks to "gamify my product", "where should I add gamification", "should I gamify X", or points the skill at a codebase/docs/URL plus a goal. Does NOT trigger for: UX review of an existing AI feature (`ai-ux-review`); designing an actual game (`team-composer` @game_designer); implementing the mechanics (suggests, then hands off); dark-pattern or engagement-maximization hacks (refused).
gtm
Phased go-to-market for startup products. Builds a GTM playbook from upstream artifacts (validation-canvas, pitch-deck, brand-workshop), produces multi-channel content, schedules cadenced tasks, enforces compliance (CAN-SPAM/GDPR/FTC), emits handoff events. Trust ramp P1→P2→P3 (read-only → scheduled → autonomous). Project-local `.gtm/`. Kill switch via HALT file. Triggers on "set up GTM", "create a launch plan", "build a marketing playbook", "kick off go-to-market", "schedule social posts", "post-launch marketing", "growth ops", "draft launch content", or having pipeline output and asking "what's next for getting users". Single-channel asks for a startup product also trigger. NOT for brand identity (`brand-workshop`), pitch decks (`pitch-deck`), lean canvas (`validation-canvas`), adversarial review (`startup-grill`), one-shot non-startup content (`marketing:content-creation`), CRM/sales. NOT Google Tag Manager (different "GTM"). NOT the idiom "going to market with findings".
handshake
A brief, opt-in calibration ritual that runs before the real work. Shows the user what the agent already knows about them, then asks ≤4 high-leverage questions about how they want to collaborate (core mode), and optionally ≤6 scoped questions about the current project (overlay mode). Writes to the existing two-tier memory store using standard `user`-type and `project`-type entries — never a parallel store. Privacy-conscious: a hard never-ask list, every question carries a stated behavioral payoff, and the user can skip any item. Slash-command-only at v1. Use when the user invokes `/handshake`, or asks to "calibrate how we work", "tune in to me", "set a working agreement", "share my preferences", or when an upstream skill suggests calibration because user memory is sparse. NOT for codebase orientation, performance-review calibration ("set Q3 goals"), or content gathering (resumes, bios, requirements, CV bullets) — those are different jobs handled by other skills.
i18n
Handle internationalization (i18n) translation files and produce culturally authentic translations. Use when: (1) editing, creating, or modifying translation files (JSON, YAML, TypeScript, JS) — especially large ones that risk exceeding token limits; (2) translating UI strings into any supported language/dialect; (3) reviewing translations for cultural accuracy and naturalness; (4) syncing translation keys across multiple locale files; (5) adding new languages or regional variants to an existing i18n setup. Triggers on "translate", "i18n", "localization", "translation file", "locale", "add language", "translation keys", or any specific locale code (e.g., th, ja, ko, zh-CN, de, fr, es, it, zh-TW, zh-HK, th-bupphe). Does NOT trigger on defining or glossing a single word/phrase inside a sentence ("what does X mean here", "define X in context") — that is the `define` skill. This skill is for i18n files and shipping localized strings, not inline word-meaning lookups.
pitch-deck
Constructs an investor-ready pitch deck as a single self-contained HTML file (Reveal.js, print-to-PDF, zero network dependencies). Interviews the founder for real content — problem, solution, market, product, business model, traction, team, competition, ask, use of funds — and enforces one-claim-per-slide plus required-slot gating. Gated on `riskiest-assumption-test`: refuses to ship a clean deck unless a RAT plan exists with populated `## Results` for the top-3 hypotheses (override ships a `[PRE-VALIDATION DRAFT]` watermark). Use when the user asks to "build a pitch deck", "create a fundraising deck", "make an investor deck", "Series A deck", "seed deck", "demo day deck", or uploads a validation-canvas / RAT plan / brand kit and asks for the pitch. Also trigger when `brand-workshop`, `validation-canvas`, and/or `riskiest-assumption-test` output is ready, or when `@startup_strategist` + `@vc_partner` are the right lenses. Use it even for a single slide.
pixel-art
Generate single-subject pixel art (scenes, characters, buildings, title cards) from a short brief, with a built-in style system (palette, density, lighting, fonts). Modes `hi-fi` (painterly high-density, default) and `lo-fi` (scanlined banner). Generative: produces a raster, runs a deterministic median-cut quantize to true grid-aligned pixel art, and can serialize one sprite to SVG. Repo-agnostic — returns artifacts, writes nothing; with no generator reachable it returns a model-agnostic prompt brief. Triggers on "pixel art", "hi-fi/lo-fi pixel", "VT323 title card", "generated pixel art as SVG" (single sprite), or a hi-density reference plus "this style". Does NOT trigger on — and declines — composed/branded SVG-asset jobs (banner, icon set, multi-asset set, needing exact text, a fixed viewBox, brand tokens, or an editable vector source it doesn't author). Not for vector illustration, logos (`brand-workshop`), photo-real, p5.js (`algorithmic-art`), or charts. Never names living artists.
riskiest-assumption-test
Walks the founder through assumption dump, risk × impact ranking, falsifiable hypothesis rewriting, and test-method selection — produces a 1-page Assumption Test Plan with the top-3 ranked hypotheses, success/kill criteria, and an interactive HTML risk × impact matrix. Use after `validation-canvas` and BEFORE `pitch-deck` — pitching on untested assumptions is sales theater. Triggers on "test my assumptions", "riskiest assumption", "RAT", "what should I validate first", "assumption mapping", "experiment design", "how do I de-risk this", "Wizard of Oz test", "fake door test", "concierge MVP", "smoke test", "5-interview rule", or when `validation-canvas.md` has populated Stress Tests and the user asks what's next. Job — "what have we proven?". Upstream half of validation closure; the downstream half loops results back into the canvas. NOT a discussion of testing in general (use `team-composer`); NOT a generic experiment platform. Produces a plan, not the experiments.
screenwright
Paints one self-contained HTML surface (page, mobile screen, or component) to a brand spec, then SEES it: renders via the Playwright MCP, screenshots, runs an axe-core accessibility audit, critiques fidelity, and fixes in a bounded loop until it passes — then hands the verified HTML back for conversion to a real stack. Triggers when the user or an agent mid-build wants a UI surface built or fixed with the render actually checked: "build this screen", "paint a component", "match this mockup", "fix this UI", "make this accessible", "the layout looks off", "screenwright this". Resolves brand from a DESIGN.md, or bootstraps one from the repo / a template / a described vibe. Does NOT trigger for: multi-page flows, routing, or full app builds; framework/stack conversion (it outputs portable HTML for the host to port); pure copywriting (use ghostwriter); brand-identity or logo work (use brand-workshop); pixel-art or illustration (use pixel-art); backend, data, or non-UI code.
skill-evaluator
Audit an existing SKILL.md for rule adherence — does the text actually land when Claude runs it? Use when the user wants a behavioral review of a shipped skill. Outputs: failure classification by fix layer (skill text / rubric / brief / fixture) and targeted rule-text diffs. Trigger ON: "audit this skill", "stress-test my skill", "does this skill actually work", "find gaps in this skill", "what's broken in this skill", "validate rule adherence", "review this skill end-to-end", or uploading a SKILL.md for behavior review. Do NOT trigger on: "build a skill", "create a skill from scratch", "benchmark this skill", "evaluate skill quality", "compare versions", "optimize trigger phrases", or "measure variance" — those are all `skill-creator`. If the request mixes both, start with `skill-creator` and chain here. Hard boundary: `skill-creator` builds, benchmarks, measures variance, and optimizes triggering; this skill does NONE of those — it asks "does the text land?".
startup-grill
Adversarially grill a startup IDEA or PITCH with a 5-seat panel of domain-aware probers (plus up to 3 domain specialists from team-composer), and ship a structured kill report ranked on two axes: severity (lethal vs material) and fixability (fixable vs unfixable). Use when the user says "grill my startup", "stress-test my pitch", "kill my idea", "pre-mortem my startup", "is this fundable", or uploads a one-pager / canvas / RAT plan / pitch deck for adversarial review. Use THIS, not `team-composer`, whenever the user wants an adversarial review with a verdict on a belief artifact, even if they say "review". CARVE-OUT: if the input is a BUILT product (a CODEBASE or live URL) that is `startup-audit`, even though they say "grill" or "kill". This skill reads BELIEF artifacts only (Lean Canvas, pitch deck, one-pager), never a codebase. Do NOT use for collaborative brainstorming, canvas construction, assumption-test design, or pitch-deck building (those route to the named constructive skills).
startup-launch-kit
Opt-in umbrella orchestrator that sequences the full five-step startup pipeline (`brand-workshop` → `validation-canvas` → `riskiest-assumption-test` → `pitch-deck` → `startup-grill`) with shared state via `kit-manifest.json`. Every pipeline skill stays independently invocable; this is convenience, not replacement: it calls each in turn, never bypasses their gates, surfaces every prompt to the founder, and records gate overrides. Use when the user asks to "build my whole startup kit", "do the full pipeline", "launch kit", "full validation pipeline", or asks Claude to coordinate the five startup skills together. Also triggers for already-built projects ("run the whole pipeline on my existing repo", "infer the kit from the code"), reusing `startup-audit`'s inference to pre-fill what code reveals. (For a fast verdict rather than the full kit, that's `startup-audit`.) Also via `/startup-launch-kit`. Does NOT trigger for single-step requests ("just the canvas", "build my deck").
storm
Research a topic from multiple perspectives, grounded in retrieval, and return a cited briefing — an agent-native port of Stanford's STORM. Discovers (or accepts) diverse perspectives, runs grounded question-asking per perspective with retrieval at each step, and curates findings into a source-attributed synthesis. Every claim is cited or tagged [unverified]. Runs standalone (`/storm <topic>`) or as a hand-off target for `team-composer`, inheriting the panel's active roles as seed perspectives. Triggers on "research X", "what's known about Y", "survey the evidence on Z", "build a grounded briefing on…", or an upstream skill needing sourced external knowledge. Does NOT trigger on single-source-of-truth fact-checking or one authoritative report — that's `deep-research`; nor on opinion, critique, planning, or assembling a team — that's `team-composer`. Perspective-breadth, not source-breadth.
sub-agent-coordinator
Orchestrate multi-agent work through fan-out, pipeline, and specialist patterns. Use this skill when delegating focused tasks to parallel sub-agents, coordinating iterations, or architecting complex workflows. Load this whenever you need spawning signals, briefing templates, coordination patterns, or communication protocols.
team-composer
Assemble the right virtual team for brainstorming, planning, or building a project or feature — tech, healthcare, finance, climate, biotech, and beyond. Use this skill whenever the user asks to "brainstorm," "discuss," "plan," "review," or "build" a project or feature and doesn't specify exact roles. Also trigger when the user says things like "bring a team," "who should be involved," "let's workshop this," "assemble a team," or asks for multi-perspective feedback on a product decision. This skill analyzes the project context and selects the optimal set of role-based personas to contribute meaningfully. Even if the user only mentions one or two roles, use this skill to fill in the gaps — a well-composed team produces better results than ad-hoc role selection.
tech-stack-recommendations
Opinionated technology stack guidance for new projects and migrations. **You MUST use this skill whenever the user is starting a new project, picking a runtime (Bun/Deno/Node), choosing a frontend framework, database, auth provider, hosting platform, styling system, mobile stack, i18n library, or AI-agent framework that the project doesn't yet have, setting up a monorepo, or migrating between stacks.** Do not assume a stack decision is too obvious to need this skill — consult it before recommending any specific technology. Does NOT trigger on: debugging existing code, maintaining or extending the current stack's existing components, library-version or package-manager migrations within the same stack, configuring tools already installed, or 'how do I do X in [framework I already use]' questions.
validation-canvas
Elicits founder beliefs block-by-block to produce a Lean Canvas (Maurya) plus Value Proposition Canvas (Osterwalder) combined artifact (`validation-canvas.md`, editable, plus `validation-canvas.html`, self-contained visual). Job — "what do we believe?". Adapts to founder experience via a 3-question intake mapping to Guided / Focused / Compressed-with-Challenge modes. Use when the user asks to "build a validation canvas", "lean canvas for my startup", "VPC", "value proposition canvas", "map out my business model", "fill in the lean blocks", or uploads a product idea and asks how to articulate the model. Also trigger when `brand-workshop` output is ready, or when `@startup_strategist` is the right lens. Use it even for one block — the others stress-test it. Replaces the prior `business-model-canvas` (BMC) skill; the Lean Canvas is the right altitude for an idea-stage founder. For nine-block BMC (board packets, regulated pitches), use `team-composer` with `@startup_strategist`.
wear-the-hat
Pick ONE role from `team-composer`'s catalog (or a user-specified role) and perform a task in that role's voice — solo, no panel discussion. Use when the user wants a single specific lens applied to a task without convening a multi-role discussion. Triggers on explicit `@role` tags ("audit middleware/ as @security_specialist"), role-embodiment phrases ("act as a security specialist", "wear the hat of the architect", "have the copywriter rewrite this", "embody"), or any prompt naming a role-shaped lens. Does NOT trigger on generic tasks that don't ask for a role lens — those route to cerby, team-composer, or sub-agent-coordinator. When a task spans multiple roles it stops, surfaces candidate single roles, and offers a hand-off to `team-composer`, never silently picking one role for a multi-role task. Reuses `team-composer/references/role-personas.md` as the canonical role catalog. For panel discussion use `team-composer`; for parallel fan-out use `sub-agent-coordinator`.
whoami
A short, conversational interview that builds a portable profile of the user — who they are and how they want to be collaborated with — so the agent can tailor its responses. Triggers on `/whoami`, "get to know me", "set up my profile", "personalize how you work with me", "calibrate how you respond", the user mentioning they are new to AI or switching agent vendors, or a request for their collaboration "character sheet" or player profile. Produces six collaboration dials, an RPG-style class + subclass with a character portrait, a memory profile, and a self-contained HTML character sheet. When a profile exists, bare `/whoami` reviews it and offers to correct or re-run; `/whoami rerun` restarts the interview. Does NOT trigger for project- or task-specific work calibration (use `handshake`), routine memory edits or CRUD (use memory-management), or brand-voice work. `whoami` is person-level and broad; `handshake` is project-level and feeds from it.
kerby
Load, install, reload, check status of, uninstall, prepare, or audit an existing repo for the kerby guardrails system. Invoke ONLY when the user explicitly mentions "kerby", "/kerby", or asks to load/unload/install/uninstall/check/ list-or-create-rulebooks/prepare (onboard an existing repo into) or audit-a-repo-against the kerby guardrails. Do NOT invoke on general coding tasks (fixing bugs, implementing features, refactoring) — kerby is a meta-system that itself governs how those tasks are done; `audit` checks a repo's conformance to the rules, it is NOT a general bug/security review. Engine sub-commands via the args parameter: `load` (default), `unload`, `reload`, `status`, `install`, `uninstall`, `rulebooks [list]|create`, `commands`, `check-updates`; loaded rulebooks add their own commands (e.g. the bundled `swe` rulebook provides `prepare` and `audit`).
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.