dufflebag

Optimize Chrome Web Store listing SEO (name, short description/summary, Overview) and marketing-site GEO using official Chrome/Google guidance plus free Google APIs. Use when the user asks for CWS SEO, Chrome Web Store listing optimization, store title/summary/description copy, keyword research for extensions, Featured eligibility copy, or GEO/AEO for an extension landing page.

Direct a cinematic, viral-ready trailer for any project — reads the repo's own docs to derive the story, consults ChatGPT (GPT-5.5 Thinking) over ai-browser-bridge to write the transcript + storyboard, batch-generates the keyframes as ChatGPT images, animates them with Higgsfield or Flow/Veo, produces voiceover + music (ElevenLabs → Higgsfield → local synth), and assembles a 9:16 master + 16:9/1:1/4:5 platform cuts with ffmpeg — behind two planpage approval gates. Use when the user wants a trailer, promo, launch/teaser/hype video, commercial, sizzle reel, or cinematic short for a project to post on LinkedIn/YouTube/Instagram/TikTok, or says "make a trailer", "make a promo", "cinematic video", or "make it go viral". For a portfolio blog cover image use write-a-post; for store/marketing copy use cws-listing-seo.

Use when refreshing, rewriting, syncing, or creating repository agent instruction files such as AGENTS.md, CLAUDE.md, GEMINI.md, Cursor rules, Kiro steering, Roo rules, Codex instructions, or root project context docs.

Audits a website or web app against web-platform best practices and ships the fixes, across seven dimensions — semantic HTML + ARIA landmarks, accessibility, images/fonts, performance, security headers (CSP/HSTS), SEO metadata, and machine-readability for AI agents (llms.txt, schema.org JSON-LD, crawler access, Google's agent-ready rules). Runs a zero-dep static scanner over a project, guides the agent's semantic pass, and provides copy-paste fixes (llms.txt + security-header templates) plus authoritative sources (MDN, web.dev, W3C APG, llmstxt.org). Use when the user wants to check or improve a site's best practices, accessibility, security headers, CSP, semantic HTML, SEO/meta tags, structured data, schema.org, or make a site agent-ready or AI-crawler-friendly. For the deep Core Web Vitals CI gate (Lighthouse CI, CrUX), defer to web-perf-ci.

Reviews code and repo structure for over-engineering, then removes the excess so the code does exactly what it needs and no more. The over-engineering companion to deslop (deslop makes code readable; deslop-v2 makes it lean). Use when the user says "deslop-v2", "over-engineered", "too much abstraction", "over-abstracted", "simplify this", "why is this so complicated", "flatten this", "too many wrappers/layers/folders", "this feels like too much", or asks to cut needless indirection, `??` fallbacks, nested ternaries, pass-through wrappers, single-use helpers, grab-bag returns, or deep folder/package nesting.

Grill the user on how a NEW/greenfield project is built — code style, structure docs, and CLI — then render an HTML plan and, on approval, write CODE-STYLE.md + a formatter config and refresh the AGENTS.md digest. Ensures PROJECT.md/CONTEXT.md/LANGUAGE.md exist (create if missing, validate if present) and grills real code idioms + formatting, not just architecture. Use when setting up or reorganizing a new project, when there is little/no code, or when defining coding style, structure, or CLI conventions from scratch. For an existing codebase, use grill-me-code-style-with-docs instead.

Coach the user WHILE building or fixing code in a repo that has a code-style.rules.json — at each real style/structure decision, explain the concept concisely (KISS) in terms they already know, show two concrete variants, make them confirm the direction before writing, then teach the mechanism and link the exemplar it mirrors. High-friction on judgment calls; silent on anything Biome already enforces. Grounds "where the business logic goes" in PROJECT.md / CONTEXT.md. Use when writing a feature/fix and the user wants to LEARN their own architecture as we go, not just delegate — or says "coach me", "teach me as we build", "grill me while we code". For reviewing a finished diff instead, use grill-me-code-style-review; to CREATE the ruleset, grill-me-code-style.

Review a large changeset (a branch, working tree, or PR — tens to hundreds of files) against a repo's code-style.rules.json so the user can trust it WITHOUT reading every diff, and LEARN the architecture from a short teaching report instead. Runs Biome (+ its grit plugins) to auto-fix the mechanical/structural channels across all files, then fans out sub-agents over the diff to check only the judgment-channel rules + the user's original intent, and returns a deviations-only report that opens with a layer/flow map of what changed and teaches the "why" behind each finding. Use when reviewing/auditing a big diff, a PR, or an AI-generated changeset, or the user says "review this", "can I trust this diff", "audit these changes". To coach while building, use grill-me-code-style-coach; to create the ruleset, grill-me-code-style.

Grill the user on how an EXISTING codebase is built — code style, structure docs, and CLI — using the real code as evidence, then render an HTML plan and, on approval, write/update CODE-STYLE.md + a formatter config and refresh the AGENTS.md digest. Ensures PROJECT.md/CONTEXT.md/LANGUAGE.md exist (create if missing, validate if present), fans out sub-agents for the most-repeated patterns, grills real code idioms + formatting with before/after, audits deps, and references official framework skills. Use when defining/updating style, structure, or CLI conventions for a repo with meaningful code. For a brand-new or empty project, use grill-me-code-style instead.

Grill the user on the TECHNOLOGY choices of a project — language, runtime, framework/meta-framework (e.g. React vs Next), and the key services/vendors — one decision at a time, so they can EXPLAIN why, not just accept it. For each: teach the tradeoff space in plain terms (honest cost vs gain, alternatives conceded), make them pick, then explain how that piece's architecture actually works. Writes each outcome to TEACH.md as a lean decision-record + a self-closing, officially-cited glossary with a code snippet per term — deduped against what's already there. Leaves LANGUAGE.md for shared domain vocabulary and hands deep multi-session concept mastery to the `teach` skill. Use when the user asks "why this stack/framework/language/service and not the other", wants to understand or be able to explain their own tech choices, or says "grill me on the stack", "why TS not Python", "why React not Next", "teach me why we chose X". For code-style conventions use grill-me-code-style.

Grilling session that challenges your plan against the existing domain model, sharpens terminology, and updates documentation (CONTEXT.md, ADRs) inline as decisions crystallise. Use when user wants to stress-test a plan against their project's language and documented decisions.

Create or edit a project's landing documentation set — README.md, AGENTS.md, CLAUDE.md, .github/copilot-instructions.md, and llms.txt — by first deciding which artifacts the repo actually needs, then grilling the user one question at a time with recommended defaults, building a compact README map, using official hyperlinks for named tools and technologies, and only then writing. Inspects the repo before asking anything discoverable. Use when the user wants to create, write, edit, rewrite, audit, polish, or "fix up" a README, AGENTS.md, CLAUDE.md, Copilot instructions, llms.txt, or any project landing/onboarding docs.

Wires automated website-performance gates into CI/CD and enforces Core Web Vitals budgets (LCP, INP, CLS). Sets up a Lighthouse CI budget check on every PR (lab), a Chrome UX Report (CrUX) real-user field check after deploy, and an optional web-vitals RUM snippet. Grills the repo to detect the stack and how the site runs (static build, dev server, or preview URL), then writes lighthouserc, the GitHub Actions workflows, and zero-dep CrUX + PSI checkers. Use when the user wants to add performance testing, Lighthouse, Lighthouse CI, PageSpeed Insights, Core Web Vitals, web-vitals, a performance budget, or a speed/perf gate to a website's CI/CD, or mentions LCP/INP/CLS regressions or a slow site.

Write a new blog post for Joseph Sabag's portfolio in his exact voice, scaffold it into clientV3/src/data/blog.ts via scripts/dev/new-post.mjs, and generate a matching cover image by driving a real ChatGPT browser conversation through ai-browser-bridge (attaching the likeness photo + an existing cover as references). Use when the user wants to write, draft, add, or publish a blog post / portfolio post, or says "write a post", "new blog post", "post in my voice", or hands over a title/theme + body for the blog.

Reviews code readability first, then applies approved cleanup that makes the full pipeline understandable in seconds. Use when the user says "deslop", "make this readable", "make this less AI", "second pass", "clean this up", "rename for clarity", "show before and after", or asks to improve code comprehension across React, TypeScript, backend, folders, imports, hooks, or functions.

Write and audit GitHub repository About metadata: one-line descriptions, homepage/demo links, and topics/tags grounded in official GitHub guidance. Use when the user asks for a GitHub repo description, repository topics, tags, About sidebar metadata, discoverability, or before/after wording for repo metadata.

Convert a PNG design into pixel-perfect code — SVG illustrations/logos, HTML/CSS UI, and animations — using a decompose → reuse-or-build → render → screenshot-diff → refine loop that measurably converges to a 1:1 match instead of eyeballing. Use when the user provides a PNG, screenshot, mockup, or inspiration image and wants it turned into SVG, HTML/CSS, a web component, or an animated illustration, or asks to match a design "pixel-perfect" / "1:1".

Render a skill's plan, review gate, or report as a beautiful, self-contained, INTERACTIVE HTML page — via the open-source planpage package (Preact components → static HTML + a local post-back server so the user can approve / adjust / flip decisions in the browser and the choice comes straight back to the agent). Use whenever a skill needs an approval gate, a decision review, or a shareable before/after report — author with the kit's components instead of hand-rolling HTML each time.

Drive the autonomous context loop for this session — arm it to auto-/compact and resume work hands-free, or pause/shut it down. Use when the user types /autorun (optionally with a number or `stop`/`exit`), says "autorun", "autopilot", "take it from here", "keep going hands-off", or asks to pause / stop / shut down / exit the auto-compact loop.

Drive the autonomous context loop for this session — arm it to auto-/compact and resume work hands-free, or pause/shut it down. Use when the user types /autorun (optionally with a number or `stop`/`exit`), says "autorun", "autopilot", "take it from here", "keep going hands-off", or asks to pause / stop / shut down / exit the auto-compact loop.