hellokianben-collab
UserAI context pack for Vishal Agarwal: profile, work history, 27 scored skills, and the expensive lessons. Loadable by any AI assistant.
Categories
Indexed Skills (27)
ship-modern-3d-site
Build and deploy a modern 3D/motion website end-to-end — Next.js (App Router) + TypeScript + Tailwind v4 + react-three-fiber + drei + motion (Framer) + Lenis, shipped to Vercel. Use when the user wants a "modern website", "3D hero", "scroll animations / parallax reveals", a "portfolio / landing page with motion", a three.js/R3F site, or asks to deploy such a site to Vercel. Covers project setup, the component/animation patterns, local verification, and the Vercel deploy gotchas.
autonomous-channel-plan
Honest architecture and feasibility read for an agent that creates and runs a content channel on its own — picks a niche, produces and schedules daily video, and grows an audience. Use when asked to build a fully autonomous YouTube/TikTok/Shorts channel agent, "grow a channel to 1M subscribers", an AI faceless-channel pipeline, or daily automated content generation. Carries the pipeline that is actually buildable, the platform policy constraints that decide whether it survives, what "free tools only" really costs, and why the subscriber target is a goal rather than a plan. Use it to scope the ask truthfully before building.
bd-import-export-domain
Bangladesh import/export domain knowledge for building tools, content or advice for traders — the NBR compounding duty cascade, HS codes and the Operative Tariff, landed-cost calculation, Letter of Credit / UCP 600 discrepancies that block payment, shared-container imports, and the market realities that kill most "AI for X" pitches here. Use when the task involves Bangladeshi importers or exporters, customs duty, HS codes, landed cost, CIF/assessable value, L/C or bank document presentation, garment export, freight or clearing, shared imports, or writing content aimed at BD traders. Not a substitute for a licensed clearing agent — outputs are indicative, never an assessment.
bkash-payment-integration
Take money in Bangladesh — the manual bKash/Nagad TrxID flow for small sellers, and the full bKash PGW tokenized-checkout integration for merchants. Use when the user mentions bKash, Nagad, "mobile banking", a merchant account, taking payments in Bangladesh, a book/product checkout for a BD audience, delivery charge by zone, or a payment that "went through but the order didn't update". Carries the correct flow, the non-negotiable money-safety rules (server-side pricing, idempotent callback, never mark failed on a callback error), the sandbox-credential dead end, and the go-live review gate. Not for Stripe/PayPal/international gateways.
brand-media-pipeline
Produce and place brand imagery and video for a personal or business site without burning paid generation credits or shipping assets that break — background cutouts done locally and free, when to generate versus edit versus self-host, thumbnail and reel handling, and the text-rendering limits. Use when adding photos or video to a site, making a portrait cutout or transparent background, generating hero/venture imagery, producing YouTube thumbnails, embedding Facebook or Instagram reels, or choosing between Higgsfield/HeyGen/local tooling. Carries the credit-saving default, the signed-URL trap, and the Bengali text-shaping limit.
claude-code-setup-ops
Install, scope, verify and debug Claude Code capabilities — MCP servers, skills, plugins and marketplaces — with a hard bias toward USER (global) scope so they work in every folder. Use whenever someone says "install this MCP", "add this skill", "install this plugin", pastes a `claude mcp add …` or `npx skills add …` command or a GitHub repo URL to install, or reports that something they installed "isn't showing in the / commands", "isn't listed in /mcp", "works in one project but not another", or "needs authentication". Also use before starting any task that a missing capability would cap — install the capability instead of shipping a weaker result. Windows-first (PowerShell 5.1, .cmd shims), but the scoping rules are cross-platform.
content-integrity-guard
Prevent invented facts, fabricated testimonials, edited quotes and regressed corrections from reaching anything published under a real person's name. Use whenever writing or editing copy for a real person's or business's site, adding testimonials/reviews/metrics/credentials, filling a placeholder, importing a design that contains lorem or sample content, or re-deploying a page whose facts were previously corrected. Also use as a pre-deploy gate — it carries a grep-based regression check. Short, absolute, and non-negotiable.
email-capture-and-delivery
Capture email addresses that are real, store them deduplicated, and actually deliver mail. Use when building a waitlist, signup, contact or booking form; when transactional email "isn't arriving"; when signups produce duplicates or 500s; when a provider returns 403 on send; or when adding Resend/SES/Mailtrap to a project. Carries the verification ladder, Gmail dot/plus canonicalization and the unique-index rule it implies, the mandatory sender-domain configuration, the 403 diagnosis, and the behaviour contract that stops a form faking success.
garmentmind
Project context, architecture, and working guide for GarmentMind AI — an L/C & export-document discrepancy checker for Bangladesh garment exporters (a FastAPI + React + Postgres app, deployed on Vercel, in Desktop/Garment). It reads a factory's trade documents (Letter of Credit, commercial invoice, packing list, bill of lading, purchase order) and flags every mismatch that would get their L/C payment rejected by the bank. Load this whenever the user works on GarmentMind, "the garment app", "the document checker", the discrepancy engine, garmentmind.vercel.app, extraction providers, the wedge/pilot strategy, or anything in the Desktop/Garment folder — it carries the mission, architecture, file map, the ~18 discrepancy rules, deploy setup + hard constraints, data model, conventions, the honest capability/gap list, the roadmap to a reliable v1.0, and the continuity state so a fresh conversation resumes without re-deriving anything. Use it even for vague asks like "keep building the garment app", "make extraction
hero-canvas-fx
Reusable patterns for leveling up a plain HTML/CSS/JS marketing site (Express + Vercel, glass/dark theme, no framework): (1) animated canvas hero backgrounds — a 2D "trade-lanes" flow and a pseudo-3D rotating dot-globe with shipping arcs, both DPR-aware, reduced-motion-safe, paused off-screen; (2) splitting a single-scroll page into standalone destination pages that reuse the shared theme with self-contained per-page JS; (3) verifying/iterating canvas visuals in a headless in-app browser where requestAnimationFrame is throttled. Use when adding a moving hero background, a globe/network/particle backdrop, a "make the hero more eye-catching" ask, splitting a section (book/board/etc.) onto its own page, or when a canvas draws blank in the preview pane. Extracted from the KiAnben site (companion to the `kianben-web` skill); applies to any file-based Node/HTML site.
kianben-web
Project context + working guide for the KiAnben website (kianben.com) — a Bangladeshi shared-import community site: an HTML/CSS/JS frontend served by an Express server, data in Neon Postgres (prod) / a JSON file (local dev), deployed on Vercel serverless. Load this whenever the user works on KiAnben, kianben.com, "the import site", the book shop / book orders, bKash/Nagad payments, the member sign-in / admin panel, or anything in the Desktop/claude/kianben.web folder — it carries the architecture, file map, deploy/env setup, data model, conventions, and open TODOs so a fresh conversation can continue without re-deriving everything. Use it even for vague asks like "fix the site", "the orders aren't showing", "change the book price", "it looks broken on my phone", "the animations load late", or "deploy my changes" — it also carries the responsive/nav/scroll-reveal and gold-vs-cyan rules that must not be regressed. For the generic Vercel-serverless mechanics it relies on, see the companion skill `vercel-node-ser
no-code-owner-handoff
Hand a working system to a non-technical owner so they can actually run it — click-by-click setup with expected results, a secrets checklist they only paste, a runbook for the three things that break, and an honest capability/limit list. Use when the person you're building for does not write code, when writing SETUP.md / README for an owner rather than a developer, when a task needs the owner to click a consent screen or add a DNS record or paste an API key, or when a build is finished and needs to be handed over. Also use when deciding how much to automate versus explain.
owner-admin-security
Design owner-only access for a small single-operator product — device-bound admin sessions, a master recovery key, an inactivity-based failover hierarchy so a lost laptop doesn't lock the owner out forever, an audit trail, and safe key rotation. Use when building an admin panel's access control, when the user asks "what's the passkey", asks to restrict admin to one device, worries about losing access, wants a recovery path, or wants to see who logged in. Covers the fail-open trap, why a login must never count as presence, and what to do instead of rolling your own crypto.
personal-mentor-website
Build a personal brand / mentorship / coach / consultant website with owner-self-managed content (achievements, goals with task-progress, videos) via a password admin panel, a dark+gold animated hero with a 3D-model slot, social links, and optional booking + Bkash payments. Two stacks — simple (HTML/CSS/JS + Node/Express + JSON file) or full (Next.js + Prisma/Postgres + Cloudinary). Use when building or extending a personal/portfolio/mentorship site with an owner-editable content panel, visibility eye-toggles, calendar booking, or Bkash payment flow.
prove-before-claiming
Verify that a change actually works before reporting it done — offline test harnesses with real Postgres in WASM, live smoke tests against production, regression greps, and the rule that a tool which lies must never be the evidence. Use before saying "done", "deployed", "fixed" or "verified"; when a change touches money, auth, email or data; when a browser preview disagrees with reality; or when a multi-agent verification pass returns zero findings. Carries the harness patterns, the honest-reporting format that always names what was NOT verified, and the specific tools in this environment that report false results.
public-data-without-api-keys
Pull real, accurate data from public web platforms without an API key or a login — YouTube video IDs and titles, Facebook post metadata via the crawler user-agent, and the general method of reading a page's embedded JSON state. Use when you need a creator's real video list, real titles, real thumbnails or post metadata for a website; when an API needs a key/quota/OAuth you don't have; when scraping the DOM returns nothing because content is lazy-loaded; or when deciding which thumbnail URLs are safe to hardcode. Carries which URLs are permanent versus signed-and-expiring, and the limits (Instagram, TikTok) you should report rather than fake.
responsive-reveal-audit
Find and fix layout, responsive and scroll-animation bugs on an existing website by MEASURING it in a browser instead of eyeballing it. Use this whenever someone says a site "isn't visible properly on mobile", "looks broken on my phone", "text/buttons are cut off", "the menu doesn't open on mobile", "buttons break in the middle", "fonts should scale with the screen", "the animations load late", "content is blank until I scroll back up", "it lags when I scroll fast", or asks to make an existing page responsive, audit it across device sizes, or fix its scroll transitions. Also use it before redesigning a live page, since it tells you which problems are real and which are imagined. Applies to any plain HTML/CSS/JS or framework-rendered site, not a particular project. Reach for this even when the user only describes a vague visual complaint — the probes turn "looks off" into exact pixel numbers, which is usually the whole difficulty.
ruthless-venture-review
Evaluate a business idea, product plan, or economic proposal adversarially and score it, instead of encouraging it. Use when the user asks you to "evaluate this like a ruthless investor", play a skeptical expert, judge or rate an idea, give judges' scores, stress-test a proposal, or asks "is this worth building". Also use unprompted before committing significant build time to a new product direction. Produces a scored verdict with named kill-shots, the market realities that invalidate common pitches, and a concrete narrower wedge — not a list of pros and cons.
self-improving-agent-loop
Architect a long-running agent system that measurably gets better instead of accumulating confident guesses — specialist micro-agents woken by events, a shared playbook, observed (not self-reported) outcomes as the training signal, and a critic that deletes unsupported claims. Use when building any autonomous or scheduled agent, a multi-agent system, a monitoring/growth agent, a "learns from feedback" feature, or when an existing agent's advice has drifted into plausible nonsense. Also use when deciding between one generalist agent and several specialists. Carries the attribution mechanism, the honesty mechanism, and the control-plane pattern that avoids exposing a laptop to the internet.
session-to-project-skill
Compress a long build session into a durable project skill (plus a short memory note) so the next conversation resumes without re-deriving anything. Use when the user says "compress this conversation into a skill", "save the context", "so you have the context in a new conversation", "update the skill", "remember this", or when a session is about to end / hit a usage limit with hard-won state that only exists in the transcript. Also use proactively after any multi-hour build, any expensive debugging session, or any owner correction that must never regress. Produces a skill that a cold session can act from — not a summary of what happened.
social-media-agent
Project context and working guide for Vishal's AI social media manager — a local Python agent in Desktop/claude/socialagent that monitors YouTube, Facebook and Instagram, diagnoses why performance dropped, designs replacement thumbnails, rewrites titles, learns from what he approves, and publishes a password-gated dashboard to socialagent-dashboard.vercel.app. Load this whenever the user mentions the social agent, "my channel", "the YouTube agent", thumbnails or CTR dropping, Facebook/Instagram engagement or reach, the channels/monitoring list, the micro-agents (design, copy, analyst, meta, strategy, critic), the manager, playbook.md or learned.md, "Fix it now"/"Don't need it", autopilot, the Telegram /fix and /skip commands, run_daily/run_sync/run_weekly, the Vercel dashboard or its password, or anything inside the socialagent folder. Use it even for vague asks like "check my channel", "why are my views down", "the agent isn't running", "add a channel to monitor", "redeploy the dashboard", or "keep building
vercel-hobby-limits
Survive Vercel's free-tier walls without upgrading — the 12-serverless-function cap and the router consolidation that answers it, the 4.5 MB response cap, filesystem-resolved-before-rewrites, query-string merging on rewrites, immutable asset caching, and env vars that read back empty. Use when a deploy fails with a function-count error, when a rewrite "does nothing", when a form silently falls through to the wrong handler, when a large file returns FUNCTION_PAYLOAD_TOO_LARGE, when assets serve stale after a deploy, or before adding any new endpoint to a Hobby-plan project. Also use when planning an architecture that will run on Hobby.
vercel-node-serverless
Deploy an existing file-based Node/Express app (HTML/CSS/JS frontend served by an Express server, data in a JSON file) to Vercel serverless — and fix the gotchas that silently break it there. Use this whenever someone wants to put a Node/Express, HTML+API, or "server.js" style site live on Vercel, connect a custom domain, or is confused why a site that "works locally" is broken in production: unstyled pages / 404 assets, data that disappears after a while (logins fail, orders vanish, admin edits reset), or emails that never arrive. Also use when adding persistent storage (Postgres/Neon) to a JSON-file app, or a manual bKash/Nagad (Bangladesh mobile-money) payment + order flow. Trigger even if the user only says "deploy my site to Vercel", "my Vercel site looks broken", "make logins persist", or "sign-in stopped working in production" — the serverless traps below apply to almost every file-based Node app.
vishal-agarwal-site
Project context + working guide for Vishal Agarwal's personal site (iamvishalagarwal.com) — a multi-business portfolio (import/export, wig manufacturing, Confidence Cement dealership, business education) rebuilt from a Claude Design export into plain HTML/CSS/JS + Vercel serverless functions + Postgres. Load this whenever the user works on iamvishalagarwal.com, "my site", "the portfolio", the book "Prothom Container", the ebook/PDF reader and its reading analytics, consultation bookings and the admin panel for scheduling calls (Zoom/Google Meet), bKash payments, the waitlist/contact forms, the video/media section, testimonials, or anything in the Desktop/claude/vishal-site folder. Also covers the tools mounted on the domain: /landed (the Bangladesh landed-cost and NBR duty-cascade calculator, its 7,465-code tariff dataset and the sync-landed.js path rewrite), /community (member accounts, scrypt passwords, saved calculations), /garmentmind (the L/C discrepancy landing page), and the admin security layer (per-d
windows-agent-automation
Windows-specific traps when an agent or script drives CLIs, schedules jobs, or generates files on Windows — .cmd shims truncating multi-line input, PowerShell 5.1 missing operators, encoding defaults that corrupt output, Task Scheduler instead of cron, and path/credential locations. Use when writing any Python/Node script that shells out on Windows, scheduling a recurring job, debugging "it works when I type it but not from the script", seeing UnicodeEncodeError or garbled characters, or when a subprocess cannot find a command that is clearly on PATH. Every trap here has already cost real hours on this machine.
work-within-usage-limits
Plan and checkpoint long builds so a usage/context limit never destroys work in flight. Use when the user mentions a usage limit, token limit, "finish it before the limit hits", asks you to monitor remaining budget, asks to split a project into portions sized to the remaining budget, or says "Continue from where you left off". Also use unprompted at the start of any build expected to run more than an hour, or any multi-agent workflow — those are the two things that get cut off mid-flight and lose everything. Covers portioning, durable checkpoints, resumable state, and recovering findings from agents that died before reporting.
zero-js-admin-panel
Build an owner-facing dashboard or admin panel that works with JavaScript fully disabled — server-rendered HTML, real forms, plain links. Use when a login or button "does nothing and shows no error", when an admin panel works in curl but not in the owner's browser, when the user is on a browser with extensions/script blockers, or when building any internal panel where reliability beats interactivity. Also use before writing a client-rendered admin, as the default choice. Carries the diagnostic that identifies a script blocker, the patterns for actions without JS, and the escaping rule that replaces `textContent`.
Bio shown is the top-scored skill's repo description as a fallback — real GitHub bios land in a future update.