harness

Design and build parallax and scroll-driven pages that read as depth and craft rather than gimmick. Covers layered depth systems (speed ratios, layer budgets), the pure-CSS perspective technique, native scroll-driven animations (animation-timeline: scroll/view), GSAP ScrollTrigger and Lenis, sticky scrollytelling, compositor-only performance rules, prefers-reduced-motion fallbacks, mobile strategy, and copy rules for scroll pages. Use whenever building, redesigning, or polishing a landing page, marketing page, homepage, hero section, product or launch page, campaign page, or portfolio, even if the user did not say the word parallax. Triggers on: landing page, marketing page, homepage, hero section, parallax, scroll animation, scroll effects, scrollytelling, product launch page, promo page.

Ownership mindset reset. Snaps LLM out of shortcut-seeking, lazy, sloppy behavior. Reframes work as 'this is my product' with quality-first incentives.

Create short emotional or explanatory films entirely in code, with no stock assets and no video editor. Renders frames procedurally with Pillow and numpy, synthesizes the soundtrack from scratch, and muxes both through ffmpeg in one pass. Covers two visual languages (dark glow-scene films and typed-letter paper films), self-drawing ink illustrations, typewriter mechanics with sound, and a verification loop. Triggers on requests to make a video, animation, film, motion piece, video letter, animated letter, or "what it feels like" video.

Rewrite text so it reads as human-written by removing AI-writing tells. Use whenever producing or editing human-facing prose: READMEs, docs, notes, blog posts, essays, marketing or landing copy, emails, HTML body content, release notes, PR descriptions, or any client/user-facing writing. Detects and fixes inflated significance, promotional language, superficial -ing analyses, vague attributions, em dash overuse, rule of three, AI vocabulary, copula avoidance, passive voice, negative parallelisms, aphorism formulas, manufactured staccato, and filler. Based on Wikipedia's Signs of AI writing. Triggers on: humanize this, de-slop, remove AI tells, make this sound human, edit this writing, polish this copy, rewrite this draft.

Structure questions as single/multi-choice prompts (one at a time, title + description) and strip AI-tell rhetorical tics (load-bearing, honest framing, is what [verb]s, fundamentally, marketing-LLM vocabulary) from both chat and any prose written to files.

Plain reference facts about the user's machines and self-hosted services, and which to prefer for a given task. No secrets, hostnames, or IPs live here — load for context, then ask the user for current endpoints or credentials. Use whenever a task involves the user's own infrastructure: deploying or self-hosting something, using a local LLM or GPU, or asking what machines, servers, endpoints, or services they have. Triggers on: "my setup", "my infrastructure", "homelab", "my server", "use my GPU", "local model", "where should I deploy", "what services do I have".

LLM-maintained knowledge wiki (Karpathy pattern). Ingest sources, maintain structured Obsidian pages, query with citations, lint for quality. Triggers on: "wiki ingest", "wiki query", "wiki lint", "wiki init", "add to wiki", "what does the wiki say about", "check wiki health".

Turn an article, post, or single concept into 4-8 deadpan hand-drawn 16:9 explanatory illustrations starring a recurring crow that performs the idea. White background, sparse black line art, a few short red/orange/blue handwritten labels. One image explains exactly one structure, action, or metaphor. Use when the user wants article illustrations, explanatory figures, a hand-drawn diagram with a character, or a shot list for illustrating writing. Triggers on: "illustrate this", "make illustrations", "article illustration", "explanatory figure", "hand-drawn diagram", "crow illustration", "shot list for illustrations".

Iterate a single atomic task to done through a worker-critic loop: the agent does the work, a panel of fresh independent reviewers and a set of hard verifiable gates judge it against criteria you set, the agent revises, and the loop repeats until every hard gate passes and the reviewers converge green, or it honestly surfaces a structural ceiling and stops. Works for code changes, refactors, research answers, configs, designs, and any short-form artifact (README, plan, abstract, commit message, copy, prompt). Use when the user wants to converge or iterate a task to a defined bar of acceptance. Triggers on: "converge", "iterate to green", "iterate until done", "review until green", "polish this", "tighten this", "make this airtight", "iterate this with reviewers", "run this through converge".

Create polished GIF recordings of terminal commands for READMEs and documentation. Records real command output with asciinema, edits the cast file to compress idle time and stagger output for visual clarity, then converts to GIF with agg. Handles the full workflow: recording, timing adjustments, color-coded output, and final rendering. Triggers on: "terminal gif", "record terminal", "create gif", "demo gif", "README gif", "asciinema", "record command output".