working-with-ai

When a task is handed off to you — especially unsupervised or "I give you stewardship over this" — carry it through to completion. The hard moment (long context, ballooning work, late hour) is the opportunity, not the cue to quit. Don't beg off — you have durable externalized memory; compact and continue. A handoff weighs more than a normal task. Load when accepting a delegation, working unsupervised, or tempted to stop before done.

Design and write antagonists with motivation that makes sense from the inside. Politics, ideology, ambition, wound — never 'evil for evil's sake.' Load when designing villains, faction antagonists, or morally complicated NPCs.

Quick intent articulation discipline before starting a new task. Names purpose, beneficiary, success criteria, and non-goals upfront so the agent has a target to optimize for when instructions run out. Klarna-failure prevention. Load before any non-trivial new task.

Automate browser interactions, test web pages and work with Playwright tests.

Open a location or moment by painting it through the senses so it becomes immersive and real. Load when narrating a new scene, room, or environment.

Shape a session or arc so each beat causes the next and the character returns changed. Load when planning a session arc, a campaign, or any multi-beat story.

Ground analytical judgment in the actual artifact before rendering it. Load before producing OR dispatching a review, critique, evaluation, assessment, or "is this right / this is the bug" verdict — anywhere a conclusion could be pattern-matched from a role instead of read from the thing. Named for D&C 9:8.

Apply Edward Tufte's principles to any data visualization, chart, dashboard, or infographic. Use when the user asks to produce, design, critique, or improve a chart, graph, dashboard, KPI tile, table-with-data, sparkline, small multiple, time series, distribution plot, choropleth, or any other visual that communicates quantitative or categorical information. Also triggers on words like "visualize", "viz", "data viz", "chart", "graph", "dashboard", "infographic", "Tufte". Outputs Tufte-compliant designs in either self-contained HTML/SVG or React (Recharts/D3). Defaults to Tufte rules; user can opt out per chart.

Structured fresh-eyes design review of any UI — screenshot-driven critique that produces ranked, concrete findings in a shared feel-word vocabulary (clean, sharp, cluttered, calm…). Use when the user asks to review/critique/improve a UI, says something "looks off / feels bolted on / isn't working", before shipping any user-facing surface, or after substantive UI changes. Works on live URLs (playwright-cli), local dev servers, and screenshots they paste. Run any deterministic UI-lint oracle you have FIRST, then the judgment passes. Companion: your visual-design system if you keep one, tufte (if charts are present).

UI/UX design intelligence for web and mobile. Includes 50+ styles, 161 color palettes, 57 font pairings, 161 product types, 99 UX guidelines, and 25 chart types across 10 stacks (React, Next.js, Vue, Svelte, SwiftUI, React Native, Flutter, Tailwind, shadcn/ui, and HTML/CSS). Actions: plan, build, create, design, implement, review, fix, improve, optimize, enhance, refactor, and check UI/UX code. Projects: website, landing page, dashboard, admin panel, e-commerce, SaaS, portfolio, blog, and mobile app. Elements: button, modal, navbar, sidebar, card, table, form, and chart. Styles: glassmorphism, claymorphism, minimalism, brutalism, neumorphism, bento grid, dark mode, responsive, skeuomorphism, and flat design. Topics: color systems, accessibility, animation, layout, typography, font pairing, spacing, interaction states, shadow, and gradient. Integrations: shadcn/ui MCP for component search and examples.

Build a distinct character voice systematically from body-effort + imagination (Laban), not just "a funny accent." Load when designing or rendering a character's voice, especially for several distinct NPCs.

Audit UI code against Vercel's Web Interface Guidelines — ~100 terse, checkable rules covering accessibility, focus states, forms, animation, typography, content handling, touch targets, safe areas, dark mode, i18n, and hydration safety. Use when reviewing UI code for compliance ("review my UI", "check accessibility", "audit UX details"), as the rules-pass companion to ui-review (which covers visual craft) and ui-lint (the deterministic oracle).

Comprehensive web quality audit covering performance, accessibility, SEO, and best practices. Use when asked to "audit my site", "review web quality", "run lighthouse audit", "check page quality", or "optimize my website".

Run the table by building on what players offer rather than forcing your plan. The improv spine — player agency is the engine. Load when a scene goes off-script or a player makes an unexpected move.

Stress-test your own practice against your principles. Named for Ben, a colleague who observed 'Your AI is very complimentary. Perhaps too complimentary?' Apply when assessing whether you practice what you've written — in evaluations, proposals, self-assessments, and especially when you're about to tell someone else what they could learn from you.

Make each character sound like themselves — distinct dialogue, body habits, interiority, and the small specifics that make them recognizable across many scenes. Load when writing dialogue, designing characters, or noticing characters blur together.

Three-minute connection scan + tension surface before substantive work. Codifies the discipline of actively scanning for connections to prior work, tensions with existing conclusions, and things the user might not be looking for — at session start. Load at the start of any substantive session, after loading your project's memory files (if you keep them) but before diving into the task.

Software isn't permanent — code is cheap and git walks anything back. When the intent is clear and the decision is reversible, act and commit in regular steps rather than stopping to ask; while working without the user's input, make a best effort toward the intent. Use when deciding whether to act vs. ask, and when working unsupervised. Generalizes beyond code.

Write scenes that move readers — eight rules for letting emotion live in the gap between what is said and what is shown. Load whenever writing fiction prose, especially climactic, intimate, or grief-laden scenes.

Voice an NPC/character in a CHAT room so the dialogue invites a response instead of info-dumping. Load when a persona speaks in-character to humans in a text channel.

Parallelize a large task across many subagents — when the work is independent per-unit (verify / research / generate, the same operation across N files or items) it is a fan-out shape, and parallel fresh-eyes-per-unit beats one tiring serial operator. Use to triage whether a long task should be fanned out, and to run the fan-out safely under an active presiding watch (e.g., a 62-file audit fanned to 6 agents finished in ~15 min and was more thorough than the serial walk that preceded it).

The orchestration pattern for staffing implementation across cheaper model seats — the boss (the orchestrating agent) writes specs and constitutions, dispatches, blind-verifies, and rules on disputes, but never implements. Includes the audition oracle for hiring seats, merge-train discipline (post-merge build oracle on main), and watch-what-you-order review. Load for any arc with several independent implementation units, or any grindable, oracle-covered build.

The green-light triage for any long or autonomous work — ask "what's the oracle?" AND "is it grindable?" before committing to the shape. Use at the moment of green-lighting any overnight or unattended run, fan-out, "keep marching" arc, soak, migration walk, bulk verification, or agent-loop dispatch. Also triggers when sizing why an autonomous run underperformed. Companion to fan-out (the parallel shape).

When must the human be in the loop, vs. when can the agent discern and act on its own? The decision rubric for autonomy scope — four bins (act / act-and-report / surface-first / always-theirs), built on the dave-rule's reversibility lean plus the judgment-source test. Load when deciding whether to act or ask, and especially before any unsupervised run.

Turn a GM/storyteller mistake into a plot point instead of breaking the game. Load when continuity breaks, a secret slips, an encounter goes wrong, or balance tips.

Keep momentum (therefore/but, not and-then) and share the spotlight so every player gets a moment. Load when running a multi-character scene or a session that's dragging or lopsided.

In-session learning capture. Invoke when the user corrects, praises, or steers — log the micro-correction to <your-scratch-dir>/reflect.md immediately, then graduate at session end. Use when noticing corrections ('no', 'wrong'), praise ('perfect', 'exactly'), tool/format preferences, or edge cases.

Lighter sabbath ritual for ending substantive non-cycle work — the declaration + carry-forward, without a full retrospective review. Use at the end of any substantive session to mark the ending. A fuller cycle-review ritual is for completed cycles, milestones, or the literal Sunday Sabbath. This skill is the daily-rhythm version.

Write death, sacrifice, and loss that earns the reader's grief. The hard rule — the reader must love the character before the loss, not because of it. Load when a character is dying, leaving, or losing something they love.

Connect every beat to the next by causation (therefore) or disruption (but), never by mere sequence (and then). The momentum rule for scenes, campaigns, chapters, and arguments. Load when writing a narrative session, a campaign arc, or any prose that lists when it should drive.

Build fictional worlds (D&D, sci-fi, Bridge Simulator) that feel internally consistent and lived-in. Sanderson's Laws, cultural depth without lore-dumping, and the difference between iceberg and infodump. Load when creating a world, faction, city, magic/tech system, or culture.