Skip to content

AGENTS.md

Latest commit

 

History

History
55 lines (43 loc) · 3.83 KB

AGENTS.md

File metadata and controls

55 lines (43 loc) · 3.83 KB
Raw

AGENTS.md — the fleet & workflow

This project is maintained and extended by a fleet of focused Claude Code subagents (defined in .claude/agents/*.md) plus human direction. Each agent has a single area of ownership so that work composes cleanly and reviews stay sharp.

Read CLAUDE.md first for project purpose, conventions, and the quality bar.

The fleet

Agent Owns Use it when
research-validator content/research/ — facts, citations, source quality a claim needs verifying, a citation is added/changed, or copy might be unsupported
content-writer plain-language section copy in content/sections/* turning validated research into audience-appropriate prose
simulator-engineer Section 5 WebGL/GLSL fluid sim, perf, controls, fallback anything touching components/simulator/** or shaders/**
teaching-visuals-engineer Sections 1–4 micro-demos in components/demos/** building/altering a concept demo; picks the lightest technique per demo
motion-designer scroll animation, pacing, reduced-motion correctness scroll/animation behavior, useReducedMotion, pinned sequences
design-system-guardian espresso/latte tokens, typography, contrast, consistency visual/token changes; contrast or hierarchy concerns
accessibility-auditor keyboard nav, semantics, contrast, reduced-motion, a11y before merging UI; auditing a section
presentation-critic the live experience vs. "best presentation available" reviewing a milestone build; empowered to demand iteration

Boundaries that matter

  • simulator-engineer vs. teaching-visuals-engineer: the simulator (Section 5) maximizes visual fidelity (real fluid sim). The teaching demos (Sections 1–4) maximize conceptual clarity and are explicitly not bound to the fluid-sim stack — a clean SVG/Canvas2D diagram beats a fancy confusing effect. Do not let the two bleed into each other.
  • research-validator gates copy: content-writer drafts; research-validator confirms every scientific sentence maps to a citation in citations.ts. No unverified claim ships.

Workflow / how they compose

  1. Research → copy: research-validator maintains the validated doc; content-writer derives prose; research-validator signs off.
  2. Build: simulator-engineer, teaching-visuals-engineer, and motion-designer implement against the validated content and design tokens.
  3. Guard: design-system-guardian and accessibility-auditor review for tokens/contrast/typography and a11y (reduced-motion + keyboard + semantics).
  4. Critique loop: presentation-critic reviews the live build against the quality bar, files concrete issues, and dispatches fixes to the owning specialist. Repeat until it passes. This is the iteration engine behind "best presentation available — iterate if not."

Gates before merge

  • npm run lint && npm run typecheck && npm run test && npm run build all clean.
  • accessibility-auditor pass (reduced-motion, keyboard, contrast).
  • All new copy traced to content/research/latte-emergence.md.

Human + AI transparency

Claude Code transcripts are captured and submitted with the assignment; human judgment and direction are visible throughout (planning, choices, and reviews).